2018 Book MultifractalDetrendedAnalysisM

Guangxi Cao
Ling-Yun He
Jie Cao
Multifractal
Detrended Analysis
Method and Its
Application in
Financial Markets
Multifractal Detrended Analysis Method and Its
Application in Financial Markets
Guangxi Cao Ling-Yun He
•
Jie Cao
Multifractal Detrended
Analysis Method and Its
Application in Financial
Markets
123
Guangxi Cao Jie Cao
School of Management Science and School of Management Science and
Engineering Engineering
Nanjing University of Information Science Nanjing University of Information Science
and Technology and Technology
Nanjing, Jiangsu Nanjing, Jiangsu
China China
Ling-Yun He
School of Economics
Jinan University
Guangzhou, Guangdong
China
ISBN 978-981-10-7915-3 ISBN 978-981-10-7916-0 (eBook)

https://doi.org/10.1007/978-981-10-7916-0
Library of Congress Control Number: 2017963978
© Springer Nature Singapore Pte Ltd. 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
part of Springer Nature
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore
Acknowledgements
We thank for the financial support from National Natural Science Foundation of
China (No. 71371100).
v
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 A Historical Evolution of Fractal Methods . . . . . . . . . . . . . . . . . 2
1.2 Application Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Long Memory Methods and Comparative Analysis . . . . . . . . . . . . 7
2.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 R/S and Modified R/S . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2 DFA Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Estimation and the Descriptive Statistics of the Time-Varying
Hurst Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 11
2.3.1 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 11
2.3.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . .. 15
2.4 Relationship Between the Two Time-Varying Hurst
Exponent Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4.1 Unit Root Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4.2 Cointegration Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.3 Granger Causality Test . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Multifractal Detrended Fluctuation Analysis (MF-DFA) . . . . . . . . . 21
3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.1 MF-DFA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.2 Partition Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Empirical Analysis on Developed-Emerging Agricultural
Futures Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... 26
3.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... 26
3.2.2 Multifractal Spectrum Analysis . . . . . . . . . . . . . . ...... 27
vii
viii Contents
3.2.3 Sources of Multifractality . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2.4 Comparative Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Empirical Analysis on Crude Oil Markets . . . . . . . . . . . . . . . . . 35
3.3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.2 Multifractality and Its Dynamical Formation
Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 36
3.3.3 Multifractal Detrended Fluctuation Analysis . . . . . . . . .. 38
3.3.4 Sources of Multifractality . . . . . . . . . . . . . . . . . . . . . . .. 39
3.3.5 Multifractal Analysis of Price Fluctuations at Different
Scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 44
3.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 46
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 46
4 Multifractal Detrended Cross-Correlation Analysis
(MF-DCCA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2 Empirical Analysis on Chinese Stock-Exchange Market . . . . . . . 52
4.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2.2 Cross-Correlation Test . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.3 Multifractal Detrended Cross-Correlation Analysis . . . . . . 56
4.2.4 Scaling Consistency Analysis . . . . . . . . . . . . . . . . . . . . . 58
4.2.5 Dynamics of Cross-Correlations Over Time . . . . . . . . . . . 59
4.2.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3 Empirical Analysis on Price-Volume Relationships in
Agricultural Commodity Futures Markets . . . . . . . . . . . . . . . . . . 65
4.3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3.2 Cross-Correlation Test . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.3 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5 Asymmetric Multifractal Detrended Fluctuation Analysis
(A-MFDFA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.1.1 A-MFDFA Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.1.2 Asymmetric GARCH Model . . . . . . . . . . . . . . . . . . . . . 83
5.2 Empirical Analysis on Shanghai-Shenzhen Stock Market . . . . . . 84
5.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.2 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Contents ix
5.3 Empirical Analysis on International Gold Markets . . . . . . . . . . . 95

5.3.1 Descriptive Statistics Analysis of Gold Price . . . . . . . . . . 95
5.3.2 Analysis of Asymmetric Scaling Behavior . . . . . . . . . . . . 97
5.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.3.4 Asymmetric Influences of Good and Bad News on Gold
Price Fluctuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6 Asymmetric Multifractal Detrended Cross-Correlation Analysis
(MF-ADCCA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.2 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.3 Empirical Cross-Correlation Analysis on
Chinese Stock Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.3.1 Asymmetric Test for Different Trends of the Chinese
Stock Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.3.2 Asymmetric Test for Different Trends of the Other
Financial Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.3.3 Multifractal of the Asymmetric Cross-Correlation . . . . . . 124
6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7 Asymmetric DCCA Cross-Correlation Coefficient . . . . . . . . . . . . . . 129
7.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.1.1 DCCA Cross-Correlation Coefficient . . . . . . . . . . . . . . . . 130
7.1.2 Statistical Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7.2 Empirical Analysis on Crude Oil Spot and Futures Markets . . . . 132
7.2.1 Data and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . 132
7.2.2 Estimation of the Cross-Correlation Coefficient . . . . . . . . 134
7.2.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.3 Empirical Analysis on Carbon and Energy Market . . . . . . . . . . . 141
7.3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.3.2 Asymmetric Cross-Correlation Coefficient Test . . . . . . . . 142
7.3.3 MF-ADCCA Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 144
7.3.4 Asymmetric Volatility-Constrained Correlation and
Volatility-Transmission Direction . . . . . . . . . . . . . . . . . . 146
7.3.5 Implications and Conclusions . . . . . . . . . . . . . . . . . . . . . 151
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8 Simulation—Taking DMCA as an Example . . . . . . . . . . . . . . . . . . 155
8.1 ARFIMA Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.2 DMCA Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.2.1 Detrended Moving-Average Cross-Correlation Analysis
(DMCA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
x Contents
8.2.2 Comparative Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

8.2.3 Empirical Analysis on the International Crude Oil Spot
Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
8.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9 Multifractal Detrend Method with Different Filtering . . . . . . . . . . . 169
9.1 Nonlinear Structure Analysis of Carbon and Energy Markets:
MFDCCA-MODWT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
9.1.1 MODWT Methodology . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.1.2 Comparative Analysis on the Performance of
MFDCCA-MODWT and MF-X-DMA . . . . . . . . . . . . . . 172
9.1.3 Empirical Results and Analysis . . . . . . . . . . . . . . . . . . . . 176
9.1.4 Original of Multifractality . . . . . . . . . . . . . . . . . . . . . . . . 186
9.1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
9.2 Multifractal Features of EUA and CER Futures Markets:
MFDFA-EMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.2.1 MFDFA-EMD Methodology . . . . . . . . . . . . . . . . . . . . . 191
9.2.3 Origins of Multifractality . . . . . . . . . . . . . . . . . . . . . . . . 201
9.2.4 Risk Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
9.2.5 Conclusion and Implications . . . . . . . . . . . . . . . . . . . . . . 205
9.3 Cross-Correlation Among Mainland China, US, and Hong
Kong Stock Markets VC-MF-DCCA . . . . . . . . . . . . . . . . . . . . . 206
9.3.1 VC-MF-DCCA Method . . . . . . . . . . . . . . . . . . . . . . . . . 208
9.3.2 Validation of the VC-MF-DCCA Method . . . . . . . . . . . . 209
9.3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
10 Risk Analysis Based on Multifractal Detrended Method . . . . . . . . . 223
10.1 Asymmetric MF-DCCA Method Based on Risk Conduction
and Its Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
10.1.1 Asymmetric MF-DCCA Method Based on Different
Risk Conduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
10.1.2 Nonlinear Causality Testing . . . . . . . . . . . . . . . . . . . . . . 227
10.1.3 Sample Selection and Descriptive Statistics Analysis . . . . 229
10.1.4 Simulation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
10.1.5 Empirical Analysis of Asymmetric
MF-DCCA Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
10.1.6 Nonlinear Granger Causality Test Based on the
Remover of Long Memory . . . . . . . . . . . . . . . . . . . . . . . 235
10.1.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Contents xi
10.2 Extreme Values Evaluation Based on Detrended Fluctuation

Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
10.2.1 Threshold Estimation Method . . . . . . . . . . . . . . . . . . . . . 241
10.2.3 Simulation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
10.2.4 Time-Clustering of Extreme Events . . . . . . . . . . . . . . . . . 246
10.2.5 Effect of Extreme Value on the Cross-Correlation of
Chinese and American Stock Markets . . . . . . . . . . . . . . . 247
10.2.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
10.2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
10.3 Research Prospect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
10.3.1 Risk Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
10.3.2 Cross Market Investment Research . . . . . . . . . . . . . . . . . 252
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Chapter 1
Introduction
For a long time, academics generally believe that returns of stock prices follow a
normal distribution and price behavior obeys a so-called ‘random-walk’ hypothesis.
This notion was first introduced by Bachelier in 1900 (Bachelier 1900), since then it
has been adopted as the essence of many asset pricing models. However, empirical
evidence suggests that returns in financial markets have fundamentally different
properties (Mandelbrot 1963, 1971, 1997). Mandelbrot and Fama proposed that
capital market returns are subject to fractal distribution on 1963. These ubiquitous
properties identified are: fat tails, volatility clustering, multi-scaling, among others.
These wide tail distributions often show a sign of long-term memory produced by a
nonlinear stochastic process. Another important context on this domain is the
efficient market hypothesis (EMH) proposed by Fama which states that stock prices
already reflect all available information useful in evaluating their value (Fama
1970). However, it has been widely criticized in the financial literature as this
hypothesis presupposes that all information is already included in the price of
assets. Another perspective is that of econophysics, which was developed in recent
years to describe the contributions of physicists to finance and economics. Some
statistical physicists study independence through the concept of scaling. The
exponent of this scaling relationship between the standard deviation of a time series
and the time increments used is the Hurst exponent. Evidence of H differences from
a half (1/2) could be interpreted as proof that returns are not independent and that
long-term memory is present (Peters 1994, 1996). This shows that the volatility of
securities prices to a certain extent, there is predictability. Rachev (2010), Paolella
(2016), Francq (2016) etc. have confirmed that the fractal distribution have freat
adaptability in the financial market, and opened up a new path for the financial
market forecast.
© Springer Nature Singapore Pte Ltd. 2018 1

G. Cao et al., Multifractal Detrended Analysis Method and Its Application
in Financial Markets, https://doi.org/10.1007/978-981-10-7916-0_1
2 1 Introduction
1.1 A Historical Evolution of Fractal Methods
Fractal methods are divided into single fractal and multifractal methods.
Single-fractal analysis is mainly the long memory (long memory) (also known as
Persistence) or anti-persistence. The long-term memory (long-range correlation) in
the financial time series is mainly judged by the Hurst index estimated by various
methods. One of the earlier and most widely used methods of calculating the Hurst
index is the RTC method (R/S). It is a nonparametric statistical method proposed by
Hurst (1951), which is applied by Mandelbrot (1963) to the analysis of time series.
Peters (1994) systematically summarizes the method of calculating the Hurst index,
the significance test, the V statistical method for calculating the average cycle
length, and the empirical criteria for selecting the sample data. R/S analysis can
distinguish between random time series and associated time series, but it depends
on the maximum and minimum values of the data sequence, which are more
sensitive to the outliers and therefore can only be applied to stationary time series
and trendless time series. Peng et al. (1994) proposed a DFA method that can detect
long-range correlations of non-stationary time series. The DFA method can elim-
inate the pseudo-correlation phenomenon in artificial non-stationary time series, and
it can judge the order of time series trend by comparison and analysis, so it is more
advantageous than R/S analysis. In addition, in addition to the use of binomial trend
as a trend, but also with the mobile local mean as a trend. Alessio et al. (2002)
proposed a DMA method for estimating the Hurst value of time series. Although
the DFA method is a very effective scaling method for detecting the long memory
of time series. However, since the statistical distribution of this statistical method
has not been found so far, the reliability of the DFA method can not be theoretically
Guarantee. Therefore, this book will use the method of numerical simulation to
study the small sample precision and asymptotic properties of DFA fractal methods
under different sample lengths, and find the optimal DFA method in the case of
small sample. With the deepening of the research, scholars have found that the
multi-scale and complexity of the financial time series lead to the single fractal can
not meet the needs of the study. Mandelbrot (1999) pointed out that, compared with
the single fractal process, the multifractal analysis theory is a better tool for
quantitatively characterizing various complex fluctuations in financial markets, with
stronger applicability. Based on the DFA (detrended fluctuation analysis) method,
Kantelhardt et al. (2002) proposed the multifractal detrended fluctuation analysis
(MFDFA) for the first time to describe the multifractal features of time series under
different time scales. In addition, Gu and Zhou (2010) extended the Moving
Average Elimination Trend Fluctuation (DMA) to Multifractal Detrend moving
Algorithm (MFDMA).
However, the above method, whether it is a single fractal or multiple fractal, is a
single time series of fractal features are described. The elimination trend method can
also be generalized for the correlation analysis of two time series to eliminate the
trend-related cross-correlation analysis (DCCA) method. DCCA is a technique for
measuring the long-term correlation of two non-stationary time series. In 2008,
1.1 A Historical Evolution of Fractal Methods 3
Podobnik and Stanley introduced the DFA method into the elimination of trend
correlation analysis. In order to quantitatively characterize the long-term correlation
of two time series, Podobnik et al. (2009) proposed the Qcc(m) test method.
Zebende et al. (2013) proposed the DCCA correlation coefficient method when
analyzing climate time series. However, the calculation of DCCA correlation
coefficient depends on the overlapping time window. Therefore, Podobnik et al.
(2011) proposed a correlation coefficient calculation method based on
non-overlapping time windows, and test and compare the DCCA correlation
coefficients of overlapping and non-overlapping time windows Statistical signifi-
cance. For the two time periods with the same time series, combined with the
elimination of the trend of cross-correlation analysis and multiple fractal, Zhou
(2008) proposed a multi-fractal elimination trend fluctuation cross correlation
analysis (MFDCCA), for the detection of two sequences of multiple fractal struc-
ture. Jiang and Zhou (2011) proposed MFXDMA method and MFDCCA method
compared with the different artificial generation of data, the advantages and dis-
advantages of the two are not the same. Cao and Xu (2016) compare MFXDMA
with MFDCCA-MODWT to find that the result of the latter is closer to the theo-
retical value when the sequence length is equal to 10,000 and the result is similar
when the sequence equals 2000. DCCA and MFDCCA have been widely used to
analyze the cross correlation between financial time series.
The fractal method can also be used for asymmetric analysis. (A-DFA) proposed
by Alvarez et al. (2009), which brings the study of time series asymmetric behavior
into an effective measure that can detect positive and negative trend sequences in
time series Asymmetric scale behavior. First of all, A-DFA is extended to the
multifractal case-AMFDFA, which can accurately characterize the multifractal
features of positive and negative trend sequences in time series. Further research
found that there are asymmetric correlations in financial markets, and because
asymmetries are prevalent in financial markets, the impact of asymmetries on
earnings and risks, especially risks, is considered in the study of portfolios, and
there are no scholars currently engaged in related research on the portfolio strategy.
In addition, there are power spectral methods (L.I. Rudin 1992), structural
function test methods (Provenzale 1992), Lyapunov index (Wolf 1985) and so can
be used to study the single fractal characteristics. Wavelet transformations,
high-height correlation functions, box-counting, partition function method, and
box-counting methods are more abundant. In practice, however, the method of
eliminating the trend class occupies an important position in the method category
of fractal features. This method can calculate the fractal characteristics of
non-stationary time series and remove the pseudo-correlation, so it is more practical
and applicable than other. The method has been greatly improved, so the research of
this topic will be carried out by DFA class method.
4 1 Introduction
1.2 Application Areas
Existing research findings are abundant. Ma et al. (2013), Wang et al. (2013), and
Wang and Xie (2013) empirically analyze the cross-correlation between the Chinese
stock market and adjacent stock markets, between price returns and trading volumes
for the CSI 300 index futures, and between the Renminbi and four major currencies,
respectively. Siokis (2013), through multifractal analysis of stock exchange crashes,
reports that temporal correlations play a substantial role during an extreme event.
Zebende et al. (2013) establish a well-defined relationship between the long-range
auto-correlation exponent and the long-range cross-correlation exponent, which will
be accomplished theoretically by differentiating the DCCA cross-correlation coeffi-
cient. Podobnik et al. (2010) study the long-range cross-correlations for multiple time
series, precisely the return time series of the NYSE members.
Many scholars have compared many different markets and investigated their multi-
fractal properties. Matia et al. (2003) investigated daily prices of 29 commodities and
2449 stocks, and found that the price returns for commodities have a significantly
broader multifractal spectrum than for stocks; Zunino et al. (2008) investigated the
multifractality degree of developed and emerging stock market indices, and found that
higher multifractality is associated with a less developed market (2008); Jiang and Zhou
(2008) also investigated the emerging and developed stock markets, and found that there
are not multifractality in the original series of the two markets, but their results on
China’s stock indexes shows that there are multifractality properties in those markets.
Matos et al. (2008) use a new method of studying the Hurst exponent with time and
scale dependency to recover the major events affecting worldwide markets which can
measure and compare the behaviors in emergent/established markets. Current studies
focused on the commodity market, stock market and some other fields.
References
E. Alessio, A. Carbone, G. Castelli, V. Frappietro, Second-order moving average and scaling of

stochastic time series. Eur. Phys. J. B 2, 197–200 (2002)
J. Alvarez-Ramirez, E. Rodriguez, E.J. Carlos, A DFA approach for assessing asymmetric
correlations. Phys. A 388, 2263–2270 (2009)
L. Bachelier, Théorie de la spéculatio. Gauthier-Villars, (1900)
G. Cao, W. Xu, Nonlinear structure analysis of carbon and energy markets with MFDCCA based
on maximum overlap wavelet transform[J]. Physica A 444, 505–523 (2016)
C. Francq, J.-M. Zakoïan, Estimating ARCH Models when the Coefficients are Allowed to be
Equal to Zero. Aust. J. Stat. 37(1), 31–40 (2016)
G.-F. Gu, W.-X. Zhou, Detrending moving average algorithm for multifractals. Phys. Rev. E. 82
(1), 011136 (2010)
H.E. Hurst, The long-term storage capacity of reserviors. Trans. Am. Soc. Civ. Eng. 116(1), 770–
799 (1951)
Z.-Q. Jiang, W.-X. Zhou, Multifractality in stock indexes: fact or fiction? Physica A 387, 3605–
3614 (2008)
References 5
Z.-Q. Jiang, W.-X. Zhou, Multifractal detrending moving average cross-correlation analysis. Phys.
Rev. E 84, 016106 (2011)
J.W. Kantelhardt, S.A. Zschiegner, E. Koscielny-Bunde, S. Havlin, A. Bunde, H.E. Stanley,
Multifractal detrended fluctuation analysis of nonstationary time series. Phys. A 316, 87–114 (2002)
F. Ma, Y. Wei, D.-S. Huang, Multifractal detrended cross-correlation analysis between the
Chinese stock market and surrounding stock markets. Phys. A 392, 1659–1670 (2013)
B.G. Malkiel, E.F. Fama, Efficient capital markets: a review of theory and empirical work.
J. Finance 25(2), 383–417 (1970)
B.B. Mandelbrot, The variation of certain speculative prices. J. Bus. 36(4), 394–419 (1963)
B.B. Mandelbrot, B. Benoit. When can price be arbitraged efficiently? A limit to the validity of the
random walk and martingale models. Rev. Econ. Stat. 225–236 (1971)
B.B. Mandelbrot, The variation of the prices of cotton, wheat, and railroad stocks, and of some
financial rates[M], Fractals and Scaling in Finance (Springer, New York, 1997), pp. 419–443
B.B. Mandelbrot, B. Benoi, A multifractal walk down Wall Street. Sci. Am. 280(2), 70–73 (1999)
K. Matia, Y. Ashkenazy, H.E. Stanley, Multifractal properties of price fluctuations of stocks and
commodities. Europhys. Lett. 61, 422–428 (2003)
J.A.O. Matos, S.M.A. Gama, H.J. Ruskin, A.A. Sharkasi, M. Crane, Time and scale Hurst
exponent analysis for financial markets. Physica A 387, 3910–3915 (2008)
M.S. Paolella, Stable-GARCH models for financial returns: fast estimation and tests for stability.
Econom. 4(2), 25 (2016)
C.K. Peng, S.V. Buldyrev, S. Havlin, M. Simons, H.E. Stanley, A.L. Goldberger, Mosaic
organization of DNA nucleotides. Phys. Rev. E 49, 1685–1689 (1994)
E.E. Peters. Fractal Market Analysis: Applying Chaos Theory to Investment and Economics,
(Wiley, 1994)
E.E. Peters. Chaos and Order in the Capital Markets: A New View of Cycles, Prices, and Market
Volatility (Wiley, 1996)
B. Podobnik, H.E. Stanley, Detrended cross-correlation analysis: a new method for analyzing two
nonstationary time series. Phys. Rev. Lett. 100, 084102-1–084102-11 (2008)
B. Podobnik, I. Grosse, D. Horvati, S. Ilic, P. Ivanov, H. Ch, E. Stanley, Quantifying
cross-correlations using local and global detrending approaches. Eur. Phys. J. B 71, 243–250
(2009)
B. Podobnik, D. Wang, D. Horvatić, I. Grosse, H.E. Stanley, Time-lag cross-correlation in
collective phenomena. Europhys. Lett. 90, 68001 (2010)
B. Podobnik, Z.Q. Jiang, W.X. Zhou, H.E. Stanley, Statistical tests for power-law cross-correlated
processes. Phys. Rev. E 84(6), 066118 (2011)
A. Provenzale, et al. Distinguishing between low-dimensional dynamics and randomness in
measured time series. Physica D: nonlinear phenomena. 58(1–4), 31–49 (1992)
S.T. Rachev, M. Hoechstoetter, F.J. Fabozzi et al., Probability and statistics for finance. John
Wiley & Sons, (2010)
L.I. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithms. Physica
D: Nonlinear Phenomena. 60(1-4), 259–268 (1992)
F.M. Siokis, Multifractal analysis of stock exchange crashes. Phys. A 392, 1164–1171 (2013)
G.-J. Wang, C. Xie, Cross-correlations between Renminbi and four major currencies in the
Renminbi currency basket. Phys. A 392, 1418–1428 (2013)
D.-H. Wang, Y.-Y. Suo, X.-W. Yu, M. Lei, Price-volume cross-correlation analysis of CSI300
index futures. Phys. A 392, 1172–1179 (2013)
Wolf, Alan, et al. Determining Lyapunov exponents from a time series. Physica D: Nonlinear
Phenomena. 16(3) 285–317 (1985)
G.F. Zebende, M.F. da Silva, A.M. Filho, DCCA cross-correlation coefficient differentiation:
theoretical and practical approaches. Phys. A 392, 1756–1761 (2013)
W.X. Zhou, Multifractal detrended cross-correlation analysis for two nonstationary signals. Phys.
Rev. E 77, 066211-1–066211-9 (2008)
L. Zunino, B.M. Tabak, A. Figliola, D.G. Pérez, M. Garavaglia, O.A. Rosso, A multifractal
approach for stock market inefficiency. Physica A 387, 6558–6566 (2008)
Chapter 2
Long Memory Methods and Comparative
Analysis
Long memory property has been empirically analyzed by many researchers in stock
markets (Peters 1991; Eisler and Kertesz 2007; Roman 2008) and exchange markets
(Muniandy et al. 2001; Tabak and Cajueiro 2006), and is still a hot topic of active
researches (Gu and Zhou 2009; Abounoori et al. 2012; Wang et al. 2011). Hurst
exponent is usually used to measure the long memory in financial time series.
Various techniques of long memory analysis have been proposed, but R/S (rescaled
range analysis) and DFA (detrended fluctuation analysis) are the most commonly
used methods. The classic R/S was proposed by hydrological analysis of Hurst
(1951), and then was applied to analysis of financial time series firstly by
Mandelbrot (1963). However, the classic R/S is sensitive to the size of sample.
Thereby, Anis and Lloyd (1976), Lo (1991) and Peters (1994) made some modi-
fication on the small sample time series. In addition, R/S is sensitive to the extreme
value depending on the maximum and minimum of the time series. To overcome
this shortcoming, DFA was proposed by Peng et al. (1994), which has been used
widely in the analysis of the long memory in stock market (Eldridge et al. 1993,
Greene and Fieltz 1997) and exchange market Ausloos (2000). Granero et al.
(2008) suggested that the estimation of modified R/S (denoted by R/S-AL) is better
than R/S, especially with short series. Moreover, Weron (2002) believe that the
DFA method turns out to be the unanimous winner than others when testing R/S, R/
S-AL and DFA on samples drawn from Gaussian white noise. However, Kristoufek
(2010) argued that the R/S still remains useful and robust even when compared to
DFA method. Therefore, to get more precise results, comprehensive use of various
methods may be a good choice in the empirical analysis.
Additionally, the long memory of financial time series may present time-varying
characteristic, for exogenous and endogenous shocks can have a significant effect
on the dynamics of financial markets. Recently, some researchers have paid
attention to the property of time-varying long memory in financial time series using
different methods. Carbone et al. (2010) suggested that the Hurst exponent of
financial series shows a much richer time-variability than that of monofractal
artificial series, using German stock index and government bond time series.
8 2 Long Memory Methods and Comparative Analysis
Cajueiro and Tabak (2005) tested time-varying long memory of stock indices for 11
emerging markets and the developed markets of US and Japan, using R/S and
modified R/S statistics. Alvarez-Ramirez et al. (2008) evaluated the time-varying
Hurst exponent for US stock markets using DFA, suggesting that the Hurst
exponent displays an erratic dynamics and the major breakthrouth of the long-term
trend is related to the huge shocks such as the collapse of the Bretton Woods
system. Wang et al. (2009) analyzed the time-varying Hurst exponent for Shanghai
stock market over time, and found that the Shanghai stock market becomes more
and more efficient after the price-limited reform.
Unfortunately, there are little literatures studying the time-varying long memory
of foreign exchange market, especially of the Chinese exchange market. In fact,
after the reform of Chinese RMB exchange rate on July 21, 2005, RMB exchange
rate is more flexible than before. Furthermore, RMB exchange rate flexibility was
narrowed on June 19, 2009 for the 2008 US financial crisis, and was increased
again on June 19, 2010 by People’s Bank of China (PBC). However, to evaluate the
efficiency of financial market, it may be bias for testing the long memory to divide
the entire sample to several sub-samples, and can not trace the efficiency evolution.
Thus, it may be a good method to estimate time-varying long memory by using
rolling windows technology.
In addition, the relationship between stock market and exchange market has been
discussed in many literatures (Aydemir and Demirhan 2009; Nieh and Yau 2010),
by using the VAR model. However, different from the traditional object of study, in
this chapter, we attempt to detect the relationship of the Chinese exchange market
and stock market, by anglicising the relationship between the time-varying Hurst
exponents of the Chinese exchange market and those of the Chinese stock market.
To the best of our knowledge, this is the first time to study the relationship of the
time-varying Hurst exponents.
Therefore, in this chapter, to obtain more precise results, we applied the R/S.
modified R/S and DFA methods comprehensively to measure time-varying long
memory so as to detect the evolution of efficiencies of the Chinese exchange market
and stock market. We find that the efficiency of the Chinese exchange market is not
improved significantly, while the efficiency of the Chinese stock market is
improved steadily, but there are little changes in recent months. Moreover, the
time-varying Hurst exponent is sensitive to the reform of enhancing flexibility of
RMB exchange rate, which improves the efficiency of the Chinese exchange market
in short term and affect little in long term. Additionally, we investigate the rela-
tionship between the time-varying Hurst exponents of the Chinese exchange market
and those of the Chinese stock market, using return and volatility series respec-
tively. As the results shown that there are long-run equilibrium relationship and
bidirectional Granger causal relationship in short term.
2.1 Methodology 9
2.1 Methodology
Assume that there is a time series xðiÞ; i ¼ 1; 2; . . .; N. The R/S, modified R/S and
DFA methods are introduced in the following.
2.1.1 R/S and Modified R/S
To estimate the Hurst exponent, the best-known classic R/S analysis was proposed
by Mandelbrot and Wallis (1969), which can be described as the following.
The time series xðiÞ of length N are divided into s sub-series of length n. For each
sub-series a ¼ 1; 2; ; s; xk;a denotes the element in sub-series a; k ¼ 1; 2; ; n.
Then:
P
n
① calculate the mean value Ea ¼ 1n xk;a , and create a cumulative series
k¼1
P
k
yk;a ¼ ðxi;a Ea Þ;
i¼1
② find the range Ra ¼ max ðyk;a Þ min ðyk;a Þ and standard deviation
1kn 1kn
P
n
2 0:5
Sa ¼ ð1n ðxk;a Ea Þ Þ ; a ¼ 1; 2; ; s;
k¼1
P
s
③ calculate the mean value of the rescaled range ðR=SÞn ¼ 1s ðRa =Sa Þ
a¼1
for all sub-series of length n. As it shown that the R/S statistics
asymptotically follows the relation ðR=SÞn / cnH , where H is called
Hurst exponent. The value of H can be obtained by running a simple
linear regression over a sample of increasing time horizons
log ððR=SÞn Þ ¼ log c þ H log n; ð2:1Þ
Equivalently, the value H can be obtained by estimating the slope of

linear regression Eq. (2.3).
If H [ 0:5, the time series is persistent (long memory). If 0\H\0:5, the time
series is anti-persistent. If H ¼ 0:5, there is no correlation or at most short-range
correlation in the time series. However, it should be noted that there is a significant
deviation from the 0.5 slope for small n. Thereby, Anis and Lloy (1976) and Peters
(1994) proposed the modified theoretical (i.e. for white noise) values of the R/S
statistics which are usually approximated by Weron (2002).
8 nP1 qffiffiffiffiffiffi
>
> n0:5 Cððn1Þ=2Þ
< n pffiffipCðn=2Þ i ; n 340;
ni
EðR=SÞn ¼ i¼1
1 qffiffiffiffiffiffi
ð2:2Þ
> n0:5 1 nP
>
: n pffiffiffiffiffiffiffi i ;
ni
n [ 340:
np=2 i¼1
where C is the Euler gamma function. The ðn 0:5Þ=2 term was added by Peters
(1994) to improve the performance for very small n. According to Weron (2002),
once Eq. (2.2) is calculated, the Hurst exponent H will be obtained by calculating
the regression on
log Hn ¼ log c þ ðH 0:5Þ log n; ð2:3Þ
where log Hn ¼ log ½ðR=SÞn EðR=SÞn . The Hurst exponent H estimated by

Eq. (2.3) is denoted by R/S-AL. Besides, the Hurst exponent H estimated by the-
oretical Eq. (2.2) and classic R/S Eq. (2.1) are denoted by R/S-Theo and R/S,
respectively.
2.1.2 DFA Method
The DFA method can be summarized as the following Peng (1994):

① construct the profile
X
i
yðiÞ ¼ ðxðtÞ xÞ; i ¼ 1; 2; . . .; N: ð2:4Þ
t¼1
P
N
where x ¼ N1 xðtÞ;
t¼1
② divide the profile fyðiÞg into s ¼ ½N=n non-overlapping windows of
length n. The local trend for each window v is evaluated by least-square
fit of the data, v ¼ 1; 2; . . .; s.
③ calculate the mean square fluctuation of the integrated and detrended
P
n
time series F 2 ðn; vÞ ¼ 1n z2 ði; vÞ, where v ¼ 1; 2; . . .; s;
i¼1
④ calculate the mean value of the root mean square fluctuation
P
s
FðnÞ ¼ 1s ½F 2 ðn; vÞ1=2 , for all sub-series of length n.
v¼1
Like in the R/S analysis, a line relationship on a double-logarithmic
graph of FðnÞ against the interval size n indicates the presence of a
power-law scaling of the form cnH. The implication of H is the same as it
for R/S analysis. Furthermore, when H [ 0:5, the larger is the Hurst
exponent H, the inefficiency is the financial market.
2.2 Data 11
2.2 Data
The RMB/USD exchange rate is the most important data in the Chinese exchange
market, and the Shanghai Stock Exchange Composite Index (SSCI) is one of the
most important stock market indices in the Chinese stock market. Thus, we select
the daily nominal RMB/USD exchange rate and the closing price of the SSCI as the
price variables of the Chinese exchange market and stock market. Given that the
reform of the Chinese exchange occurred on 21 July 2005, the sample interval is
from 1 August 2005 to 20 October 2011. A total of 1392 pairs of common price
data were adopted after removing the data from different trading dates in the RMB
exchange market and stock market. The original sample data were obtained from
FC Station 3.0 (http://icbc-zx1.finchina.com/product-2.asp).
Returns are computed by rt ¼ log ðPt Þ log ðPt1 Þ, with Pt being the closing
price at time t. Daily volatility of returns are denoted by absolute return, introduced,
according to Cajueiro and Tabak (2005). Thereby, for both return series and
volatility series, the number of empirical data is 1391 pairs. The descriptive
statistics of returns and volatilities of RMB/USD and SSCI (denoted by RER, CSR,
V-RER and V-CSR respectively) are provided in Table 2.1.
In Table 2.1, for all time series, the skewness and kurtosis are different from 0,
showing significant departures from normality. The Jarque-Bera statistics Engle and
Granger (1987) are significant at 1% significant level, suggesting that, for all series,
the normality assumption of returns and volatilities can be rejected. Moreover, the
degree of departures from normality for volatility series is more serious than for
return series. The Jarque-Bera Q statistics with all lags are all significant on the 1%
significant level, indicating that all series may present long memory.
2.3 Estimation and the Descriptive Statistics

of the Time-Varying Hurst Exponent
For time series of returns rt (or volatilities vt) (rt−w+1, rt−w+2…, rt) with time win-
dows with length w, time varying Hurst exponent at time t is denoted by Ht. Then,
Ht + i can be calculated from the time series (rt−w+i+1, rt−w+i+2…, rt+i). We perform
the estimation of the time-varying Hurst exponents, following Cajueiro and Tabak
(2005), with the length of time window w = 1000.
2.3.1 Estimation
Figures 2.1 and 2.2 present the Hurst exponents estimated for returns and volatil-
ities of RMB/USD exchange rate by using R/S, R/S-AL, DFA and R/S-Theo
methods. All estimated Hurst exponents are larger than 0.5 and change over time,
12
Table 2.1 Descriptive statistics for returns and volatilities of SSCI and RMB/USD
Mean Max Min S.D. Ske Kur J-B Q(3) Q(15) Q(30)
CSR 0.0006 0.0903 −0.1276 0.0202 -0.50 6.44 732.94*** 12.77*** 39.27*** 57.31***
*** *** ***
RER −0.0002 0.0036 −0.0043 0.0009 −0.63 6.19 535.44 56.82 77.38 97.43***
*** *** ***
V-CSR 0.0143 0.1276 0.0000 0.0143 2.12 9.88 3787.31 73.10 401.32 713.77***
*** *** ***
V-RER 0.0006 0.0043 0.0000 0.0007 2.00 7.66 2186.86 542.9 1996.1 3693.3***
Notes *,** and *** denote 10%,5% and 1% significant level respectively; Symbols “Max”, ‘Min’,“S.D.”, “Ske”, “Kur” denote Maximum, Minimum, Stv.Dev,
Skewness and Kurtosis respectively; The J-B denotes Jarque-Bera statistics testing for the normality assumption. Q(i) denotes the value of Jarque-Bera
statistics with i lags
2 Long Memory Methods and Comparative Analysis
2.3 Estimation and the Descriptive Statistics of the Time-Varying … 13
0.7
0.68 11/08/2010
0.66
0.64
0.62
H
0.6
0.58
0.56 DFA
R/S-AL
0.54 11/08/2010
R/S
R/S-Theo
0.52
1000 1050 1100 1150 1200 1250 1300 1350 1400
Sample points
Fig. 2.1 Time-varying Hurst exponents evaluated for the returns of RMB/USD
0.95
0.9
0.85
0.8
H
0.75 22/06/2010
0.7
DFA
0.65 R/S-AL
R/S
0.6 R/S-Theo
0.55
1000 1050 1100 1150 1200 1250 1300 1350 1400
Sample points
Fig. 2.2 Time-varying Hurst exponents evaluated for the volatilities of RMB/USD
indicating that long memory exist in return series and volatility series of the
Chinese exchange market. Furthermore, the breakpoint can be seen in Figs. 2.1 and
2.2 respectively (especially in Fig. 2.2). The breakpoint in Fig. 2.1 is 11 August,
2010, while in Fig. 2.2 it is 22 June 2010 which is precisely the 2nd trading date
after the increasing flexibility of RBM exchange rate announced by PBC. After
increasing of the RMB exchange rate flexibility, Hurst exponents of the RMB/USD
returns or volatilities increase gradually again in long term. This indicates that the
adjustment of the RMB exchange rate flexibility improves the efficiency in short
term, but it does not change the nature of the inefficiency of the Chinese exchange
market for the Chinese financial environment has not been improved significantly.
Figures 2.3 and 2.4 plot the time-varying Hurst exponents estimated for return
series and volatility series of China’s stock market. Hurst exponents are all larger
than 0.5 and change over time, implying that the long memory of the Chinese stock
market is time-varying, either for return series or for volatility series.
0.7
DFA
0.68 R/S-AL
R/S
0.66 R/S-Theo
0.64
0.62
H
0.6
0.58
0.56
0.54
0.52
1000 1050 1100 1150 1200 1250 1300 1350 1400
Sample points
Fig. 2.3 Time-varying Hurst exponents estimated for the returns of SSCI
0.9
DFA
R/S-AL
0.85 R/S
R/S-Theo
0.8
0.75
H
0.7
0.65
0.6
0.55
1000 1050 1100 1150 1200 1250 1300 1350 1400
Sample points
Fig. 2.4 Time-varying Hurst exponents estimated for the volatility of SSCI
Additionally, from Figs. 2.1–2.4, we can find that evolution of Hurst exponents
of the Chinese exchange market is relatively slower than that of the stock market
overall. The change rate of Hurst exponents of RMB/USD exchange rate is smaller.
Time-varying Hurst exponents of the Chinese stock market decrease slowly before
May 2011 and decrease steady after May 2011, for either return series or volatility
series (especially for volatility series). These imply that the efficiency of the
Chinese foreign market is not improved significantly, but the efficiency of the
Chinese stock market is improved steadily. In fact, after the reform of China’s split
share structure on 2005, the market-oriented level in the Chinese stock market is
improved, promoting the efficiency improve of the Chinese stock market. However,
with the full-blown of debt crisis in Europe on 2011, stock market went down and
caught in a bear market, which influence the efficiency of the Chinese stock market.
On the other hand, some reforms of foreign exchange market have been
2.3 Estimation and the Descriptive Statistics of the Time-Varying … 15
implemented in China, and to some extent, the flexibility has been increased. The
reform of the Chinese exchange rate regime is progressive, so the market-oriented
level is not increased so rapidly that its efficiency is not improved significantly.
2.3.2 Descriptive Statistics
Tables 2.2 and 2.3 provide the descriptive statistics of time varying Hurst expo-
nents estimated for return and for volatility series of the Chinese exchange market
and stock market, respectively, by R/S, R/S-AL and DFA methods (Hurst exponent
estimated by R/S-Theo is a constant, so it is not presented in Table 2.1). In
Tables 2.2 and 2.3, the means indicate that, for each financial market, the Hurst
exponents of volatility series are larger than those of return series. This indicates
that volatility of RMB/USD exchange market and stock market present longer
memory than their return. Moreover, either for return or for volatility series, the
Hurst exponents of the Chinese exchange market are larger than those of the
Chinese stock market (also can be found from Figs. 2.1–2.4), implying that Chinese
exchange market is more inefficient than stock market.
Furthermore, Table 2.2 shows that the skewness of Hurst exponents of RER is
negative while that of CSR is positive, indicating that the distribution of time
varying Hurst exponent estimated for return series of the Chinese foreign exchange
market and stock market are not normal distribution, and present “fat tail” feature.
The kurtosis of Hurst exponents is not equal to 3, and the values of Jarque-Bera
statistics are significant at 1% significant level. This implies that, either for the
Chinese exchange market or for the Chinese stock market, we can reject the nor-
mality assumption of Hurst exponents estimated for return series. In addition,
Table 2.3 shows the same results to the time-varying Hurst exponents estimated for
the volatility series of the Chinese exchange market and stock market.
Table 2.2 Descriptive statistics for time varying Hurst exponent of returns
RER CSR
DFA R/S-AL R/S DFA R/S-AL R/S
Mean 0.623 0.583 0.655 0.576 0.559 0.631
Median 0.629 0.585 0.657 0.571 0.558 0.630
Max 0.654 0.620 0.694 0.621 0.590 0.664
Min 0.571 0.524 0.594 0.535 0.529 0.599
S.D. 0.018 0.019 0.020 0.024 0.012 0.013
Ske −1.122 −0.360 −0.356 0.262 0.182 0.183
Kurt 3.747 2.699 2.641 1.729 2.555 2.533
J-B 91.33*** 9.95*** 10.39*** 30.87*** 5.39* 5.75*
Notes J-B denote the Jarque-Bera statistic
*, ** and *** respectively denotes 10%, 5% and 1% significant level
Table 2.3 Descriptive statistics for time varying Hurst exponent of volatilities
V-RER V-CSR
DFA R/S-AL R/S DFA R/S-AL R/S
Mean 0.953 0.855 0.931 0.754 0.695 0.773
Median 0.970 0.861 0.937 0.766 0.708 0.787
Max 0.997 0.890 0.968 0.828 0.739 0.819
Min 0.868 0.801 0.875 0.682 0.632 0.707
S.D. 0.037 0.022 0.023 0.044 0.035 0.036
Ske −0.657 −1.044 −1.036 −0.106 −0.347 −0.350
Kur 1.979 3.114 3.115 1.626 1.493 1.502
J-B 45.20*** 71.43*** 70.29*** 31.57*** 44.96*** 44.64***
2.4 Relationship Between the Two Time-Varying Hurst

Exponent Series
We discuss the relationship between the time-varying Hurst exponents of the

Chinese exchange market and those of the Chinese stock market, using integration
method and Granger cause test (Granger 1980; Dickey and Fuller 1979). In this
section, the time-varying Hurst exponent is estimated by the DFA method, which is
based on two reasons. On the one hand, Weron (2002) suggests that DFA method is
better than R/S and R/S-AL methods by analyzing the simulated data. On the other
hand, our empirical results above also indicate that the Hurst exponents estimated
by DFA is usually a middle of those estimated by R/S-AL and R/S.
2.4.1 Unit Root Test
Table 2.4 presents the results of ADF (Augmented Dickey-Fuller) tests (Schwarz
1978) for time-varying Hurst exponents estimated for returns and volatilities of the
Chinese exchange market and stock market. Optimal lag length is automatic cal-
culated based on SIC (Schwarz Information Criterion) (Johansen 1988) with
maximum lag length 16.
As the result shown in Table 2.4, the level series of time-varying Hurst expo-
nents all have unit root and their 1st difference series do not have unit root at 1%
significant level. This indicates that these four types of time varying Hurst exponent
series are all I(1) process. Thereby, there maybe exist long-run equilibrium rela-
tionship between the time-varying Hurst exponents of the Chinese exchange market
and those of the Chinese stock market.
2.4 Relationship Between the Two Time-Varying Hurst Exponent Series 17
Table 2.4 ADF test for time-varying Hurst exponent series

Level series 1st difference series
t-statistics Lag length t-statistics Lag length
RER −2.6381 16(C, L) −3.7829*** 16(C, 0)
CSR −1.3026 16(C, L) −3.7902*** 15(C, 0)
V-RER −2.1171 10(C, L) −4.1830*** 9(C, 0)
V-CSR −3.2644* 13(C, L) −5.3270*** 12(C, 0)
Notes (C, L) denotes the constant and linear trend are included in the test equation, and (C, 0)
denotes that the constant is included in the test equation while the linear trend is not included in;
Null hypothesis is that series has a unit root
Table 2.5 Cointegration test for Hurst exponent series of returns and volatilities
Hypothesized Eigenvalue Trace 0.05 Max-Eigen 0.05
No. of CE(s) statistic Critical statistic Critical
value value
returns None 0.026 14.2217 18.3977 10.1837 17.1477
At most 1 0.0104 4.0380** 3.8415 4.0380** 3.8415
volatilities None 0.0326 17.4768 18.3977 12.7826 18.3977
At most 1 0.0121 4.6941** 3.8415 4.6941** 3.8415
Notes Lags interval (in first differences) is from 1 to 5; ** denotes rejection of the hypothesis at the
5% level; H_CSR and H_RER denote Hurst exponent of CSR and RER respectively. H_V-CSR
and H_V-RER denote Hurst exponent of volatility series of CSR and RER respectively
2.4.2 Cointegration Test
The results of Johansen cointegration test (Granger 1969) for time-varying Hurst
exponent series of returns and volatilities are shown in Table 2.5. Either for return
series or for volatility series, the statistics of trace test and max-eigenvalue test can
not reject the hypothesis “there is no cointegrating equation” and can not reject the
hypothesis “there exist at most one cointegrating equation” at the 5% significant
level. This indicates that there is no cointegrating equation between the time
varying Hurst exponent series of the Chinese exchange market and those of the
Chinese stock market, either for return series or for volatility series.
2.4.3 Granger Causality Test
Granger causality test can detect the short-term interaction among some series while
cointegration test can only assess whether the long-run relationship exists or not.
For four types of time varying Hurst exponent series are all non-stationary, we
Table 2.6 Granger causality test of Hurst exponents (Block exogeneity Wald test)
Hypothesis Chi-sq. df
H_RER does not Granger cause to H_CSR 21.964*** 2
H_CSR does not Granger cause to H_RER 15.007*** 2
H_V-RER does not Granger cause to H_V-CSR 17.891*** 2
H_V-CSR does not Granger cause to H_V-RER 27.628*** 2
Notes The lag length in VEC model is 2; Intercept (no trend) is in cointegrating equation. “df”
denotes degree of freedom
*** denotes rejection of the hypothesis at 1% level, ** denotes rejection of the hypothesis 5%
level, * denotes rejection of the hypothesis at 10% level
choose the block exogeneity Wald test based on VEC model as the Granger
causality test.
Table 2.6 shows the results of the Granger causality test. The Chi-squared
statistics are all significant at 1% significant level, indicating that there are bidi-
rectional Granger causality relationships, in short term, between the RMB exchange
market and stock market in China, either for return series or for volatility series.
This implies that the efficiency of the Chinese exchange market has significant
effect on the efficiency of the Chinese stock market, and vice versa.
2.5 Conclusions
We investigated the evolution of time-varying Hurst exponents of the Chinese

exchange market and stock market, using R/S, modified R/S (R/S-AL) and DFA
methods. The results show that Hurst exponent estimated by DFA is middle on
those estimated by R/S and by R/S-AL, while the Hurst exponent estimated by R/S
is usually the largest. This seems to confirm the simulation results with Weron
(2002). The distribution of the time-varying Hurst exponent estimated for the
Chinese financial markets is not equal to normal distribution, which is similar to the
results in Ref. (Kristoufek 2010). Moreover, the Hurst exponent estimated for
volatility series is usually larger than that estimated for return series, which means
that volatility series present longer memory than return series in the Chinese
exchange market and stock market. In addition, we also find that the breakpoint
occurring in the evolution of time varying Hurst exponent of the Chinese exchange
market is related to the reform of the Chinese RMB exchange rate flexibility on 22
June 2010. This indicates that time-varying Hurst exponent can detect the crash or
breakpoint in financial market, similar to the results of Wang et al. (2010, 2011).
Besides, we find that efficiency of China’s exchange market is not improved sig-
nificantly, although it was improved sharply in short term after the reform of the
RMB exchange rate flexibility in June 2010. At the same time, the efficiency of the
Chinese stock market is improved steadily, but there are little changes in recent
months. Therefore, to improve the efficiency of the Chinese exchange market and
2.5 Conclusions 19
stock market, the market-oriented level of them need be improved through

implementing the economic and financial reform.
In addition, we studied the relationship between the time-varying Hurst expo-
nents of the Chinese exchange market and those of the Chinese stock market, using
cointegration and Granger causality test. As the results shown, there is no long-run
equilibrium relationship, but exist bidirectional Granger causal relationship in short
term between the time-varying Hurst exponents of these two financial markets,
indicating that the efficiency of these two financial markets interact in short term but
not in long term. Thereby, when making policy, the government should be careful,
considering the influence among the financial markets.
References
E. Abounoori, M. Shahrazi, S. Rasekhi, An investigation of forex market efficiency based on

detrended fluctuation analysis: a case study for Iran. Phys. A 39, 3170–3179 (2012)
A.A. Anis, E.H. Lloyd, The expected value of the adjusted rescaled Hurst range of independent
normal summands. Biometrica 63, 111–116 (1976)
M. Ausloos, Statistical physics in foreign exchange currency and stock market. Phys. A 285, 48–
65 (2000)
O. Aydemir, E. Demirhan, The relationship between stock prices and exchange rates: evidence
from Turkey. Int. Res. J. Finance Econ. 23, 207–215 (2009)
D.O. Cajuerio, B.M. Tabak, Ranking efficiency for emerging equity marketsII. Chaos, Solitons
Fractals 23, 671–675 (2005)
D.A Dickey, W.A. Fuller, Distribution of the estimators for autoregressive time series with a unit
root. J. Am. stat. Assoc. 74(366a), 427–431 (1979)
Z. Eisler, J. Kertesz, Liquidity and the multiscaling properties of the volume traded on the stock
market. Europhys. Lett. 77 (2007)
M.R. Eldridge, C. Bernbarde, I. Mulvey, Evidence of chaos in the S&P 500 cash index. Adv.
Futures Options Res. 6, 179–192 (1993)
R.F. Engle, C.W.J. Granger, Co-integration and error correction: representation, estimation and
testing. Econometrica 55, 251–276 (1987)
M.A. Granero, J.E. Segovia, J. Perez, Some comments on Hurst exponent and the long memory
processes on capital markets. Phys. A 387, 5543–5551 (2008)
W.J. Granger, Clive, Investigating causal relations by econometric models and cross-spectral
methods. Econometrica: Journal of the Econometric Society. 424–438 (1969)
W.J. Granger, Clive, R. Joyeux, An introduction to long-memory time series models and fractional
differencing. Journal of time series analysis. 1(1), 15–29 (1980)
M.T. Greene, B.D. Fieltz, Long term dependence in common stock returns. J. Financ. Econ. 4,
249–339 (1997)
G.F. Gu, W.X. Zhou, Emergence of long memory in stock volatility from a modified Mike-Farmer
model. Europhys. Lett. 86 (2009)
H.E. Hurst, Long-term storage capacity of reservoirs. Trans. Amer. Soc. Civil Eng. 116, 770–808
(1951)
S. Johansen, Statistical analysis of cointegration vectors. J. Econ. Dyn. Control 12, 231–254
(1988)
L. Kristoufek, Rescaled range analysis and detrended fluctuation analysis: finite sample properties
and confidence interals. AUCo Czech Econ. Rev. 4, 315–329 (2010)
A.W. Lo, Long-term memory in stock market prices. Econometrica. 59, 1279–1313 (1991)
B. Mandelbrot, New methods in statistical economics. J. Polit. Econ. 71(5), 421–440 (1963)
B. Mandelbrot, J.R. Wallis, Robustness of the rescaled range and the measurement of long-run
statistical dependence. Water Resour. Res. 5, 967–988 (1969)
S.V. Muniandy, S.C. Lim, R. Murugan, Inhomogeneous scaling behaviors in Malaysian foreign
currency exchange rates. Phys. A 301, 407–428 (2001)
C. Nieh, H. Yau, The impact of Renminbi appreciation on stock prices in China. Emerg. Markets
Finance Trade 46, 16–26 (2010)
C-K. Peng, et al. Mosaic organization of DNA nucleotides. Phys. Rev. E. 49(2), 1685 (1994)
E.E. Peters, Chaos and Order in Capital Markets: A New View of Cycles, Prices and Market
Volatility (Wiley, 1991)
E. Peters, Fractal Market Analysis: Applying Chaos Theory to Investment and Economics (Wiley,
USA, 1994)
H.E. Roman, M. Porto, C. Dose, Skewness, long-time memory, and non-stationarity: application
to leverage effect in financial time serires. Europhys. Lett. 84 (2008)
G. Schwarz, Estimating the dimension of a model. Ann. Stat. 6, 461–464 (1978)
B.M. Tabak, D.O. Cajueiro, Assessing inefficiency in euro bilateral exchange rates. Phys. A 367,
319–327 (2006)
Y. Wang, L. Liu, R. Gu, Analysis of efficiency for Shenzhen stock market based on multifractal
detrended fluctuation analysis. Int. Rev. Financ. Anal. 18, 271–276 (2009)
Y. Wang, C. Wu, Z. Pan, Multifractal detrending moving average analysis on the US Dollar
exchange rates. Phys. A 390, 3512–3523 (2011)
R. Weron, Empirical Science of Financial Fluctuations (Springer-Verlag, Japan, 2002)
Chapter 3
Multifractal Detrended Fluctuation
Analysis (MF-DFA)
The study of financial or crude oil markets is largely based on current main stream
literature, whose fundamental assumption is that stock price (or returns) follows a
normal distribution and price behavior obeys ‘random-walk’ hypothesis (RWH),
which was first introduced by Bachelier (1900), since then it has been adopted as
the essence of many asset pricing models. However, some important results in
econophysics suggest that price (or returns) in financial or commodity markets have
fundamentally different properties that contradict or reject RWH. These ubiquitous
properties identified are: fat tails (Gopikrishnan 2001), long-term correlation
(Alvarez 2008), volatility clustering (Kim 2008), fractals multifractals (He et al.
2007), chaos (Adrangi 2001), etc. Nowadays, RWH has been widely criticized in
the finance and econophysics literature as this hypothesis fails to explain the market
phenomena.
After investigating the prices of cotton, wheat and so on, Mandelbrot provided
earliest empirical evidence that agricultural commodity spot prices do not obey
RWH by means of fractal geometry (Mandelbrot 1963, 1969). Since then, fractal
geometry has been widely applied in finance and market research domains. Peters
(1991, 1994) introduced fractal theory into the capital market research, and pro-
vided empirical evidence of the mono-fractal properties in many financial markets
by means of R/S analysis. As Mono-fractals can not describe the multi-scale and
subtle substructures of fractals in complex systems, many measures are applied to
investigate the multifractality, such as height–height correlation function Barabasi
et al. (1991), Multifractal Detrended Fluctuation Analysis (MF-DFA) (Kantelhardt
2002; Tabak and Cajueiro 2007; Alvarez 2008), the partition function method
(Telesca 2005; Kumar and Deo 2009; Yuan et al. 2009), etc.

22 3 Multifractal Detrended Fluctuation Analysis (MF-DFA)
3.1 Methodology
3.1.1 MF-DFA
Assume that there is a time series fyðiÞg, i ¼ 1; 2; . . .; N, where N is the length of

the series, then the MF-DFA method can be summarized as follow:
① Construct the profile
X
i
YðiÞ ¼ ðyðtÞ yÞ; i ¼ 1; 2; . . .; N; ð3:1Þ
t¼1
P
where y ¼ N1 Nt¼1 yðtÞ:
Divide the profile Y(i) into Ns ¼ ½N=s non-overlapping segments of equal
length s. Since the length N of the series is often not a multiple of the considered time
scale s, a short part at the end of the profile may remain unused. In order not to disregard
this part of the series, the same procedure is repeated starting from the opposite end.
Thereby, 2Ns segments are obtained altogether. And then calculate the local trends for
each of the 2Ns segments by kth order polynomial fit. Then the variance is given by
1X s
F 2 ðs,vÞ ¼ fY½ðv 1Þs þ i Y v ðiÞg2 ð3:2Þ
s i¼1
for each segment v, v ¼ 1; 2; . . .; N s , and
1X s
F 2 ðs; vÞ ¼ fY½N ðv Ns Þs þ i Y v ðiÞg2 ð3:3Þ
s i¼1
for each segment v, v ¼ Ns þ 1; . . .; 2Ns . Here, Y v ðiÞ denote the fitting polynomial
with order k in segment v.
② By averaging over all segments, the q-order fluctuation function is
obtained as follows
( )1=q
1 X2Ns
Fq ðsÞ ¼ ½F 2 ðs; vÞq=2 ð3:4Þ
2Ns v¼1
3.1 Methodology 23
for any q 6¼ 0, and

( )
1 X2Ns
F0 ðsÞ ¼ exp ln½F 2 ðs; vÞ ð3:5Þ
4Ns v¼1
If the series is power-law correlated, with the increasing of s, it should obey the
power-law:
Fq ðsÞ shðqÞ ð3:6Þ
Through the least-square fit, the slope of ln Fq ðsÞ and ln s is the generalized
Hurst exponent h(q), if h(q) is a constant which is independent on q, then the series
is monofractal; or it is multifractal. When q = 2, MF-DFA is turn to be DFA, and
h(2) is the well-known Hurst exponent.
If the series is stationary, the detrending procedure is not required. Thus the DFA
is replaced by the FA, and the variance can be simplified as
2
FFA ðs; vÞ ¼ ½YðvsÞ Yððv 1ÞsÞ2 ð3:7Þ
Through Eqs. (3.4) and (3.6), we can obtain

( )1=q
1 X2Ns
jYðvsÞ Yððv 1ÞsÞjq shðqÞ ð3:8Þ
2Ns v¼1
or it can be written as
X
2Ns
jYðvsÞ Yððv 1ÞsÞjq sqhðqÞ1 ð3:9Þ
v¼1
The partition function can be defined as
X
2Ns
Zq ðsÞ ¼ jYðvsÞ Yððv 1ÞsÞjq ssðqÞ ð3:10Þ
v¼1
Then we can obtain
sðqÞ ¼ qhðqÞ 1 ð3:11Þ

By Legendre Transformation, we can obtain the relationship between s(q) and

f(a):
dsðqÞ
a¼
dq ð3:12Þ
f ðaÞ ¼ qa sðqÞ
Through Eqs. (3.11) and (3.12), we can obtain
a ¼ hðqÞ þ qh0 ðqÞ

ð3:13Þ
f ðaÞ ¼ qða hðqÞÞ + 1
If there exist multifractal properties, the generalized Hurst exponent h(q) can be
fitted by the following function (Koscielny et al. 2006):
1 lnðaq þ bq Þ
hðqÞ ¼ ; ða [ bÞ ð3:14Þ
q q ln 2
where a and b stand for fitted parameters. Through Eq. (3.14) we can obtain h(q),
and then through Eq. (3.13) we can obtain multifractal spectrum f(a). In multi-
fractal models, the strength of multifractality can be described by the width of
spectrum Da. Through Eqs. (3.13) and (3.14), we can obtain:
lnða=bÞ
Da ¼ amax amin ¼ hð1Þ hð þ 1Þ ¼ ð3:15Þ
ln 2
Thus the Da can be estimated by the parameters a, b.
3.1.2 Partition Function
To keep our description as self-contained as possible, let us review briefly the

models (Du and Ning 2008; Jiang and Zhou 2008):
Let us suppose T(i) (i = 1, 2, …, L) to be the time series of crude oil prices,
where L is the length of the series. Let us define its t-returns
rðiÞ ¼ logðTði þ tÞ=TðiÞÞ; ð1 t 250Þ ð3:16Þ
Then we divide it into N parts with equal size s. Let us assume the length of the
whole time series as unit length 1, then each part is dðd ¼ 1=NÞ. For the jth part, we
define the mass probability Pj ðdÞ
3.1 Methodology 25
Ps
rððj 1Þs þ iÞ
Pj ðdÞ ¼ i¼1
PL ð3:17Þ
i¼1 rðiÞ
If the analyzed time series develops multifractal properties, its mass probability
reveals power-law scaling:
Pj ðdÞ da ; ðd ! 0Þ ð3:18Þ
If a remains the same values, the series is mono-fractal; otherwise let Na ðdÞ be
the number of areas covered with same a, which satisfies
Na ðdÞ df ðaÞ ; ðd ! 0Þ ð3:19Þ
where a is called the singular exponent, and f ðaÞ is the singular spectrum, a.k.a. the
multifractal spectrum. As it is difficult to obtain the multifractal spectrum directly, it
is usually calculated by means of partition function, which defined as:
X
N
Sq ðdÞ ¼ Pi ðdÞq ð3:20Þ
i¼1
where q can range from −∞ to +∞ theoretically. When q is positive, the partition

function Sq ðdÞ is mainly influenced by the large value of P, which can reflect the
large price fluctuations; when q is negative, the partition function is mainly influ-
enced by the small value of P, which can reflect the small fluctuations. If the time
series is the multifractal, the partition function reveals the power-law scaling in a
small scale d, given by
Sq ðdÞ dsðqÞ ; ðd ! 0Þ ð3:21Þ
Or be rewritten as:
ln Sq ðdÞ
sðqÞ ¼ lim ð3:22Þ
d!0 ln d
According to Eq. (3.21), we can calculate the s(q) through the slope of
ln Sq ðdÞ ln d curve. If s(q) associated with a linear plot in q, the time series is
mono-fractal; otherwise the series is multifractal. By Legendre Transformation, the
relationship between f(a) and s(q) is given by
dsðqÞ
a¼ ð3:23Þ
dq
Therefore, another way to characterize a multifractal series is the singularity

spectrum f(a) defined by
f ðaÞ ¼ aq sðqÞ ð3:24Þ
Through Eqs. (3.20), (3.22) and (3.23), we can obtain

PN
Pi ðdÞq ln P ki
a ¼ lim Pi¼1
N q ¼ lim ð3:25Þ
i¼1 Pi ðdÞ ln d
d!0 d!0 ln d
According to Eq. (3.25), if there exists a linear relationship between ki and lnd,
we can estimate a by the least square method; then, we can obtain the multifractal
spectrum f(a) according to Eq. (3.24).
3.2 Empirical Analysis on Developed-Emerging

Agricultural Futures Markets
We chose Wheat, Soy Meal, Soy Bean and Corn futures contracts from US and
China’s agricultural markets as the representatives of the emerging and developed
markets, and applied the famous MF-DFA to study the multifractal properties. We
found that there are multifractality in the two markets; though shuffling the fluc-
tuations, we investigated the dynamical resources of multifractality; and at last we
applied the average of s(q) to obtain the multifractal spectrum of whole markets,
and compared the multifractal strength of the two countries, or the multifractal
strength of emerging and developed agricultural futures markets.
3.2.1 Data
Let us suppose T(i) (i = 1, 2, …, L) to be the time series of crude oil prices, where
L is the length of the series. Let us first define the fluctuation as
f ðiÞ ¼ jlnðTði þ DtÞ=TðiÞÞj; ði L DtÞ ð3:26Þ
In our case, because our data are daily records, namely, Dt ¼ 1.

We used the daily closing prices of Hard Winter wheat futures market from
December 28th, 1993 to May 11th, 2009 (L = 3106) market from China’s
Zhengzhou Commodity Exchange, and Soy Meal futures from July 17th, 2000 to
May 11th, 2009 (L = 2108), No. 1 Soy Bean futures from March 15th, 2002 to May
3.2 Empirical Analysis on Developed-Emerging Agricultural Futures Markets 27
Table 3.1 The summary statistics of wheat, soy meal, soybean and corn
Mean Std. Dev. Skewness Kurtosis Jarque-Bera
China
Wheat 1466.5 221.06 0.0472 2.1911 85.841*
Soy meal 2395.8 599.40 0.8386 3.2724 253.58*
Soybean 3119.2 762.98 1.1765 3.7002 432.94*
Corn 1494.8 211.28 −0.0712 1.7408 75.272*
USA
Wheat 376.32 115.73 2.6349 12.679 46,476*
Soy meal 191.00 48.480 1.5067 6.0921 7113.5*
Soybean 658.47 160.83 2.0803 9.8035 24,332*
Corn 275.17 72.509 2.3703 12.419 42,517*
Note *Means reject the null hypothesis that the sample comes from a normal distribution at the
significance of 0.01
11th, 2009 (L = 1724), Corn futures from September 22nd, 2004 to May 11th,
2009 (L = 1125) market from China’s Dalian Commodity Exchange. To compare
the difference, we also chose the daily closing prices of Wheat futures (L = 9184),
Soy Meal futures (L = 9158), Soy Bean futures (L = 9182) and Corn futures
(L = 9177) from January, 2nd to May 11th, 2009 from CBOT. All our data are
taken from Reuter© database. In our discussion, the size s ranges from 12 to [N/6]
with the computation step 10; the degree of polynomial m = 1, 2, 3; q ranges
from −60 to 60 with the step 0.1.
To get a better understanding of our data sets, we provided summary statistics of
the four futures (see Table 3.1), from which we can see clear departure from a
normal distribution.
3.2.2 Multifractal Spectrum Analysis
By means of the above-mentioned model, we first of all obtained lnFq(s) versus

lns relationships of China’s and US agricultural futures markets (see Figs. 3.1 and
3.2). From the figures we can find the relationships are linear, which implies that
there exist power-law relationships. Then we obtained the relationships between
q and s(q) (see Fig. 3.3), and find that the relationships between s(q) and q are
nonlinear. It is also obvious that h(q) is nonlinearly dependent on q, and decreases
while q increases (see Fig. 3.4). Figure 3.5 presents us the nontrivial multifractal
spectra. All these pieces of empirical evidence imply that multifractality properties
can be found both in China’s and US agricultural futures markets. Especially, when
q = 2, all the Hurst exponents of China’s agricultural futures markets are greater
than 0.5 (see Table 3.2), for example, h(2) = 0.7612 ± 0.0293 for China’s wheat,
h(2) = 0.8204 ± 0.1386 for US wheat, which imply that all the markets do not
(a) (b)
-1 -2
-2 -3
-3
-4
-4
lnFq(s)
lnFq(s)
-5 -5
-6 -6 q=-60
-7
-7 q=-10
-8
-9 -8 q=-5
2 3 4 5 6 7 2 3 4 5 6
lns lns q=0

(c) (d)
-2 -3 q=5
-3 -4 q=10
-4 q=60
lnFq(s)
lnFq(s)
-5
-5
-6
-6
-7 -7
-8 -8
2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5
lns lns
Fig. 3.1 Shows the relationship between lns and lnFq(s) in China’s agricultural futures market
when m = 3. Panels a–d illustrate the relationships between lnFq(s) and lns in wheat, soy meal,
soy bean and corn futures markets respectively
obey random-walk and show persistent properties. We can also find that the Hurst
exponents of US agricultural futures markets are greater than China’s, which
suggests than US agricultural futures markets are more persistent than China’s.
3.2.3 Sources of Multifractality
In current literature, two major sources of multifractality are widely acknowledged

which can be found in various time series. One is long-range temporal correlation
for small and large fluctuations, the other is fat-tailed probability distribution of
increments (Matia et al. 2003; Kwapien et al. 2005). Usually, two procedures can be
applied to indentify the contributions of two sources and to indicate the multi-
fractality strength, that is, shuffling procedure, and phase randomization (Matia
et al. 2003). To identify which causes constitute the major contributions of mul-
tifractality in the markets, we shuffled the original time series to destroy any
temporal correlations, aka, long-range or short-range memories in the markets,
while the shuffled data still remain the exactly the same fluctuation distributions
without any memory. If the generalized Hurst exponent h(q) does not change after
(a) (b)
-1 0
-1
-2
-2
-3
-3
lnFq(s)
lnFq(s)
-4
-4
-5
-5
-6 -6 q=-60
-7 -7 q=-10
-8 -8
2 3 4 5 6 7 8 2 3 4 5 6 7 8 q=-5
q=0
lns (d) lns
(c) -1 q=5
0
-1 -2 q=10
-2 -3
q=60
lnFq(s)
lnFq(s)
-3
-4
-4
-5
-5
-6 -6
-7 -7
-8 -8
2 3 4 5 6 7 8 2 3 4 5 6 7 8
lns lns
Fig. 3.2 Illustrates the relationship between lns and lnFq(s) in US agricultural futures markets
when m = 3. Panels a–d illustrate the relationships between lnFq(s) and lns in wheat, soy meal,
soy bean and corn futures markets respectively
we shuffled the series, it implies that there is no influence by nonlinear temporal

correlation, and the multifractality comes from fat-tailed distribution; if the gen-
eralized Hurst exponent h(q) becomes a constant 0.5, the series is no longer mul-
tifractal, which means that the multifractality comes mainly from nonlinear
temporal correlation; if it is still multifractal but the strength of multifractality
becomes weaker, both the causes have contributions in the multifractality forma-
tion. The shuffling procedure consists of the following steps (Matia et al. 2003):
Step 1: we generate pairs (m, n) of random integer numbers, which satisfies m,
n N, where N is the length of the time series to be shuffled;
Step 2: we interchange entries m and n of the time series;
Step 3: we repeat the first and second steps for 20N times. It is critical to ensure
that ordering of entries in the time series is fully shuffled, thus the
long-range or short-range memories, if any, will be destroyed. The
shuffling is repeated with different random seeds to avoid the systematic
errors caused by random number generators.
From the results for the shuffled cases in Table 3.2, we can find that all the
spectra widths become significantly narrower except that those of soy meal markets,
that is, Da from 0.4531 ± 0.0249 (original) to 0.4665 ± 0.0555 (shuffled) for
(a) (b)
40 40
20 20
0 0
(q)
(q)
-20 -20
-40 -40
China China
USA -60 USA
-60
-80 -80
-80 -60 -40 -20 0 20 40 60 80 -80 -60 -40 -20 0 20 40 60 80
q q
(c) (d)
40 40
20 20
0 0
(q)
(q)
-20 -20
-40 -40
USA China
-60 America -60 USA
-80 -80
-80 -60 -40 -20 0 20 40 60 80 -80 -60 -40 -20 0 20 40 60 80
q q
Fig. 3.3 Depicts q * s(q) curve when m = 3, where q ranges from −60 to 60 with the step 3,
which we can find nonlinear relationships between s(q) and q in both China’s and US agricultural
futures markets. Panels a–d illustrate the relationships between s(q) and q in wheat, soy meal, soy
bean and corn futures markets respectively
China’s soy meal market, and from 0.4857 ± 0.0707 (original) to 0.4169 ± 0.0670
(shuffled) for US soy meal market. Thereby, nonlinear temporal correlation con-
stitutes the major contributions in the multifractality formation in all markets except
soy meal.
Especially, as for the shuffled fluctuations, when q = 2, all the Hurst exponents
of China’s and US agricultural futures markets are around 0.5 (see Table 3.2), such
as h(2) = 0.4950 ± 0.0062 (China’s wheat) and 0.5121 ± 0.0082 (US wheat), h
(2) = 0.5119 ± 0.0234 (China’s soy meal) and 0.4727 ± 0.0134 (US soy meal),
h(2) = 0.4978 ± 0.0130 (China’s soybean) and 0.4864 ± 0.0161 (US soybean), h
(2) = 0.5223 ± 0.0314 (China’s corn) and 0.4745 ± 0.0105 (US corn). These
results clearly prove that the shuffled series obey random walk (Peters 1991, 1994).
Compared with the results for original cases, we can find that except for soy meal
markets, after the shuffling procedure successfully destroyed the nonlinear temporal
correlation in the original series, the multifractality also becomes significantly
weaker or even vanishes while at the same time the distributions remain exactly
(a) (b)
1.4 1.1
1.2 China 1.0

China
USA .9 USA
1.0
.8
h(q)
h(q)
.8
.7
.6
.6
.4 .5
(b)
.2 .4
-80 -60 -40 -20 0 20 40 60 80 -80 -60 -40 -20 0 20 40 60 80
q q
(c) (d)
1.4 1.4
1.2
China 1.2
China
1.0 USA
USA
1.0
.8
h(q)
h(q)
.6 .8
.4
.6
.2 (d)
0.0 .4
-80 -60 -40 -20 0 20 40 60 80 -80 -60 -40 -20 0 20 40 60 80
q q
Fig. 3.4 Shows the q * h(q) relationships when m = 3, where q ranges from −60 to 60 with the
step 1. Panels a–d illustrate the relationships between h(q) and q in wheat, soy meal, soy bean and
corn futures markets respectively
unchanged; thereby, the fat-tailed probability distribution is by no means the

plausible explanation for market multifractality formation, no matter in China’s and
US agricultural futures markets. As for soy meal markets, there might be other
unknown factors which determine the multifractality formation in these markets.
3.2.4 Comparative Analysis
Although we shed light on the plausible cause for multifractality formation in those
markets, the following question is still waiting to be answered: Are the multifractal
strengths in those markets of the transition and emerging economies weaker (or
stronger) than those of the developed ones?
In order to compare China’s and US agricultural futures markets as a whole, we
applied the average of s(q) proposed in Matia et al. (2003):
(a) (b)
1.2 1.2 China
China USA
1.0 1.0
USA
.8 .8
.6 .6
f (α)
f (α)
.4 .4
.2 .2
0.0 0.0
-.2
.2 .4 .6 .8 1.0 1.2 1.4 .4 .5 .6 .7 .8 .9 1.0 1.1
(c) (d)
1.2 1.2 China
China
1.0 USA
USA 1.0
.8 .8
.6 f (α)
f (α)
.6
.4 .4
.2 .2
0.0 0.0
0.0 .2 .4 .6 .8 1.0 1.2 1.4 .2 .4 .6 .8 1.0 1.2 1.4
Fig. 3.5 Illustrates the relationships between a and f(a), where q ranges from −60 to 60 with the
step 0.5. Panels a–d illustrate the relationships between f(a) and a in wheat, soy meal, soy bean
and corn futures markets respectively
1X 4
sav ðqÞ ¼ si ðqÞ ð3:27Þ
N i¼1
Therefore, we can obtain

P4
dsav ðqÞ 1 d i¼1 si ðqÞ 1X 4
aav ¼ ¼ ¼ ai ðqÞ ð3:28Þ
dq N dq N i¼1
Then we can obtain fav(a) by means of Eqs. 3.23 and 3.24, thereby we can get
the multifractal spectra of China’s and US markets as two whole markets (see
Figs. 3.3, 3.4, 3.5 and 3.6). We estimate widths of spectra by Daav aav (−60)
−aav (60). The numerical results in Table 3.3 tell us that in general the spectrum
widths of US markets as a whole are narrower than those of China’s ones with
different orders of polynomial m. The nontrivial finding presents us a further
hypothesis: the multifractal strengths in the developed economies may be weaker
than those in the transition or emerging ones. At least, the answer is positive for the
special and representative cases of China’s and US agricultural futures markets. Of
course, many other efforts and empirical or theoretical results from other peers may
Table 3.2 Generalized Hurst exponents and the width of multifractal spectrum
Original Shuffled
Wheat Soy Soybean Corn Wheat Soy Soybean Corn
meal meal
m=1
China h(2) 0.7905 0.7804 0.8055 0.7449 0.5012 0.5353 0.4910 0.5537
h(−60) 1.3733 0.9858 1.1859 1.0297 0.7075 0.7920 0.7195 0.8011
h(60) 0.4241 0.5699 0.5465 0.4414 0.2244 0.4073 0.1070 0.3662
⊿a 0.9744 0.4500 0.6759 0.6224 0.5224 0.4197 0.6426 0.4737
USA h(2) 0.9590 0.8858 0.8426 0.8558 0.5057 0.4861 0.5025 0.4604
h(−60) 1.0874 1.1123 1.1017 1.0379 0.6283 0.6647 0.6355 0.6550
h(60) 0.8257 0.7164 0.5994 0.7122 0.4016 0.4000 0.2871 0.2507
⊿a 0.2891 0.4336 0.5375 0.3539 0.2604 0.2974 0.3812 0.4383
m=2
China h(2) 0.7683 0.7206 0.6807 0.6987 0.4895 0.5079 0.4916 0.5250
h(−60) 1.3754 0.9580 1.1133 1.0227 0.8025 0.7697 0.6483 0.8024
h(60) 0.4692 0.5642 0.3240 0.4593 0.1969 0.3405 0.1153 0.3747
Da 0.9318 0.4313 0.8171 0.6021 0.6364 0.4577 0.5623 0.4578
USA h(2) 0.7705 0.8086 0.8564 0.7555 0.5203 0.4608 0.4818 0.4782
h(−60) 1.0190 1.0459 1.1728 1.1253 0.6692 0.7167 0.6380 0.6301
h(60) 0.5189 0.5258 0.5736 0.4059 0.3867 0.2625 0.2542 0.2778
Da 0.5332 0.5564 0.6292 0.7436 0.3163 0.4839 0.4168 0.3887
m=3
China h(2) 0.7249 0.6748 0.6383 0.6803 0.4943 0.4925 0.5108 0.4883
h(−60) 1.3220 0.9223 1.0604 1.0943 0.7746 0.7922 0.6636 0.8676
h(60) 0.4102 0.4759 0.2389 0.4475 0.1589 0.2920 0.1531 0.3550
Da 0.9406 0.4780 0.8461 0.6807 0.6470 0.5220 0.5425 0.5472
USA h(2) 0.7316 0.7929 0.8195 0.7921 0.5102 0.4712 0.4749 0.4850
h(−60) 0.9938 0.9966 1.1802 1.1804 0.6780 0.7202 0.6370 0.6112
h(60) 0.4638 0.5644 0.5412 0.4512 0.3382 0.2810 0.2541 0.2893
Da 0.5578 0.4672 0.6692 0.7559 0.3710 0.4695 0.4164 0.3594
be called for in this field to accept or reject this hypothesis. No matter the final
answer is positive or negative; the findings on this issue will definitely offer us
better understandings on the dynamics of markets.
(a) (b)
1.2 1.2 China
China
USA
1.0 USA 1.0
.8 .8
f ( α)
f ( α)
.6 .6
.4 .4
.2 .2
0.0 0.0
m =1 (b) m=2
.4 .5 .6 .7 .8 .9 1.0 1.1 1.2 .4 .5 .6 .7 .8 .9 1.0 1.1 1.2
α α
(c) 1.2
China
1.0 USA
.8
f (α)
.6
.4
.2
0.0
m=3
.2 .4 .6 .8 1.0 1.2
Fig. 3.6 Shows the aav fav ðaÞ curve, where q ranges from −60 to 60 with the step by 0.1 with
different orders of polynomial k
Table 3.3 The estimated width of multifractal spectrum Daav

The order of polynomial m China America
m=1 0.6815 0.4047
m=2 0.6965 0.6180
m=3 0.7399 0.6159
Table 3.4 The descriptive statistics of Brent & WTI crude oil prices
Mean Std. Dev. Skewness Kurtosis Jarque-Bera
Brent 31.585 23.101 1.9896 7.0143 7584.0*
WTI 32.003 22.875 2.0830 7.5004 9405.4*
*
Note Means reject the null hypothesis that the sample comes from a normal distribution at the
significance of 0.01
3.2.5 Conclusions
In this section, we investigated the multifractal properties in China’s and US

agricultural futures markets from the comparative perspective. Our nontrivial
empirical findings can be summarized as follows:
First of all, multifractality is found in all those markets.
Secondly, nonlinear temporal correlation constitutes the major contribution in
multifractality formation instead of fat-tailed distribution.
Thirdly, the width of multifractal spectrum of US agricultural futures markets as
a whole is narrower than that of China’s.
Of course, there are still some questions waiting to be answered in our future
works: are there any more causes of multifractality in the analyzed markets along
with nonlinear temporal correlation and fat-tailed distribution, especially for the
case of soy meal markets? Is our hypothesis valid for other cases? Many other
pieces of further empirical evidence and theoretical proofs are needed from other
commodity or financial markets in more emerging or transition economies.
3.3 Empirical Analysis on Crude Oil Markets
In this section, we firstly applying MF-SSA as well as MF-DFA methods, we

provided and confirmed the mutually supporting empirical evidence of the exis-
tence of multifractality in crude oil markets; secondly, we proposed the explanation
of the dynamical formation of multifractality in the markets; thirdly, to get deeper
insights into the underlying price behaviors in crude oil markets, we furthered our
study to investigate the nonlinear dynamical mechanisms, especially of the price
fluctuations at different scales.
3.3.1 Data
We used the daily spot FOB prices (dollars per barrel) of Brent and WTI from May
20, 1987 to Oct. 14, 2009 including 5656 (Brent) and 5754 (WTI) daily records
respectively (data source: databases of U.S. Energy Information Administration,
(http://tonto.eia.doe.gov/dnav/pet/xls/PET_PRI_SPT_S1_D.xls). To better describe
our data in terms of business, we defined the characteristic delay tk = 1, 5, 21, 63,
125, 250, which represents returns of business day, week, month, season, half-year,
year respectively. The size s ranges from 10 to [L/2] with the computation interval
10; the q in the partition function ranges from −60 to 60 with the interval 1.
To get a better understanding of our data sets, we provided summary statistics of
the four futures (see Table 3.4), from which we can see clear departure from a
normal distribution.
(a) (b)
600 600
400 400
200 200
ln S q ( δ )
ln S q ( δ )
0 0
-200 -200
-400 -400
-7 -6 -5 -4 -3 -2 -1 0 -7 -6 -5 -4 -3 -2 -1 0
lnδ lnδ
Fig. 3.7 Illustrates the relationships between lnSq(d) and lnd at the characteristic delay tk = 1,
where q ranges from −60 to 60 from up to bottom with the computation step 10. Panels a and
b depict the relationships between lnSq(d) and lnd for Brent and WTI crude oil markets
respectively
3.3.2 Multifractality and Its Dynamical Formation

Mechanisms
We investigated the relationship bewteen lnSq(d) and lnd with characteristic delay
tk = 1 (see Fig. 3.7). According to Fig. 3.7, we can know that lnSq(d) exhibits a
linear relation with lnd, which implys Sq(d) and d obey the power law (Eq. 3.21),
there are the similar results when tk = 1, 5, 21, 63, 125, 250, which is the evidence
of fractals.
In order to investigate the multifractal features in Brent and WTI crude oil
markets, we obtained the relationships between s(q) and q (see Fig. 3.8), and found
that the relationships are nonlinear at different characteristic delays tk, which implies
(a) (b)
60 60
(b)
40 40
20
20
0
0
-20 tk=1 tk=1
τ(q )
τ(q )
-20
-40 tk=5 tk=5
tk=21 -40 tk=21
-60
tk=63 -60 tk=63
-80
tk=125 tk=125
-100 -80
tk=250 tk=250
-120 -100
-6 0 -4 0 -2 0 0 20 40 60 -60 -40 -20 0 20 40 60
q q
Fig. 3.8 Depicts nonlinear relationships between s(q) and q at different characteristic delays
(tk = 1, 5, 21, 63, 125, 250) for a Brent and b WTI original price series
3.3 Empirical Analysis on Crude Oil Markets 37
(a) (b)
(c) (d)
Fig. 3.9 Depicts empirical results of f(a) as a function of a at different characteristic delays
(tk = 1, 5, 21, 63, 125, 250) for the original and shuffled series. Panels a–d plot the results of Brent
(original), WTI (original), Brent (shuffled) and WTI (shuffled) respectively. It is obvious that the
spectra for the original series are wider than those of shuffled cases
that both the two markets reveal multifractality characteristics. Meanwhile, we

obtained the multifractal spectra for the two markets (see Fig. 3.9). According to
Eq. (3.12), we can obtain df ðaÞ=da ¼ q. We thereby divided the multifractal
spectrum into left- and right-half part by means of maximum extreme values of f(a);
the left-half corresponds to the section where q > 0, which reflects the price
behaviors influenced by the large price fluctuations; while the right one corresponds
to the section where q < 0, which describes the impacts of the small price fluctu-
ations on the price behaviors. In order to describe the strength of multifractality, we
define the width of the multifractal spectrum by

ln Pmin ln Pmax lnðPmin = ln Pmax Þ
Da ¼ amax amin ¼ lim ¼ lim ð3:29Þ
d!0 ln d ln d d!0 ln d
Table 3.5 The width of Characteristic Original series Shuffled series

multifractal spectrum Da time delays WTI Brent WTI Brent
tk = 1 0.4393 0.3710 0.1422 0.1590
tk = 5 0.5175 0.4221 0.1332 0.1648
tk = 21 0.5558 0.4929 0.1555 0.1345
tk = 63 0.8259 0.7628 0.1283 0.1353
tk = 125 0.7804 0.7895 0.1441 0.1184
tk = 250 0.8435 0.8895 0.1051 0.1338
So that Da reflects the difference between minimum probability of large fluc-

tuations and the maximum probability of small fluctuations.
From the above-mentioned results, we found that the Brent and WTI crude oil
markets reveal distinct nontrivial multifractal features. However, it is still necessary
to identify which underlying sources or mechanisms constitute the markets’ mul-
tifractality. Thus, we shuffled the original series to destroy the nonlinear temporal
correlation and market memory, while the distribution of the fluctuations remains
exactly the same. Therefore, if no multifractal feature can be found after we shuffled
the original multifractal series, the nonlinear temporal correlation contributes the
major part in the formation of multifractality in the markets; otherwise, if multi-
fractality remains exactly the same strength, that is, the width of multifractal spectra
of the shuffled series remain the same with the original ones, the non-Gaussian
distribution is the major reason for the underlying mechanism of multifractality;
furthermore, if multifractality remains but the strength is significantly weaken,
namely, the width becomes narrow, both of the reasons have nontrivial impacts on
the formation of multifractality in the markets.
Applying the method mentioned above to shuffled series, we obtained the
relationships between s(q) and q (see Fig. 3.8), multifractal spectra (see Fig. 3.9).
Then we calculated the width of spectra for both original and shuffled series at
different characteristic time delays (see Table 3.5).
3.3.3 Multifractal Detrended Fluctuation Analysis
By means of the above-mentioned model, we first of all obtained lnFq(s) versus

lns relationships of Brent & WTI crude oil markets (see Figs. 3.10 and 3.11). From
the figures we can find the relationships are linear, which implies that there exist
power-law relationships, which imply that the two markets display fractal features.
To further investigate the existence of multifractality in the markets, we obtained
the relationships between q and s(q) (see Fig. 3.12), and find that the relationships
between s(q) and q are nonlinear. Thus, we obtained another piece of evidence of
multifractality.
(a) (b)
(a) (b)
Fig. 3.10 The relationship between lns and lnFq(s) in WTI market when m = 1, 2, 3, 4
respectively
It is also obvious that h(q) is nonlinearly dependent on q, and decreases while

q increases (see Fig. 3.13). Figure 3.14 presents us the nontrivial multifractal
spectra, which are the other pieces of empirical evidence for the multifractality
properties in WTI & Brent markets. Especially, when q = 2, all the Hurst exponents
of the markets are greater than 0.5 for all different orders (see Table 3.5), which
imply that the markets do not obey random-walk and show persistent properties.
We can also find that the Hurst exponents of WTI are greater than those of Brent,
which suggests that the former are more persistent than the latter.
From the above-mentioned discussions, we found that the results from both
methods positively support that there exist fractals and multifractals in both Brent
and WTI crude oil markets.
3.3.4 Sources of Multifractality
In general, there are two major sources of multifractality which can be found in
various time series. One is nonlinear temporal correlation for small and large
(a) (b)
(a) (b)
Fig. 3.11 The relationship between lns and lnFq(s) in Brent market when m = 1, 2, 3, 4
respectively
(a) (b)
Fig. 3.12 The q and s(q) relationships, where q ranges from −60 to 60 with the step 3, which we
can find nonlinear relationships between s(q) and q in both markets. Panels a and b illustrate the
relationships between s(q) and q in WTI & Brent markets respectively
(a) (b)
Fig. 3.13 The q * h(q) relationships when m = 3, where q ranges from −60 to 60 with the step
1. Panels a and b depict the results for Brent & WTI fluctuations
(a) (b)
Fig. 3.14 The relationships between a and f(a), where q ranges from −60 to 60 with the step 0.5.
Panels a and b depict the results for Brent & WTI fluctuations
fluctuations; the other is fat-tailed probability distribution of increments

(Norouzzadeh and Rahmani 2006; Kwapien et al. 2005), while both of the causes
need scaling exponents to measure the small and large price fluctuations. Usually,
two procedures can be applied to indentify the contributions of two sources and to
indicate the multifractality strength, that is, shuffling procedure, and phase random-
ization (Norouzzadeh and Rahmani 2006). To identify which causes constitute the
major contributions of multifractality, we shuffled the original time series to destroy
any temporal correlations, aka, long-range or short-range memories in the markets,
while the shuffled data still remain the exactly the same fluctuation distributions
without any memory.
To identify which causes constitute the major contributions of multifractality in

the markets, we shuffled the original time series to destroy any temporal correla-
tions, aka, long-range or short-range memories in the markets, while the shuffled
data still remain the exactly the same fluctuation distributions without any memory.
If the generalized Hurst exponent h(q) does not change after we shuffled the series,
it implies that there is no influence by nonlinear temporal correlation, and the
multifractality comes from non-Gaussian distribution; if the generalized Hurst
exponent h(q) becomes a constant 0.5, the series is no longer multifractal, which
means that the multifractality comes mainly from nonlinear temporal correlation;
if it is still multifractal but the strength of multifractality becomes weaker, both the
causes have contributions in the multifractality formation.
The shuffling procedure consists of the following steps (Matia et al. 2003):
firstly, we generate pairs (m, n) of random integer numbers, which satisfies m,
n N, where N is the length of the time series to be shuffled; then, we interchange
entries m and n of the time series. To ensure that ordering of entries in the time
series is fully shuffled, we repeat the first and second steps for 20N times so that
long-range or short-range memories, if any, will be destroyed. The shuffling is
repeated with different random seeds to avoid the systematic errors caused by
random number generators.
According to our discussion, we can find in Table 3.5 that after we shuffled the
original series, the width of spectra become significantly narrow compared with the
results for the original ones; for example, the width Da fall into the interval [0.1051,
0.1555] (shuffled WTI series), which is significantly narrowed compared with the
interval [0.4393, 0.8435] (original WTI series). As we know, for monofractal time
series, the spectrum for the series is one point, i.e., the width Da ¼ 0; otherwise, the
time series analyzed is multifractal characterized with a nontrivial positive width
ðDa [ 0Þ. Therefore, the multifractality persists after we shuffled the original series
because the spectrum widths of shuffled series do not approximate to zero and are
nontrivially positive in both the markets. We can summarize that the major cause in
the markets is due to long range correlation; while since the widths of spectra are
significantly narrowed but the multifractality persists, the formation of multifrac-
tality is also influenced by non-Gaussian distribution.
From Fig. 3.15 and Table 3.6, we found that the Hurst exponents for the shuffled
fluctuations are around 0.5, which imply that the shuffling procedure effectively
destroyed the correlation and the shuffled series are close to random walk motions.
In Table 3.6, the generalized Hurst exponents h(q) (q = 2, −60, +60) for original
Brent & WTI fluctuations do change after we shuffled the series, it implies that the
nonlinear temporal correlation do have major contributions in the multifractality
formation; furthermore, the generalized Hurst exponents h(q) for shuffled cases
approximate (but not equal) to 0.5, that is, the shuffled series are close to random
walk process but remains weak multifractal. Compared with the results for original
cases, we can find that after the shuffling procedure successfully destroyed the
nonlinear temporal correlation in the original series, the multifractality also
becomes significantly weaker while at the same time the distributions remain
exactly unchanged; thereby, the non-Gaussian probability distribution is by no
1.0
WTI
.9 Brent
Shuffled WTI
Shuffled Brent
.8
h (2)
.7
.6
.5
.4
1 2 3 4 5 6 7 8
m
Fig. 3.15 Hurst exponents for original and shuffled fluctuations. It is noted that all the exponents
for shuffled cases are around 0.5, which imply that the shuffled series are close to random walk
motions
means the primary plausible explanation for market multifractality formation, no

matter in Brent or WTI market, while the nonlinear temporal correlation constitutes
the major contributions in the markets’ multifractality formation.
Interestingly, our results from both methods again consistently suggest that the
dynamical cause of the existence of market multifractality derives mainly from
nonlinear temporal correlation instead of non-Gaussian probability distribution.
Furthermore, from the results for the shuffled cases in Table 3.6, we can also find
that the spectra widths of WTI are narrower than those of Brent except when m = 1,
2. Let us average over all the spectra widths under different m, we obtained the
average width 0.462825 (WTI) and 0.4961375 (Brent), which may suggest that
multiscaling microstructure of price fluctuations in Brent market may be more
complex than that in WTI.
Table 3.6 Generalized Hurst exponents and the width of multifractal spectrum
The order of polynomial m Original Shuffled
h(2) h(−60) h(60) Da h(2) h(−60) h(60)
m=1
WTI 0.9665 1.1147 0.7111 0.4373 0.4777 0.6585 0.2716
Brent 0.9269 0.9372 0.6294 0.3425 0.4896 0.6563 0.1924
m=2
WTI 0.9619 1.0386 0.7014 0.3715 0.5003 0.6589 0.2458
Brent 0.8980 0.9185 0.6320 0.3202 0.4861 0.6811 0.1960
m=3
WTI 0.9083 0.9841 0.6583 0.3623 0.5070 0.6817 0.1860
Brent 0.8622 0.9729 0.6195 0.3871 0.4921 0.6634 0.2072
m=4
WTI 0.8332 0.9345 0.5739 0.3963 0.5174 0.6945 0.1953
Brent 0.8014 0.9720 0.5407 0.4649 0.5061 0.6792 0.2326
m=5
WTI 0.7736 0.8962 0.5054 0.4267 0.5262 0.7223 0.1916
Brent 0.7506 0.9603 0.4700 0.5236 0.5186 0.6909 0.2298
m=6
WTI 0.7381 0.9328 0.4498 0.5205 0.5319 0.7439 0.1939
Brent 0.7160 0.9659 0.4242 0.5752 0.5288 0.6941 0.2439
m=7
WTI 0.7091 0.8959 0.3915 0.5492 0.5362 0.7541 0.2189
Brent 0.6809 0.9762 0.3576 0.6520 0.5377 0.7016 0.2438
m=8
WTI 0.6869 0.9747 0.3819 0.6388 0.5423 0.8019 0.2174
Brent 0.6592 1.0112 0.3439 0.7036 0.5445 0.7449 0.2445
3.3.5 Multifractal Analysis of Price Fluctuations

at Different Scales
In order to investigate the impacts of large and small price fluctuations on the price
behaviors, we divided the multifractal spectrum into left- and right-half part by
means of maximum extreme values of f(a), i.e., f ða0 Þ ¼ maxf ðaÞ. The left-half
a0 amin corresponds to the section where q > 0, which reflects the price behaviors
of the large price fluctuations; while the right one amax a0 corresponds to the
section where q < 0, which describes the price behaviors of the small fluctuations.
Thus, we discussed the left- and right-half spectra changes with the t respectively,
which can provide us better insights into the underlying price dynamic behaviors.
As the a is a decreasing function of q, we only need to calculate two points to
obtain Da. We take maxjqj ¼ 60 because of the following two considerations:
(a) (b)
Fig. 3.16 Depicts the evolution of a the left-half spectra and b the right-half spectra. As we can
see, the dynamics of left-half spectra differs significantly from the right-half spectra
(1) The slope of f ðaÞ is q, when jqj approximates to 60, the absolute values of the
slope of the spectrum in the two endpoints are 60, the angle formed by the
tangent and the x-axis is arctan 60 ¼ 1:5541p=2, which is nearly vertical, a is
approximately convergent to a constant;
(2) If q, the exponent of partition function, is too large, we will inevitably expe-
rience computation overflow; actually, a too large q is also not necessary.
Based on the reasons above, the three points, namely, q = ±60 and q = 0 are
chosen to obtain Da. Our numerical results are depicted in Fig. 3.16. Compared
with the width of left- and right-half spectra changes with different delay, we can
find that:
(1) As for the large price fluctuations (see Fig. 3.16a): For Brent and WTI markets,
there are critical points tc , which is tc = 73(WTI) and tc = 117(Brent). The
widths of left-half spectra gradually increase before the critical points as the
delay t increasing, while after these points, the width begins to decrease.
(2) As for the small fluctuations (see Fig. 3.16b): the width of right-half spectra
increases slowly as the delay t increases in the long run in both markets.
(3) The left-half spectra fluctuate much more smoothly than the right-half ones,
which imply that there exist different underlying mechanisms due to different
scales of price fluctuations in the formation of multifractality in the markets.
3.3.6 Conclusions
In this section, we investigated the multifractality features and the causes in the
most important crude oil markets, that is, Brent and WTI. We acquired the fol-
lowing nontrivial findings:
First of all, empirical evidence from both MF-DFA and MF-SSA confirmed that
there exist multifractal features in both the markets.
Secondly, by shuffling the original time series, we destroyed the underlying
nonlinear temporal correlation, that is, any market memories were eliminated; but
the distribution of price fluctuations remains exactly the same after the shuffling
procedure. We found that the multifractal strength is significantly weaken but
remains; therefore, we identified that long range correlation mechanism constitutes
major contributions in the formation in the multifractality of the markets; while at
the same time, the market dynamics is also influenced by non-Gaussian distribution.
Thirdly, in monofractal sense, WTI is more persistent than Brent; while in
multifractal sense, the width of multifractal spectrum of WTI market is narrower
than that of Brent.
Finally, by tracking the evolution of left- and right-half spectra, the dynamics of
large fluctuations is significantly different from that of the small ones, which
implies that there exist different underlying mechanisms in the formation of mul-
tifractality in the markets.
Of course, there are still some questions waiting to be answered in our future
works: are there any more causes of multifractality in the analyzed markets along
with nonlinear temporal correlation and non-Gaussian distribution? For example, as
one of the authors found that the expectations and timing strategies of speculators
can distort crude oil price behaviors and market dynamics (He et al. 2009), the
following questions are intriguing to be answered: what are the roles of market
participant’ expectation and transaction behavior in the multifractality formation?
Are they the other possible causes for multifractality? What are the underlying
mechanisms that result in the different impacts of large fluctuations and small ones
on the price dynamics?
References
B. Adrangi, et al. Chaos in oil prices? Evidence from futures markets. Energy Econ. 23(4), 405–425
(2001)
J. Alvarez-Ramirez, J. Alvarez, E. Rodriguez, Short-term predictability of crude oil markets: a
detrended fluctuation analysis approach. Energy Econ. 30, 2645–2656 (2008)
L. Bachelier, Theory of speculation (1900), in The Random Character of Stock Market Prices
(MIT Press, Cambridge, MA, 1964)
A.-L. Barabasi, T. Vicsek, Multifractality of self-affine fractals. Phys. Rev. A 44, 2730–2733
(1991)
G. Du, X. Ning, Multifractal properties of Chinese stock market in Shanghai. Phys. A: Stat. Mech.
Appl. 387(1), 261–269 (2008)
References 47
P. Gopikrishnan, V. Plerou, X. Gabaix, L.A.N. Amaral, H.E. Stanley, Price fluctuations and
market activity. Phys. A 299, 137–143 (2001)
L.-Y. He, Y. Fan, Y.-M. Wei, The empirical study on fractal features and long-run memory
mechanism in the petroleum pricing systems. Int. J. Glob. Energy Issues 27, 492–502 (2007)
L.-Y. He, Y. Fan, Y.-M. Wei, Impact of speculator’s expectations of returns and time scales of
investment on crude oil price behaviors. Energy Econ. 31(1), 77–84 (2009)
Z.-Q. Jiang, W.-X. Zhou, Multifractal analysis of Chinese stock volatilities based on the partition
function approach. Phys. A: Stat. Mech. Appl. 387(19–20), 4881–4888 (2008)
Multifractal detrended fluctuation analysis of nonstationary time series. Phys. A: Stat. Mech.
Appl. 316, 87–114 (2002)
E. Koscielny-Bunde et al., Long-term persistence and multifractality of river runoff records:
detrended fluctuation studies. J. Hydrol. 322(1–4), 120–137 (2006)
S. Kumar, N. Deo, Multifractal properties of the Indian financial market. Phys. A: Stat. Mech.
Appl. 388(8), 1593–1602 (2009)
J. Kwapien, P. Oswiecimka, S. Drozdz, Components of multifractality in high-frequency stock
returns. Phys.: Stat. Theor. Phys. 350, 466–474 (2005)
B.B. Mandelbrot, New methods in statistical economics. J. Polit. Econ. 71, 142–440 (1963)
B. Mandelbrot, J.R. Wallis, Robustness of the rescaled range and the measurement of long-run
statistical dependence. Water Resour. Res. 5, 967–988 (1969)
K. Matia, Y. Ashkenazy, H.E. Stanley, Multifractal properties of price fluctuations of stocks and
P. Norouzzadeh, B. Rahmani, A multifractal detrended fluctuation description of Iranian rial-US
dollar exchange rate. Phys. A: Stat. Mech. Appl. 367, 328–336 (2006)
G. Oh, S. Kim, C. Eom, Long-term memory and volatility clustering in high-frequency price
changes. Phys. A: Stat. Mech. Appl. 387(5), 1247–1254 (2008)
E.E. Peters, in Chaos and Order in Capital Markets: A New View of Cycles, Prices and Market
Volatility (Wiley, New Jersey, 1991)
E. Peters, Fractal Marker Analysis: Applying Chaos Theory to Investment and Economics (Wiley,
USA, 1994)
B.M. Tabak, D.O. Cajueiro, Are the crude oil markets becoming weakly efficient over time? A test
for time-varying long-range dependence in prices and volatility. Energy Econ. 29(1), 28–36
(2007)
L. Telesca, V. Lapenna, M. Macchiato, Multifractal fluctuations in seismic interspike series. Phys.
A: Stat. Mech. Appl. 354, 629–640 (2005)
Y. Yuan, X.-T. Zhuang, X. Jin, Measuring multifractality of stock price fluctuation using
multifractal detrended fluctuation analysis. Phys. A: Stat. Mech. Appl. 388(11), 2189–2197
(2009)
Chapter 4
Multifractal Detrended Cross-Correlation
Analysis (MF-DCCA)
A large number of studies have conducted empirical analyses of the long-range

auto-correlations in several financial markets, especially the stock markets
(Papaionnou and Karytinos 1995; Lim 2007; Cajueiro and Tabak 2008) and foreign
exchange markets (Tabak 2006; Wang 2011). To reveal the multifractal features of
two cross-correlated non-stationary signals, Zhou (2008) proposed multifractal
detrended cross-correlation analysis (MF-DCCA, or called MF-DXA), a combi-
nation of the MF-DFA and DCCA methods. Since then, DCCA and MF-DCCA
have been discussed on methodology (Jiang and Zhou 2011; Kristoufek 2011;
Hedayatifar 2011) and used widely in the detection of cross-correlations between
two financial series in several literature [35–39]. Concerning the methodology,
Jiang and Zhou created MF-X-DMA (2011) which is an extension of (Zhou 2008)
using MF-DMA (Gu and Zhou 2010) and DMA (Alessio 2002). Kristoufek pro-
posed MF-HXA (Jiang and Zhou 2011) based on the height-height correlation
analysis of Barabasi and Vicsek (1991), and Hedayatifar et al. (2011) extended the
MF-DCCA to the method of coupling detrended fluctuation analysis (CDFA) for
the case when more than two series are correlated to each other. In addition,
recently, some researchers discussed the effects of different trends on DCCA or
MF-DCCA and proposed a variety of different DCCA or MF-DCCA methods using
different trend-filtering method (Zhao et al. 2011, 2012; Horvatic et al. 2011; Song
and Shang 2011). Concerning the empirical analysis, with the exception of the
previous references, Wang et al. (2010a, b) investigated the cross-correlations
between Chinese A-share and B-share markets using the MF-DCCA method and
found that multifractality existed. He and Chen (2011a, b, c) suggested that mul-
tifractal cross-correlation features are significant in both Chinese and US agricul-
tural futures markets using the MF-DCCA method. Moreover, they established
theoretical proof of the relationship between the bivariate cross-correlation expo-
nent and the generalized Hurst exponents for the time series of respective variables
(He and Chen 2011a, b, c). Yuan et al. (2012) found that the cross-correlation
between the Chinese stock price and trading volume is multifractal.

50 4 Multifractal Detrended Cross-Correlation Analysis (MF-DCCA)
4.1 Methodology
Assume that there are two time series fxðiÞg and fyðiÞg, i ¼ 1; 2; . . .; N, where N is
the length of the series, then the MF-DCCA method can be summarized as follow:
① Construct the profile
X
i X
i
XðiÞ ¼ ðxðtÞ xÞ; YðiÞ ¼ ðyðtÞ yÞ; i ¼ 1; 2; . . .; N; ð4:1Þ
t¼1 t¼1
PN PN
where x ¼ N1 t¼1 xðtÞ and y ¼ N1 t¼1 yðtÞ.
② The profile {XðiÞ} and {YðiÞ} are divided into Ns ¼ ½N=s
non-overlapping windows (or segments) of equal length s. Since the
record length N does not need to be a multiple of the considered
time-scale s, a short part at the profile will remain in most cases. In order
to take into account this part of the record, the same procedure is
repeated starting from the other end of the record. Thus, 2Ns segments
are obtained altogether.
③ The local trend X v ðiÞ and Y v ðiÞ for each segment v is evaluated by
least-square fit of the data, v ¼ 1; 2; . . .; 2Ns . Then, we determine the
variance
1X s
F 2 ðs; vÞ ¼ jXððv 1Þs þ iÞ X v ðiÞj jYððv 1Þs þ iÞ Y v ðiÞj; ð4:2Þ
s i¼1
for each segment v, v ¼ 1; 2; . . .; N s , and
1X s
F 2 ðs; vÞ ¼ jXðN ðv Ns Þs þ iÞ X v ðiÞj jYðN ðv Ns Þs þ iÞ Y v ðiÞj;
s i¼1
ð4:3Þ
for each segment v, v ¼ Ns þ 1; . . .; 2Ns . Here, X v ðiÞ and Y v ðiÞ denote the fitting
polynomial with order k in segment v (conventionally called MF-DCCA-k).
Traditionally, it is chosen that 2k þ 2 s N=4.
4.1 Methodology 51
④ By averaging over all segments, the q-order fluctuation function is

obtained as follows
( )1=q
1 X2Ns
Fq ðsÞ ¼ ½F 2 ðs; vÞq=2 ð4:4Þ
2Ns v¼1
for any q 6¼ 0, and

( )
1 X2Ns
F0 ðsÞ ¼ exp ln½F 2 ðs; vÞ ð4:5Þ
4Ns v¼1
for q ¼ 0, where the index variable q can take any real value. For q ¼ 2, the
standard DCCA procedure is retrieved.
⑤ Finally, the scaling behavior of the fluctuations is determined by ana-
lyzing log-log plots of Fq ðsÞ versus s for each value of q. If two series
fxðiÞg and fyðiÞg are long-range cross-correlated, Fq ðsÞ will increase for
large values of s, as a power-law.
Fq ðsÞ sHxy ðqÞ : ð4:6Þ
This can be presented as follows
log Fq ðsÞ ¼ Hxy ðqÞ logðsÞ þ log A: ð4:7Þ
The scaling exponent Hxy ðqÞ is known as the generalized cross-correlation

exponent, describing the power-law relationship between two time series.
Especially, if the time series x1 is identical to x2, MF-DCCA is equivalent to
MFDFA. Moreover, if scaling exponent Hxy ðqÞ is independent on q, the
cross-correlations between two time series are monofractal. If scaling exponent
Hxy ðqÞ is dependent on q, the cross-correlations between two time series are mul-
tifractal. Furthermore, for positive q, Hxy ðqÞ describes the scaling behavior of the
segments with large fluctuations. On the contrary, for negative q, Hxy ðqÞ describes
the scaling behavior of the segments with small fluctuations.
The bivariate Hurst exponent Hxy ð2Þ has similar properties and interpretation as
a univariate Hurst exponent (Kristoufek 2011). If scaling exponent Hxy ð2Þ [ 0:5,
the cross-correlations between two time series are long-range persistent1. If scaling
1
There are different descriptions of the implications. Kristoufek consider “if Hxy ð2Þ [ 0:5, the
series are cross-persistent so that a positive (a negative) value of Dxt Dyt is more mare statistically
probable to be followed by another positive (negative) value of Dxt þ 1 Dyt þ 1 .” Podobnik and
Stanley, and Yuan et al. suggest that “long-range cross correlations between two stocks imply
that each stock separately has long memory of its own previous values and, additionally, has a
long memory of previous values of the other stock”, and Podobnik and Stanley also argue that
exponent Hxy ð2Þ\ 0:5, the cross-correlations between two time series are
anti-persistent. If scaling exponent Hxy ð2Þ ¼ 0:5, there is no cross-correlations or at
most short-range cross-correlations between two time series.
In order to measure the time-varying degree of multifractality, the financial risk
measures, DH and Da was proposed by Yuan and Zhuang (2009, 2012) as the
following
DH ¼ Hmax ðqÞ Hmin ðqÞ: ð4:8Þ
The greater is DH, the stronger is the degree of multifractality. Therefore,

according to this idea, if HðqÞ is replaced by Hxy ðqÞ in Formula (4.8), then we can
obtain DHxy which measure the degree of multifractality of cross-correlation.
According to Shadkhoo and Jafari, the similar relationship between classical
multifractal scaling exponents sxy(q) and q can be given by
sxy ðqÞ ¼ qHxy ðqÞ 1 ð4:9Þ
If sxy(q) is linear with q, the cross-correlation of the correlated series is

monofractal; otherwise, it is multifractal. By the means of a Legendre transfor-
mation, we can obtain the following relationships
0
a ¼ Hxy ðqÞ þ qHxy ðqÞ
ð4:10Þ
f ðaÞ ¼ qða Hxy ðqÞÞ þ 1
And the strength of multifractality can also be estimated by the width of mul-
tifractal spectrum, which is given by
Da ¼ amax amin ð4:11Þ
4.2 Empirical Analysis on Chinese Stock-Exchange

Market
On 21 July 2005, the People’s Bank of China reformed the exchange rate regime by
establishing a managed floating exchange rate regime based on market supply and
demand with reference to a basket of currencies. Since then, the flexibility and
impact of the Chinese Yuan (RMB) exchange rate have become a hot topic of
discussion among financial researchers and managers. Naturally, the relationship
between the Chinese RMB exchange market and stock market is an important case.
Since 2005, the RMB exchange rate regime has also been adjusted continuously to
“power-law cross-correlations indicating that a large increment in one variable is more likely to be
followed by large increment in the other variable.”
4.2 Empirical Analysis on Chinese Stock-Exchange Market 53
accommodate changes in the domestic and international economic environment. In

July 2008, China halted the RMB’s rise to cope with the global financial crisis. On
19 June 2010, the People’s Bank of China decided to enhance its flexibility and
proceeded to further reform the RMB exchange rate regime. Therefore, a problem
worthy of study is whether the changes in the Chinese exchange rate regime in July
2008 and June 2010 affected the relationship between the Chinese exchange market
and stock market after the exchange rate reform on 21 July 2005.
Therefore, we applied the MF-DCCA method to empirically analyze the
cross-correlations between the RMB exchange market and stock market in China.
First, ours is the first study to apply the MF-DCCA method in the empirical analysis
of multifractal cross-correlations between the RMB exchange market and stock
market in China. Second, using the rolling windows method, we investigate the
time-varying features of multifractal cross-correlations. Third, our empirical results,
to some extent, do not support the theoretical results of the relationship between the
bivariate cross-correlation exponent and the generalized Hurst exponents for time
series of respective variables (He and Chen 2011a, b, c). We arrive at the following
reliable conclusions: First, the cross-correlation between the Chinese exchange
market and stock market is significant. Second, multifractality exists in such
cross-correlations. Moreover, the cross-correlated behavior of small fluctuations is
found to be more persistent than that of large fluctuations. Third, multifractal
cross-correlations vary with time and are sensitive to the reform of the RMB
exchange rate regime. Furthermore, the reduction of RMB exchange rate flexibility
in July 2008 strengthened the persistence of cross-correlations and decreased the
degree of multifractality, whereas the enhancement of RMB exchange rate flexi-
bility in June 2010 weakened the persistence of cross-correlations and increased the
degree of multifractality. Finally, the theoretical proof put forward in (He and Chen
2011a, b, c) is discussed.
4.2.1 Data
The exchange rate between the RMB and US dollar (USD) is the most important
type of data in the Chinese foreign exchange market. The Shanghai Stock Exchange
Composite Index (SSCI) is one of the most important stock market indices that
reflect fluctuations in the Chinese stock market. Thus, we select the daily nominal
RMB/USD exchange rate and the closing price of the SSCI as the price variables
representing the price data of the foreign exchange market and stock market in
China. Given that the reform of the Chinese exchange occurred on 21 July 2005,
the sample interval is from 1 August 2005 to 20 October 2011. A total of 1392 pairs
of common price data were adopted after removing the data from different trading
dates in the RMB exchange market and stock market. The original sample data
were obtained from FC Station 3.0 (http://icbc-zx1.finchina.com/product-2.asp).
Returns are computed by rt ¼ logðPt Þ logðPt1 Þ, with Pt being the closing
price index at time t. Therefore, for the return series, the empirical data numbered
.10 .004
.05 .002
.00 .000
-.05 -.002
-.10 -.004
CSR RER
-.15 -.006
250 500 750 1000 1250 250 500 750 1000 1250
Fig. 4.1 Returns of the SSCI and RMB/USD exchange rate
Table 4.1 Descriptive statistics for the returns of SSCI and RMB/USD
Mean Max Min S. D Ske Kur J-B Q(3) Q(15) Q(30)
CSR 0.0006 0.0903 −0.1276 0.0202 −0.50 6.44 732.94*** 12.77*** 39.27*** 57.31***
RER −0.0002 0.0036 −0.0043 0.0009 −0.63 6.19 535.44*** 56.82*** 77.38*** 97.43***
Note 1*, ** and *** denotes 10%, 5% and 1% significant level respectively; Symbols “Max”, “Min”, “S. D”, “Ske”,
“Kur” denote Maximum, Minimum, Stv. Dev, Skewness and Kurtosis respectively; The J-B denotes Jarque-Bera
statistics [45] testing for the normality assumption. Q(i) denotes the value of Jarque-Bera statistics with i lags
1391 pairs. Returns of RMB/USD and SSCI (denoted by RER and CSR) are pre-
sented in Fig. 4.1, whereas the descriptive statistics of the two return series are
provided in Table 4.1.
As Fig. 4.1 shows, fluctuation persists in the return series for the RMB exchange
market and stock market in China. However, for RMB/USD returns, structural
changes can be seen between 2009 and 2010 (to be more precise, the changes
occurred before 21 June 2010). These changes may be attributed to the adjustment
of RMB exchange rate flexibility on 19 June 2009 and 19 June 2010.
Table 4.1 indicates that both skewness and kurtosis for the two time series show
significant departures from normality. The Jarque–Bera statistics for normality test
suggests that for all of series, the normality assumption of returns can be rejected.
The Jarque–Bera Q statistics with lags are all significant at the 1% significant level,
indicating that all series may present a long-range persistence (or long memory).
4.2.2 Cross-Correlation Test
We applied the new cross-correlation test proposed by Podobnik et al. (2009) to test
for the significant level of cross-correlation between the Chinese exchange market
and stock market.
For two stochastic processes fxðiÞg and fyðiÞg sharing the same length N, let
cross-correlation function
PN
k¼i þ 1 xk yki
Ci ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN 2 PN 2 : ð4:12Þ
k¼1 xk k¼1 yk
Then the cross-correlations test statistic
Xm
Ci2
Qcc ðmÞ ¼ N 2 ; ð4:13Þ
i¼1
Ni
which is approximately v2 ðmÞ distributed with m degrees of freedom. If there are no

cross-correlations between two time series, the cross-correlation test agree well with
the v2 ðmÞ distribution. If the cross-correlations test exceeds the critical value of the
v2 ðmÞ distribution, then the cross-correlations are significant at special significant level.
Figure 4.2 shows the cross-correlation statistics (logarithmic form) for the
returns of the Chinese exchange market and stock market. The degrees of freedom
vary from 1 to 1000. The logarithms of the critical values for v2 ðmÞ distribution at
the 5% level of significance are also shown in Fig. 4.2 for comparison. The results
show that for the returns of the Chinese exchange market and stock market, the
empirical results of Qcc(m) are larger than the critical values of v2 ðmÞ distribution at
a 5% level of significance. Therefore, the null hypothesis of no cross-correlations
can be rejected in its entirety. This finding indicates that long-range
cross-correlations exist between the exchange market and stock market in China.
Furthermore, to affirm our results obtained above, we also applied another new
method of the cross-correlation coefficient recently proposed in (Zebende 2011;
Vassoler and Zebende 2012; Podobnik et al. 2011). The results are shown in
30
critical values
log(Qcc(m))
25
log(Qcc(m))
20
15
10
-2.4
-5
0 0.5 1 1.5 2 2.5 3
log(m)
Fig. 4.2 Cross-correlation statistics Qcc(m) (logarithmic form)

Fig. A1 in Appendix. According to Table 1 in Podobnik et al. (2011), the critical

values for DCCA cross-correlation coefficient for each window size n is not larger
than 0.4 when T is 1000 (or 2000). However, from Fig. A1, we can find that the
minimum of the DCCA cross-correlation coefficient qDCCA of the Chinese exchange
market and stock market versus window size n is about 0.7. Therefore, this confirms
the results of Qcc(m) test that long-range cross-correlations exist between the
Chinese exchange market and stock market.
4.2.3 Multifractal Detrended Cross-Correlation Analysis
The cross-correlation test based on the statistics in Eq. (4.13) can only test for the
presence of cross-correlation qualitatively. To present the cross-correlations quan-
titatively, the MF-DCCA method is used to estimate the cross-correlation exponent.
Figure 4.3 shows the log-log plots of log(Fq(s)) versus log(s) for the Chinese
exchange market and stock market as q ¼ 10; 9; 8; . . .; 10 when polynomial
order k = 3 (i.e., MF-DCCA-3, when k = 1, 2, 4, and 5, the results are qualitatively
similar). Although larger fluctuations of log(Fq(s)), especially for q 4 and q 5,
are observed for large values of log(s), the estimated coefficients Hxy(q) and constant
log(A) in Eq. (4.7) are all significant at 1% significant level by the method of linear
least squares. For different q, each curve is linear, suggesting that power-law
cross-correlations exist between the Chinese exchange market and stock market.
In Fig. 4.4, the relationship between Hxy(q) and q is displayed, with q varying
from −10 to 10. We can find that the scaling exponents Hxy(q) decrease from larger
than 1 to smaller than 0.6. Thereby, Hxy(q) is not a constant, which indicates that
multifractality exists in the cross-relations between the Chinese exchange market
-0.5
-1
-1.5
-2
log(Fq(s))
-2.5
-3
-3.5
-4
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
log(s)
( Curves from the bottom to the top are corresponding to the plots with q =10, 9, ..., -9, -10. )
Fig. 4.3 Log-log plot of log(Fq(s)) versus log(s) for the Chinese exchange market and stock
market with q varying from 10 to −10 (k = 3)
1.3
MF-DCCA-1
1.2 MF-DCCA-2
MF-DCCA-3
1.1 MF-DCCA-4
MF-DCCA-5
1
Hxy(q)
0.9
0.8
0.7
0.6
0.5
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 4.4 MF-DCCA scaling exponents Hxy(q) for the Chinese exchange market and stock market
with q varying from −10 to 10 and fitting polynomial order k varying from 1 to 5
and stock market. Moreover, the scaling exponents Hxy(q) for q < 0 are larger than
those for q > 0, although they are all larger than 0.5. This indicates that the
cross-correlated behavior of small fluctuations is more persistent than that of large
fluctuations. For the two intervals q 0 and q 2, Hxy(q) functions are monotoni-
cally decreasing, which is consistent with the results of the references (Zhou 2008;
Wang et al. 2010a, b; He and Chen 2011a, b, c; Yuan et al. 2012), but Fig. 4.4 also
displays a “U” curve for the values of Hxy(q) for 2 q 2. The key is that
Hxy(q) functions are monotonically increasing for 0 q 2, indicating that the
persistence of the cross-correlations between the Chinese RMB/USD exchange
market and stock market are increase with the “normal” fluctuations2 enhancing. In
fact, narrowed fluctuations of the RMB/USD exchange rate during July 2008–June
2010 may cause them more persistent and strengthen the relationship between the
Chinese exchange market and stock market. Furthermore, large fluctuations of
the Chinese stock market are likely to occur and present more persistent feature for
the global financial crisis and the slowdown in growth of the Chinese economy
during 2008–2010. As a result, small fluctuations of the RMB/USD exchange rate
perhaps lead to a large fluctuation of the Chinese stock market. Thus, it is reasonable
that the persistence of the cross-correlations increases gradually as the fluctuations of
these two financial markets increasing with q varying from 0 to 2. Therefore, we
suggest that the reform of the RMB exchange rate regime and the financial crisis in
2008 may be one of the factors affecting the persistence of the cross-correlation
between the Chinese RMB/USD exchange market and stock market.
Additionally, from Fig. 4.4, we can find that the results of multifractal
cross-relations estimated by polynomial order k = 1, 2, 3, 4, 5 are similar. Notably,
when k = 2, 3, 4 and 5, the difference among estimated results is very small.
2
In order to facilitate the description, we denote the fluctuations dominate the scaling exponent
Hxy(q) for q\0, q [ 2 and 0 q 2 are small fluctuation, normal fluctuation and large fluc-
tuation, respectively.
4.2.4 Scaling Consistency Analysis
If the scaling exponents estimated are constant or very close when the time scales
s vary, then they are consistent, indicating that cross-correlations exist as a feature
of indefinite memory. In other words, the length of persistence is indefinite.
Similarly, if time scale s* exists, in which the scaling exponents are significantly
differ between s > s* and s s*, then the scaling exponents are inconsistent. The
time scale s* is called “crossover.”
Figure 4.3 provides the log-log plots of log(Fq(s)) versus time scale log(s) for
the return series of the Chinese exchange market and stock market. The values of
q vary from −10 to 10. The results show two “crossovers” located at approximately
s1 = 47 and s2 = 157 (log(s*) = 1.67 and 2.2, respectively).
To prove the intuitive judgment above, we provide the scaling exponents Hxy(2)
for different scale intervals [1, s] versus s varying from 29 to 347 in Fig. 4.5
(k = 3). The scaling exponents Hxy(2) decrease with time scale s for s\s1 and
s [ s2 , whereas the scaling exponents Hxy(2) increase with time scale s for
s1 \s\s2 . These findings prove the crossover results in Fig. 4.3.
Besides, for the limited size of the sample, we have to admit that there may be a
finite-size effect in the determination of the scaling exponents (Zhou 2009, 2012) in
our empirical analysis. Resulting from the decreasing of the scaling exponents
Hxy(2) with time scale s increasing when s [ s2 , the scaling exponent Hxy(2) of the
cross-correlation between the Chinese exchange market and stock market maybe
decrease if the sample size enlarged. Moreover, according to the references (Zhou
2012; He and Chen 2011a, b, c), the curves of Fq(s) versus s in log-log scales of
MF-X-DMA have smaller fluctuations than MF-X-DFA (i.e. MF-DCCA in this
book). Thereby, based on more sufficient data and MF-X-DMA method, the “U”
curve displayed in Fig. 4.4 may be flattened and Hxy(q) may monotonically
decrease. This may be a valuable research worthy of doing in the future.
0.76
0.74
0.72
Hxy(2)
0.7
0.68
0.66
0.64
1.4 1.6 1.8 2 2.2 2.4 2.6 2.8
log10(s)
Fig. 4.5 Plots of Hxy(2) for different scale intervals [1, s] versus s varying from 29 to 347 for
returns of the Chinese exchange market and stock market with (k = 3)
4.2.5 Dynamics of Cross-Correlations Over Time
The rolling windows method is used to investigate the time-varying feature of the
cross-correlations between the Chinese exchange market and stock market. A total
of 250 business days (approximately a year) was selected as the length of each
rolling window (Wang 2010a, b). However, the Hurst exponent estimated through
DFA is affected by the choice of minimum or maximum scales (Kristoufek 2010),
which are limited by the sample size. In this chapter, the length of each window is
fixed at 500 days (approximately two years) to obtain a more precise Hurst
exponent estimation.
Figure 4.6 provides the plots of the scaling exponents Hxy(q) for q = −6, −4, −2,
0, 2, 4, 6, with a moving window. The window length is fixed at 500 days. The
slide step is one day. The shaded area in Fig. 4.6 denotes the period from 10 July
2008 (in July 2008, China halted the RMB’s rise to cope with the global economic
crisis. Since then, the RMB has been pegged at approximately 6.83 per dollar) to 19
June 2010 (on this day, the Chinese government decided to proceed with further
reform of the RMB exchange rate regime to enhance its flexibility), indicating the
period wherein the flexibility of RMB exchange rate was limited by the Chinese
government.
From Fig. 4.6, we can find that the scaling exponent Hxy(q) increased gradually
in the period more than half of the flexibility limited period of the RMB exchange
rate, and then decreased sharply in the later period, as denoted by the shaded area.
Hxy(q) then decreased slowly after 19 June 2010 (1096th point). Notably, Hxy(2)
was approximately 0.7 before the end of 2009 but gradually approached 0.5 after
that time and fluctuated to approximately 0.55 after 19 June 2010. These findings
indicate that the cross-correlations between the Chinese exchange market and stock
market grew stronger over time in the limited flexibility period of the RMB
0.9
0.8
0.7
0.6
Hxy(q)
0.5
0.4 q = -6
q = -4
0.3 q = -2
q= 0
0.2
q= 2
0.1 q= 4
q= 6 10 July, 2008 - 19 June, 2010
0
500 600 700 800 900 1000 1100 1200 1300 1400
sample points
Fig. 4.6 Scaling exponents Hxy(q) for the exchange market and stock market with a rolling
window. The window length is fixed at 500 days, and q = −6, −4, −2, 0, 2, 4, 6. The line with
marker D is Hxy(2), and the line of Hxy = 0.5 is also provided
exchange rate but grew weaker after the enhancement of the flexibility of RMB
exchange rate. We therefore suggest that the reform fostering the flexibility of the
RMB exchange rate regime is very important for the efficiency of the RMB
exchange market. Although the Chinese stock market has become gradually but
increasingly efficient since the reform in 2005, the cross-correlations between the
Chinese exchange market and stock market were determined primarily by the
flexibility regime of the RMB exchange rate.
Additionally, from Fig. 4.6, we also notice that the difference of the scaling
exponents Hxy(q) for q = −6, −4, −2, 0, 2, 4, 6 varied in different periods, indi-
cating that the degree of multifractality differed over time. The difference is smaller
in the shaded area than in other areas. To obtain more precise results, we apply the
financial risk measures, DHxy in Eq. (4.8) to measure the degree of multifractalilty
of cross-correlations. The evolution of DHxy for cross-correlations of the Chinese
exchange market and stock market are provided in Fig. 4.7.
Figure 4.7 shows that DHxy , decreasing from an initial value of 0.3, is approx-
imately 0.1 in the period from 10 June 2010 to 19 June 2010, whereas it increased
sharply after 10 June 2010 and then fluctuated to approximately 0.38. This finding
proves the results in Fig. 4.6, implying that the degree of multifractality is lower
after the period of limited flexibility in July 2008 and higher after the enhancement
of flexibility on 19 June 2010. Therefore, we believe that reducing the flexibility of
RMB exchange rate can decrease the multifractality of cross-correlations between
the Chinese exchange market and stock market, indicating that the risk of portfolio
arbitrage in these two financial institutions can be reduced. On the contrary,
enhancing the flexibility of RMB exchange rate will increase the risk of portfolio
arbitrage in the Chinese exchange market and stock market, which can be reflected
by the higher degree of multifractality in cross-correlations.
0.8
0.7
DELTA(Hxy)
0.6
0.5
0.4
0.3
0.2
0.1
0
500 600 700 800 900 1000 1100 1200 1300 1400
sample points
Fig. 4.7 Time evolution of DHxy

4.2.6 Discussion
4.2.6.1 Rolling Windows
Grech and Mazur argued that the local exponent at a given time t depends on the
time-window length (Grech and Mazur 2004). In various studies with different
objectives, the lengths of rolling windows vary. Cajueiro and Tabak used the length
of several years to analyze the evolution of long-term correlation, whereas Wang
et al. only used 250 data points as the length of each time-window. In a recent study
(Vasilescu et al. 2010a, b) on Lidar measurements conducted in the Black Sea
waters, the structure functions and the singular measure algorithms were found to
have possible application with reliable results for time series of slightly more than
800 points (Cristescu et al. 2012). However, for shorter data sets (less than
700 points), the results are no longer reliable. This hypothesis is used explicitly or
implicitly in relevant studies on financial time series (Schmitt 1998; Matteo 2007;
Schmitt et al. 2000; Stanley and Plerou 2001). Furthermore, when rather short time
series are used, the universal multifractal hypothesis might be misleading.
Therefore, we decided to adjust the length of a window to 1000, which corresponds
to approximately four years, to achieve a robust comparison of the previous results
estimated with a length of 500. The graphical representation of the scaling expo-
nents Hxy(q) for the Chinese exchange market and stock market is shown in
Fig. 4.8. The results in Fig. 4.8 are very similar to the previous results estimated by
a rolling window length of 500 (Fig. 4.6), implying the robustness of the results.
However, as shown in Fig. 4.8, the information on 500 time points are loosened
when the window length is 1000 compared with information based on a window
length of 500, which may not detect the dynamics of cross-correlations in the period
of limited flexibility for RMB exchange rate. Therefore, in the previous analysis
1.05
0.95
0.9
0.85
Hxy(q)
0.8
0.75
0.7
0.65
0.6
0.55
1000 1050 1100 1150 1200 1250 1300 1350 1400
sample points
Hxy(-6) Hxy(-4) Hxy(-2) Hxy(0) Hxy(2) Hxy(4) Hxy(6)
Fig. 4.8 Scaling exponents Hxy(q) with window moving. The window length is fixed at
1000 days, and q = −6, −4, −2, 0, 2, 4, 6. The line with marker D denotes Hxy(2)
(Sect. 4.2.5), we also used a rolling window length of 500 to investigate the
dynamics of cross-correlations.
4.2.6.2 Relationship Between Bivariate Cross-Correlation Exponents

and the Generalized Hurst Exponents
Zhou observed that for two time series constructed by binominal measure from the
p-model, the following relationship exists (Zhou 2008):
Hxy ðqÞ ¼ ½Hxy ðqÞ þ Hxy ðqÞ=2: ð4:14Þ
However, He and Chen (He and Chen) proved that if s ! 1, the relationship
satisfies the following inequality:
Hxy ðqÞ ½Hxy ðqÞ þ Hxy ðqÞ=2: ð4:15Þ
Therefore, to assess the relationship between bivariate cross-correlation expo-

nents and the generalized Hurst exponents, Fig. 4.9 provides the cross-correlation
exponents Hxy(q) estimated via MF-DCCA and the generalized Hurst exponents
Hxx(q) and Hyy(q) of the Chinese exchange market and stock market estimated via
MF-DFA. For comparison, we also present the value of ½Hxx ðqÞ þ Hyy ðqÞ=2 in
Fig. 4.9. If Hxx(q) (resp. Hyy(q)) is a constant, the Chinese exchange (resp. stock)
market is monofractal; otherwise, it is multifractal.
From Fig. 4.9, we can find that Hxx(q) and Hyy(q) are vary with q, indicating that
the Chinese exchange market and stock market are multifractal. In addition, in order
to discuss the Eq. (4.15) accurately, Fig. 4.10 provides the difference between
½Hxx ðqÞ þ Hyy ðqÞ=2 and Hxy(q). From Figs. 4.9 and 4.10, we can find that Hxy(q) is
always less than ½Hxx ðqÞ þ Hyy ðqÞ=2 for q\0, but Eq. (4.15) does not hold for
q [ 0. Thus, Eq. (4.15) is not verified by empirical analysis of the Chinese
exchange market and stock market. Moreover, we find that the absolute values of
Magenta— Hxx(q)
1.8 Green— Hyy(q)
Black—Hxy(q)
1.6 Blue—(Hxx(q)+Hyy(q))/2
1.4
H(q)
1.2
0.8
0.6
0.4
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 4.9 H(q) as a function of q between the Chinese exchange market and stock market
the difference between ½Hxx ðqÞ þ Hyy ðqÞ=2 and Hxy(q) are significantly larger for
q < 0 than for q > 0. Moreover, the difference almost close to 0 when q > 0.
4.2.6.3 Implications
Our results reveal important modeling and policy implications for the Chinese
exchange market and stock market.
On the one hand, there are three modeling implications. First, the cross-correlations
between the return series of the exchange market and stock market in China are
multifractal and are thus nonlinear. In effect, conventional linear models, such as the
linear regression model, vector-regression model, and correlation coefficient analysis,
cannot be used to model the dynamics of the cross-correlations between the RMB
exchange market and stock market in China. Second, the cross-correlations exhibit
time-varying features, implying that traditional models with a constant coefficient
cannot capture the nature of the relationship between the RMB exchange market and
stock market in China. Third, the degree of persistence of large fluctuations (for q < 0)
differ from that of small fluctuations (for q > 0), indicating that the MF-DCCA
method can be modified and proposed as a statistic measurement or parameter for
detecting the cross-correlation among the different fluctuations in the two financial
series.
On the other hand, several policy implications for Chinese financial market can
be derived. As shown in Fig. 4.11, the Hurst exponents of the Chinese stock market
are smaller than those of the RMB exchange market, an indication that the Chinese
stock market is becoming more efficient than the Chinese exchange market.
Accordingly, the Hurst exponents of the Chinese exchange market increased
gradually in the limited flexibility period of RMB exchange rate (from July 2008 to
June 2010), implying that the efficiency of the RMB exchange market gradually
decreased in this period. After enhancing the flexibility of RMB exchange rate, the
Hurst exponent decreased sharply, indicating an improvement in the efficiency of
0.2
k=1
k=2
k=3
0.15
k=4
[Hxx(q)+Hyy(q)]/2-Hxy(q)
k=5
0.1
0.05
-0.05
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 4.10 [Hxx(q) + Hyy(q)]/2−Hxy(q) versus q varying from −10 to 10

0.75
Hxx(2)
0.7 Hyy(2)
0.65
Hxx(2) and Hyy(2)
0.6
0.55
0.5
0.45
0.4
600 700 800 900 1000 1100 1200 1300 1400
sample points
Fig. 4.11 Time-varying Hurst exponents Hxx(2) and Hyy(2) evaluated for the return series of the
exchange market and stock market in China. The rolling window length is 600
the RMB exchange market. Therefore, compared with the analysis of

cross-correlations in Fig. 4.6, we believe that the reform fostering the flexibility of
RMB exchange rate regime significantly affected the cross-correlations between the
Chinese exchange market and stock market. This finding implies that, to improve
the efficiency of the RMB exchange market and weaken the cross-correlations
between the Chinese exchange market and stock market, the government needs to
do a great deal of work in the future reforms of the RMB exchange rate regime.
4.2.7 Conclusions
In this section, we investigate the cross-correlations between the exchange market

and stock market in China. We find the cross-correlations to be generally significant
based on the analysis of the significance of the statistic Qcc(m) and the
cross-correlation coefficient. The empirical results obtained through the MF-DCCA
method imply that multifractality exists in the cross-relations between the Chinese
exchange market and stock market and that the cross-correlated behavior of small
fluctuations is more persistent than that of large fluctuations. Based on the scaling
consistency analysis, we learn that the business cycle is approximately 157 days,
which is an important crossover for the cross-correlation between the Chinese
exchange market and stock market.
Moreover, using the rolling windows method, we find that the cross-correlations
between the Chinese exchange market and stock market vary at given times and are
sensitive to reforms of the RMB exchange rate. The cross-correlations grow
stronger over time in the limited flexibility period of the RMB exchange rate, but
grow weaker after the enhancement of the flexibility of the RMB exchange rate.
Moreover, the degree of multifractality in cross-correlations varies with time and is
sensitive to the reform of the RMB exchange rate. The reduction of the flexibility of
the RMB exchange rate in July 2008 decreased the degree of multifractality in
cross-correlations of the Chinese exchange market and stock market, and the
enhancement of the flexibility of the RMB exchange rate in June 2010 increased the
multifractality. This finding indicates that limiting the flexibility of the RMB
exchange rate will reduce the risk of portfolio arbitrage, and conversely, enhancing
flexibility of RMB exchange rate will increase the risk of portfolio arbitrage in the
Chinese exchange market and stock market.
Finally, several relevant discussions are provided to verify the robustness of our
empirical analysis and to discuss the relationships between the bivariate
cross-correlation exponent and the generalized Hurst exponents for the time series
of the Chinese exchange market and stock market. Furthermore, the modeling and
policy implications of this work are presented.
4.3 Empirical Analysis on Price-Volume Relationships

in Agricultural Commodity Futures Markets
4.3.1 Data
The data used in this section are the complete historical records available of the
daily prices and trading volumes taken from representative commodity futures
markets in China and US (data source: Reuter© Database), that is, hard winter
wheat futures prices from Dec. 28th, 1993 to Mar. 12th, 2010 (L = 3280) from
China’s Zhengzhou Commodity Exchange, and soy meal futures prices from Jul.
17th, 2000 to Mar. 12th, 2010 (L = 2313), No. 1 soybean futures market from Mar.
15th, 2002 to Mar. 12th, 2010 (L = 1924), corn futures market from Sep. 22nd,
2004 to Mar. 12th, 2010 (L = 1325) from China’s Dalian Commodity Exchange.
Meanwhile, the data of daily closing prices and volumes of wheat futures market
(L = 7219, from Jul. 1st to Mar. 12th, 2010), soy meal futures market (L = 7226,
from Jul. 1st to Mar. 12th, 2010), soybean futures market (L = 7225, from Jul. 1st
to Mar. 16th, 2010) and corn futures market (L = 7169, from Jul. 1st to Mar. 12th,
2010) are also chosen from Chicago Board of Trade (CBOT).
Then we define the normalized price (or volume) fluctuations as:
Tði þ DtÞ TðiÞ

gðtÞ ¼ ð4:16Þ
r
where Dt ¼ 1 day and r is the standard deviation for Tði þ DtÞ TðiÞ over the
whole time series.
We choose the commodity futures markets from the two countries based on the
following reasons: (i) all these markets are very influential globally or regionally;
and (ii) China’s markets are the promising and fast-growing ones qualified to be the
representatives of the emerging markets, while the counterparts in the USA be
representatives of the developed markets; (iii) the knowledge on commodity futures
markets in China and US are so globally important as to be publicly and intensively

concerned for academic communities, policy makers and investors all over the
world. Therefore, our results may also be applicable reference to other developed
and developing economies.
4.3.2 Cross-Correlation Test
In order to quantify bivariate cross-correlation between price and volume, we

applied a new cross-correlation test proposed by Podobnik et al. (2009) that
mentioned in Sect. 4.2.2.
By this method proposed in Podobnik et al. (2009), we obtained the
cross-correlations statistic Qcc(m) (see Fig. 4.12). The results show that for all
China’s and US agricultural futures markets, the numerical results of Qcc(m) are
larger than the critical values of v2(m) distribution at 5% level of significant, and
that the different between Qcc(m) and the critical values can describe the strength of
cross-correlations, which suggests that the long-range cross-correlations are sig-
nificant between all price-volume relationships (the solid symbols).
We also applied Fourier phase-randomization process to quantify that there is
linear or non-linear cross-correlation Podobnik et al. (2009). As we know, Fourier
phase-randomization can destroy the nonlinear correlation but preserves their linear
properties. We phased randomize the series fyðjÞg and then obtained the surrogate
0 0
series fyðjÞ g. Then for each pair of series fxðiÞg and fyðjÞ g, we recalculated the
statistic Qcc(m). If Qcc(m) decreases or even is below the critical value of
v2(m) distribution, the non-linear cross-correlations can be identified; otherwise,
there are mainly linear cross-correlations. The algorithm of phase randomization is
as follows (Small and Tse 2003):
(a) (b)
Fig. 4.12 The cross-correlations statistic Qcc(m) for different degrees of freedoms m before (solid
symbols) and after (open symbols) Fourier phase-randomization in a China’s and b US agricultural
futures market. The black line shows the critical values for the v2(m) distribution at 5% level of
significant
4.3 Empirical Analysis on Price-Volume Relationships … 67
Step 1: Performing the discrete Fourier transform of the time series;

Step 2: Shuffling the phases of the complex conjugate pairs;
Step 3: Taking the inverse Fourier transformation.
After Fourier phase-randomization, non-linear cross-correlations are destroyed.
The open symbols in Fig. 4.12 trend towards or even below the critical values of
v2(m) distribution, which is the evidence that there are strong non-linear
cross-correlations between price-volume relationships. Therefore, MF-DCCA is
applied to quantify the non-linear cross-correlation.
4.3.3 Results and Discussions
In our opinion, the nonlinear dependency and its underlying dynamical mechanisms
between price and volume as a topic of research have received a lot of attention for
several reasons:
Firstly, the knowledge of the nonlinear dependency of the price-volume rela-
tionship may help one use the information obtained from one quantity (variable) to
infer underlying dynamics or properties for the other. Therefore, a researcher or a
technical analyst can complement her understandings on market dynamics by
obtaining more comprehensive knowledge from the integrated point of view rather
than from separated quantities. It is very important to study the relationships
between two variables as both are the joint products of a single market mechanism
and discussion on only one of these variables cannot be complete without a
simultaneous discussion of another variable.
Secondly, the presence of nonlinear dependence in the form of cross-correlation
between price and volume, two of the characteristic quantities to measure the
trading activities in the market, is of interest to investors who wish to base their
trading strategies on thorough understanding of market dynamics; actually, if the
dependency exists, one can not simply jump to conclusion which variable is the
cause of the volatility of the other, since they are mutually interacted.
Thirdly, if the presence of nonlinear dependence exists, scholars and policy
makers may use the measure of the cross-correlation exponents as an indication of
the degree of market efficiency; for example, if the cross-correlation exponent is 0,
the two variables are completely uncorrelated, relevant technical analyses would be
useless; otherwise, if the cross-correlation exponent is significantly deviated from 0,
thereby the market is by no means perfectly efficient, since one may infer the
dynamics of one variable from that of the other.
Fourthly, price-volume relationship has especially significant implications for
research into futures market. Price variability affects the volume of trade in futures
contracts. These implications provide us profound insights into the impacts of
market speculation on futures prices.
Based on the above-mentioned method and discussion, we intend to (i) identify
whether nonlinear dependency exists in the price-volume relationships in the
markets; and (ii) test empirically the multifractal scaling properties in these rela-
tionships; (iii) quantify the strengths of the multifractality in the temporally
cross-correlated relationships (if any). If the nonlinear dependency and multifrac-
tality do exist, it may help us identify, explain and control the underlying physical
mechanisms that dominate and govern the market dynamics; thus, it would greatly
deepen our insights into effectiveness of the technical trading strategies in a market
with nonlinear dependency and multifractality (He and Zheng 2010).
By the method mentioned above, we obtained the price and volume relationships
between Fxy(q,s) and the scale s (see Figs. 4.13 and 4.14). The linear relationship
between lnFxy(q,s) and lns shows that for all China’s and US agricultural futures
markets, the price and volume relationships are power-law cross-correlated. From
the figures one may find that there are crossovers, for instance, the corn futures in
Panel (d) of Fig. 4.13. These crossovers may be caused by the trend. The DCCA or
MF-DCCA method can only eliminate the polynomial trend; but for the real
financial series, there are always other trends, e.g. periodic trends. Academic
(a) (b)
(c) (d)
Fig. 4.13 The nonlinear relationships between lnFxy(q,s) and lns in China’s agricultural futures
markets when polynomial order m = 3. Panel a, b, c and d illustrate the results for hard winter
wheat, soy meal, soybean and corn futures markets respectively
(a) (b)
(a) (b)
Fig. 4.14 The nonlinear price-volume relationships between lnFxy(q,s) and lns in US agricultural
futures markets when polynomial order m = 3. Panel a, b, c and d illustrate the results for hard
winter wheat, soy meal, soybean and corn futures markets respectively
records also reports that such trends have impacts on the detection of power-law
relationship (Podobnik et al. 2009; Hu et al. 2001).
Then we obtained the cross-correlation exponent hxy(q) and q (see Figs. 4.15 and
4.16). To make a comparison, we also estimated the generalized Hurst Exponent h
(q) of each individually analyzed futures market by means of MF-DFA (see
Figs. 4.15 and 4.16). If exponent h is a constant, the analyzed market is
monofractal, otherwise it is multifractal. From these plots we can find that the
price-volume relationships are nonlinear and multifractal; for different q, there are
different exponents h; that is, for different q, there are different power-law
cross-correlations.
(a) (b)
(c) (d)
Fig. 4.15 The nonlinear relationships between h(q) and q in China’s agricultural futures markets
when polynomial order m = 3. Panel a, b, c and d illustrate the results for hard winter wheat, soy
meal, soybean and corn futures markets respectively
The numerical results for the cross-correlation exponents and generalized Hurst
exponents when q = 2 are listed in Table 4.2. We also calculated the average of
generalized Hurst exponents, that is, ðhxx þ hyy Þ=2 (the pink curves in Figs. 4.15
and 4.16), which is an average of separately analyzed price and volume series in
China’s and US markets (the red curves and green curves in Figs. 4.15 and 4.16).
From Table 4.2, we find that if we simply judge the market efficiency by means of
one variable, we may be misled by two completely contradicted conclusions: (i) in
terms of prices, all the results of generalized Hurst exponents significantly
approximate to 0.5, which implies that the markets are seemingly efficient; but
(ii) in terms of trading volumes, those are significantly less than 0.5, which is the
clear evidence that the markets are more anti-persistent rather than perfectly effi-
cient. That, in our view, is one of the reasons why researchers frequently draw quite
contradicted conclusions over one specific market. If we simply draw conclusions
from one single measure, our conclusions may be biased and thereby systematic
errors may occur. But if we incorporate the two characteristic quantities, which are
two major measures of a certain market, all the bivariate cross-correlation
(a) (b)
(c) (d)
Fig. 4.16 The nonlinear relationships between h(q) and q in US agricultural futures markets when
polynomial order m = 3. Panel a, b, c and d illustrate the results for hard winter wheat, soy meal,
soybean and corn futures markets respectively
exponents are found to be less than 0.5, namely, all the analyzed markets are
anti-persistent rather than perfectly efficient. Our discussion may provide a way of
avoiding self-contradiction and systematic errors in market analyses by incorpo-
rating different aspects of specific markets.
By means of Eq. (4.9), the multifractal exponent sxy(q) is estimated (see
Figs. 4.17 and 4.18). From the figures one can find that s is nonlinearly dependent
on q, which is another piece of empirical evidence that nonlinear dependency and
multifractality exist in price-volume relationships for all the agricultural commodity
futures markets.
72
Table 4.2 Hurst exponents and multifractal spectra widths ⊿a

Country Market Variable h(2) ⊿a
m=1 m=2 m=3 m=1 m=2 m=3
China Wheat Price 0.5369 ± 0.0101 0.5224 ± 0.0109 0.4950 ± 0.0069 0.6300 0.6751 0.7795
Volume 0.2365 ± 0.0154 0.2757 ± 0.0127 0.3079 ± 0.0107 0.9046 0.8524 0.8332
Bivariate 0.3729 ± 0.0167 0.4233 ± 0.0137 0.4628 ± 0.0117 0.4049 0.3926 0.3204
Soy meal Price 0.5893 ± 0.0115 0.5878 ± 0.0126 0.5857 ± 0.0057 0.3488 0.4271 0.4039
Volume 0.2826 ± 0.0149 0.3362 ± 0.0122 0.3661 ± 0.0103 2.1212 2.1205 2.1008
Bivariate 0.4775 ± 0.0128 0.5134 ± 0.0136 0.5391 ± 0.0123 0.9772 1.0082 0.9388
Soybean Price 0.5151 ± 0.0164 0.4952 ± 0.0166 0.5286 ± 0.0141 0.6213 0.4840 0.6259
Volume 0.3127 ± 0.0168 0.3572 ± 0.0187 0.3565 ± 0.0225 2.6275 2.5617 2.5943
Bivariate 0.4668 ± 0.0147 0.4846 ± 0.0139 0.4997 ± 0.0147 1.1074 0.9907 1.0301
Corn Price 0.5084 ± 0.0164 0.4737 ± 0.0166 0.4762 ± 0.0159 0.4237 0.5393 0.5586
Volume 0.3176 ± 0.0194 0.3351 ± 0.0183 0.3131 ± 0.0169 4.3059 4.1631 4.1619
Bivariate 0.4724 ± 0.0232 0.4537 ± 0.0209 0.4347 ± 0.0236 2.1779 2.1440 2.2155
US Wheat Price 0.4462 ± 0.0085 0.4490 ± 0.0072 0.4433 ± 0.0062 0.3126 0.3598 0.4397
Volume 0.1557 ± 0.0077 0.1944 ± 0.0069 0.2246 ± 0.0060 0.8031 0.7840 0.7915
Bivariate 0.3167 ± 0.0071 0.3364 ± 0.0068 0.3488 ± 0.0065 0.5543 0.5195 0.5477
Soy meal Price 0.4939 ± 0.0045 0.5019 ± 0.0042 0.5075 ± 0.0042 0.3269 0.3908 0.4005
Volume 0.1296 ± 0.0088 0.1696 ± 0.0091 0.2029 ± 0.0089 0.6830 0.6617 0.6265
Bivariate 0.3172 ± 0.0055 0.3327 ± 0.0063 0.3482 ± 0.0068 0.4096 0.3946 0.3939
Soybean Price 0.5222 ± 0.0063 0.5309 ± 0.0056 0.5256 ± 0.0050 0.3393 0.3162 0.3881
Volume 0.1670 ± 0.0118 0.2233 ± 0.0119 0.2639 ± 0.0111 1.0592 1.0176 1.0121
Bivariate 0.3417 ± 0.0091 0.3754 ± 0.0094 0.3961 ± 0.0097 0.5562 0.5167 0.5554
Corn Price 0.5326 ± 0.0087 0.5544 ± 0.0083 0.5371 ± 0.0086 0.3605 0.3083 0.3465
Volume 0.1827 ± 0.0092 0.2291 ± 0.0077 0.2599 ± 0.0063 1.4027 1.3830 1.4257
Bivariate 0.3757 0.0083 0.4163 0.0057 0.4266 0.0059 0.7092 0.6583 0.7067
4 Multifractal Detrended Cross-Correlation Analysis (MF-DCCA)
± ± ±
(a) (b)
(c) (d)
Fig. 4.17 The nonlinear relationships between s(q) and q in China’s agricultural futures markets
when polynomial order m = 3. Panel a, b, c and d illustrate results for hard winter wheat, soy
To better describe the nonlinear dependency, we further investigate to multi-

fractal strength by means of multifractal spectra. First, all of the slopes for different
q and the multifractal spectra are estimated by means of Eqs. (4.10) and (4.11) (see
Table 4.2 and Figs. 4.19, 4.20). It is widely known that the multifractal spectrum of
monofractality is a point, namely, the width of multifractal spectrum is zero if the
system under study is monofractal. Actually the width of multifractal spectrum can
be regarded as an estimate of multifractal strength. The numerical results of the
widths are listed in Table 4.2. The widths of cross-correlation multifractal spectra
for price-volume relationships and individually analyzed price (volume) series in all
markets are significantly nonzero, which imply that there are clear departures from
random walk process for all the cases.
(a) (b)
(c) (d)
Fig. 4.18 The nonlinear relationships between s(q) and q in China’s agricultural futures markets
when polynomial order m = 3. Panel a, b, c and d illustrate results for hard winter wheat, soy
Based on the above-mentioned results, the nonlinear dependency and multi-

fractality can be clearly found between price and volume relationships in China’s
and US agricultural commodity futures markets. Therefore, a researcher or a
technical analyst can complement her understandings on market dynamics by
obtaining more comprehensive knowledge from the integrated point of view rather
than from separated quantities; and as the dependency exists, one can not simply
jump to conclusion which variable is the cause of the volatility of the other, since
they are mutually interacted.
(a) (b)
(c) (d)
Fig. 4.19 The multifractal spectra between China’s agricultural futures markets when polynomial
order m = 3. Panel a, b, c and d illustrate multifractal relationships between f(a) and a in hard
4.3.4 Conclusions
In this section, we discussed the nonlinear dependency between characteristic

market quantities (variables): trading volume and market price. To take China’s and
US agricultural commodity futures markets as the representatives of emerging and
developed economies respectively, we investigate the nonlinear dependency and
multifractality in the price-volume relationships by applying a new methodology
Multifractal Detrended Cross-Correlation Analysis (MF-DCCA), an algorithm to
analyze two spatially or temporally correlated time series. Our nontrivial empirical
findings can be summarized as follows:
First of all, there exists nonlinear dependency in the form of cross-correlations in
all price-volume relationships, which suggests that the price and volume are
mutually interacted.
Secondly, clear departures from zero are found for all of the bivariate
cross-correlations, which imply that that none of those markets, no matter in China
(a) (b)
(c) (d)
Fig. 4.20 The multifractal spectra between US agricultural futures markets when polynomial
order m = 3. Panel a, b, c and d illustrate multifractal relationships between f(a) and a in hard
or in US, are so-called “efficient markets” from the perspective of price-volume

relationships.
Thirdly, multifractality is significant in all the price-volume relationships mea-
sured by bivariate cross-correlations of agricultural futures markets in two
economies.
Furthermore, we also discuss the relationship between bivariate cross-correlation
exponent and the generalized Hurst exponent for time series of respective variables.
Our theoretical proof shows that if the scale s ! 1, bivariate cross-correlation
exponent is smaller than and equal to the average of the generalized Hurst expo-
nents of the two individually market quantities.
References 77
References

A.-L. Barabasi, P. Szepfalusy, T. Vicsek, Multifractal spectra of multi-affine functions. Phys. A:
Stat. Theor. Phys. 178, 17–28 (1991)
D.O. Cajueiro, B.M. Tabak, Testing for long-range dependence in world stock markets. Chaos,
Solitons Fractals 37, 918–927 (2008)
C.P. Cristescu, C. Stan, E.I. Scarlat, T. Minea, C.M. Cristescu, Parameter motivated mutual
correlation analysis: application to the study of currency exchange rates based on intermittency
parameter and Hurst exponent. Phys. A 391, 2623–2635 (2012)
D. Grech, Z. Mazur, Can one make any crash prediction in finance using the local Hurst exponent
idea? Phys. A 336, 133–145 (2004)
G.-F. Gu, W.-X. Zhou, Detrending moving average algorithm for multifractals. Phys. Rev. E 82
(1), 011136 (2010)
L.-Y. He, S.-P. Chen, A new approach to quantify power-law cross-correlation and its application
to commodity markets. Phys. A 390, 3806–3814 (2011a)
L.-Y. He, S.-P. Chen, Multifractal detrended cross-correlation analysis of agricultural futures
markets. Chaos, Solitons Fractals 44, 355–361 (2011b)
L.-Y. He, S.-P. Chen, Nonlinear bivariate dependency of price-volume relationships in agricultural
commodity futures markets: a perspective from Multifractal Detrended Cross-Correlation
Analysis. Phys. A 390, 297–308 (2011c)
L.-Y. He, F. Zheng, Detecting fractal/multifractal and asymmetric properties in an artificial
quote-driven financial market. Fractals 18, 87–99 (2010)
L. Hedayatifar, M. Vahabi, G.R. Jafari, Coupling detrended fluctuation analysis for analyzing
coupled nonstationary signals. Phys. Rev. E 84, 021138 (2011)
D. Horvatic, H.E. Stanley, B. Podobnik, Detrended cross-correlation analysis for non-stationary
time series with periodic trends. EPL 94, 18007 (2011)
K. Hu, P.C. Ivanov, Z. Chen, P. Carpena, H. Eugene, Stanley, Effect of trends on detrended
fluctuation analysis. Phys. Rev. E 64, 011114 (2001)
Z.-Q. Jiang, W.-X. Zhou, Multifractal detrending moving average cross-correlation analysis. Phys.
Rev. E 84, 016106 (2011)
L. Kristoufek, Rescaled range analysis and detrended fluctuation analysis: finite sample properties
and confidence intervals. AUCO Czech. Econ. Rev. 4, 315–329 (2010)
L. Kristoufek, Multifractal height cross-correlation analysis: a new method for analyzing
long-range cross-correlations. EPL 95, 68001 (2011)
K.P. Lim, Ranking market efficiency for stock markets: a nonlinear perspective. Phys. A 376, 445–
454 (2007)
T.D. Matteo, Multi-scaling in finance. Quant. Finance 7, 21–36 (2007)
G. Papaionnou, A. Karytinos, Nonlinear time series analysis of the stock exchange: the case of an
emerging market. Int. J. Bifurcat. Chaos 5, 1557–1584 (1995)
E.E. Peters, in Chaos and Order in Capital Markets: A New View of Cycles, Prices and Market
Volatility (Wiley, New Jersey, 1991)
B. Podobnik, I. Grosse, D. Horvati, S. Ilic, P. Ivanov, H. Ch, E. Stanley, Quantifying
(2009)
B. Podobnik, Z.-Q. Jiang, W.-X. Zhou, H.E. Stanley, Statistical tests for power-law
cross-correlated processes. Phys. Rev. E 84, 066118 (2011)
F. Schmitt, D. Schertzer, S. Lovejoy, Multifractal analysis of foreign exchange data. Appl.
Stochast. Models Data Anal. 15, 29–53 (1998)
F. Schmitt, D. Schertzer, S. Lovejoy, Multifractal fluctuations in finance. Int. J. Theor. Appl.

Finance 3, 361–364 (2000)
M. Small, C.K. Tse, Detecting determinism in time series: the method of surrogate data. IEEE
Trans. Circuits Syst. 1, Fundam. Theory Appl. 50 (2003)
J. Song, P.-J. Shang, Effect of linear and nonlinear filters on multifractal detrended
cross-correlation analysis. Fractals 19, 443–453 (2011)
H.E. Stanley, V. Plerou, Scaling and universality in economics: empirical results and theoretical
interpretation. Quant. Finance 1, 563–567 (2001)
B.M. Tabak, D.O. Cajueiro, Assessing inefficiency in euro bilateral exchange rates. Phys. A 367,
319–327 (2006)
J. Vasilescu, C.P. Cristescu, L. Belegante, Multifractal analysis of fluorescence Lidar time series of
Black Sea waters. J. Optoelectron. Adv. Mater. 12, 1414–1420 (2010a)
J. Vasilescu, L. Marmureanu, E. Carstea, C.P. Cristescu, Oil spills detection from fluorescence
Lidar measurements. UPB Sci. Bull. Ser. A 72, 149–154 (2010b)
R.T. Vassoler, G.F. Zebende, DCCA cross-correlation coefficient apply in time series of air
temperature and air relative humidity. Phys. A 391, 2438–2443 (2012)
Y. Wang, L. Liu, R. Gu, J. Cao, H. Wang, Analysis of market efficiency for the Shanghai stock
market over time. Phys. A 389, 1635–1642 (2010a)
Y. Wang, Y. Wei, C. Wu, Cross-correlations between Chinese A-share and B-share markets. Phys.
A 389, 5468–5478 (2010b)
Y. Wang, C. Wu, Z. Pan, Multifractal detrending moving average analysis on the US Dollar
exchange rates. Phys. A 390, 3512–3523 (2011)
Y. Yuan, X. Zhuang, Measuring multifractality of stock price fluctuation using multifractal
detrended fluctuation analysis. Phys. A 388, 2189–2197 (2009)
Y. Yuan, X. Zhuang, Z. Liu, Price-volume multifractal analysis and its application in Chinese
stock markets. Phys. A (2012). https://doi.org/10.1016/j.physa.2012.01.034
G.F. Zebende, DCCA cross-correlation coefficient: quantifying level of cross-correlation. Phys.
A 390, 614–618 (2011)
X.-J. Zhao, P.-J. Shang, A.-J. Lin, G. Chen, Multifractal Fourier detrended cross-correlation
analysis of traffic signals. Phys. A 390, 3670–3678 (2011)
X.-J. Zhao, P.-J. Shang, C. Zhao, J. Wang, R. Tao, Minimizing the trend effect on detrended
cross-correlation analysis with empirical mode decomposition. Chaos, Solitons Fractals 45,
166–173 (2012)
W.-X. Zhou, Multifractal detrended cross-correlation analysis for two nonstationary signals. Phys.
Rev. E 77, 066211 (2008)
W.-X. Zhou, The components of empirical multifractality in financial returns. EPL 88, 28004
(2009)
W.-X. Zhou, Finite-size effect and the components of multifractality in financial volatility. Chaos,
Chapter 5
Asymmetric Multifractal Detrended
Fluctuation Analysis (A-MFDFA)
The presence of multifractality suggests the inefficiency (Cajueiro and Tabak

2004a, b, 2008; Cajueiro et al. 2009; Tabak and Cajueiro 2007; Wang et al. 2010),
volatility predictability (Wei and Wang 2008), crash predictions (Wei and Wang
2008; Grech and Pamula 2008), and complexity (Matia et al. 2003; Kumar and Deo
2009; Norouzzadeh and Jafari 2005) of the market. Multifractal analysis has been
widely applied in stock markets (Czarnecki and Grech 2010; Jiang and Zhou 2008)
to investigate the intermittent nature of turbulence. Multifractal detrended fluctua-
tion analysis (MF-DFA) (Kantelhardt et al. 2002), which is based on detrended
fluctuation analysis (DFA) (Peng et al. 1994), is widely used to detect long-range
autocorrelations and multifractality in stock markets for nonstationary time series
(Eldridge et al. 1993; Greene and Fieltz 1997). Numerous studies have analyzed
multifractality in the Chinese stock market (Wei and Huang 2005; Jiang and Zhou
2008; Wang et al. 2009; Yuan and Zhuang 2009; Du and Ning 2008; Bai and Zhu
2010), but relatively few have focused on the asymmetry of multifractal scaling
behavior. Even though a small number of studies (Zhou et al. 2009) have focused
on this field, researchers usually investigate the different correlation features by
categorizing the period of stock markets into bull and bear, in which the sample
division is subjective. Thus, studying asymmetric correlation in the entire sample
interval is necessary. This study intends to fill this research gap.
Correlation asymmetry affects portfolio diversification, risk management, and
policymaking (Demirer 2003). The presence of asymmetric correlation in the stock
market is not surprising because of the expected asymmetric response to economic
news. Recent studies suggest that asymmetric correlations exist in stock returns
(Longin and Solnik 2001; Ang and Chen 2002; Bae et al. 2003). Ang and Bekaert
(2002) utilized a two-regime-switching model to determine the connection between
low returns and high correlation. Longin and Solnik define a new concept called
“exceedance correlation” and find a high correlation between large negative returns
and zero correlation between large positive returns. Ang and Chen use exceedance
correlation test and show that asymmetric correlation exists in different types of

80 5 Asymmetric Multifractal Detrended Fluctuation Analysis …
domestic portfolios. Hong and Zhou propose a model-free method and confirm that
asymmetry exists in the United States stock market. These methods can detect the
presence of asymmetric correlations. However, these methods depend on
assumptions, such as model and threshold value selection.
Alvarez-Ramirez et al. (2009) extended the DFA in 2009 by developing
asymmetric DFA (A-MFDFA) method to explore the asymmetries in the scaling
behavior of time series. We call the A-MFDFA method “A-MFDFA” in this chapter
because A-MFDFA can analyze multifractal scaling behavior. The A-MFDFA
method can not only assess multifractality in different correlations but can also
detect the asymmetry of scaling behavior in time series with uptrends and down-
trends. Alvarez-Ramirez et al. show that different scaling properties exist if the
signal trend is positive or negative. However, Alvarez-Ramirez et al. only show
intuitively the relationship between asymmetric behavior and intrinsic correlations.
They also discuss only the skewness of data distribution in their samples rather than
the origins of asymmetric behavior.
5.1 Methodology
5.1.1 A-MFDFA Method
Let fxðtÞg be time series t ¼ 1; 2; . . .; N, where N is the length of the series; thus,
A-MFDFA method can be summarized with the following steps.
Step 1: We construct the profile as
X
j
yðjÞ ¼ ðxðtÞ xÞ; j ¼ 1; 2; . . .; N; ð5:1Þ
t¼1
P
where x ¼ N1 Nt¼1 xðtÞ.
Step 2: We divide the time series fxðtÞg and its profile fyðtÞg into Nn ¼
intðN=nÞ non-overlapping subtime series of n lengths. A short part of the profile
will remain in most cases because the record length N does not have to be a multiple
of the considered time-scale n. This procedure is repeated starting from the other
end of the record to consider the remaining part of the record. Thus, 2Nn segments
are obtained. Let Sj ¼ fsj;k ; k ¼ 1. . .; ng be the jth subtime series of length n and
Yj ¼ fyj;k ; k ¼ 1; . . .; ng be the integrated time series (i.e., profile) in the jth time
interval, j ¼ 1; 2; . . .; 2Nn . Thus, in the jth segment, k ¼ 1; 2; . . .; n, we have
sj;k ¼ xððj 1Þn þ kÞ; yj;k ¼ yððj 1Þn þ kÞ; ð5:2Þ
for j ¼ 1; 2; . . .; Nn and
5.1 Methodology 81
sj;k ¼ xðN ðj Ns Þn þ kÞ; yj;k ¼ yðN ðj Ns Þn þ kÞ; ð5:3Þ
for j ¼ Ns þ 1; . . .; 2Nn . 5 n N=4 is traditionally selected based on the recom-

mendations of Peng et al. (1994).
Step 3: For each subtime series Sj ¼ fsj;k ; k ¼ 1. . .; ng and its profile time series
Yj ¼ fyj;k ; k ¼ 1; . . .; ng, we compute the corresponding local least-squares line fits
LSj ðkÞ ¼ asj þ bsj k and LYj ðkÞ ¼ aYj þ bYj k, where k represents the horizontal
coordinate and i ¼ 1; 2. The fits LSj ðkÞ and LYj ðkÞ represent the linear trends for the
jth subtime series Sj and its integrated time series Yj , respectively. The linear fit
LSj ðkÞ is used only to discriminate via slope bSj whether the trend of the subtime
series Sj is positive or negative. The linear fit LYj ðkÞ is used to detrend the integrated
time series Yj . We determine the fluctuation functions
1X n
Fj ðnÞ ¼ ðyj;k LYj ðkÞÞ2 ð5:4Þ
n k¼1
for each segment j ¼ 1; 2; . . .; 2Nn .

Step 4: To assess asymmetric cross-correlation scaling properties, the average
fluctuation functions are considered in cases in which the time series xðtÞ has
piecewise positive and negative linear trends. This trend discrimination is made by
using the sign of the slope bSj ; that is, bSj [ 0 (resp. bSj \0) indicates that the time
series xðtÞ has a positive (resp. negative) trend in the subtime series Sj .
The directional q-order average fluctuation functions are computed by
!1=q
1 X2Nn
signðbSj Þ þ 1
Fqþ ðnÞ ¼ ½Fj ðnÞq=2 ; ð5:5Þ
M þ j¼1 2
!1=q
1 X2Nn
½signðbSj Þ 1
Fq ðnÞ ¼
½Fj ðnÞq=2 ; ð5:6Þ
M j¼1 2
P n signðbSj Þ þ 1 P n ½signðbSj Þ1

where M þ ¼ 2N j¼1 2 and M ¼ 2Nj¼1 2 are the number of sub-
time series with positive and negative trends, respectively. Assuming that bSj 6¼ 0
for all j ¼ 1; 2; . . .; 2Nn , then M þ þ M ¼ 2Nn .
The traditional MF-DFA is similarly performed by computing the average
fluctuation function
!1=q
1 X2Nn
Fq ðnÞ ¼ ½Fj ðnÞq=2 : ð5:7Þ
2Nn j¼1
Therefore, if power-law cross-correlations exist, then the scaling or power-law

relationship satisfies
þ
Fq ðnÞ nHðqÞ ; Fqþ ðnÞ nH ðqÞ
; Fq ðnÞ nH ðqÞ
; ð5:8Þ
where HðqÞ, H þ ðqÞ, and H ðqÞ respectively denote the overall, upward, and
downward scaling exponents. The scaling behavior of the fluctuations in Eq. (5.8)
is determined by analyzing the log–log plots of Fq ðnÞ, Fqþ ðnÞ, and Fq ðnÞ versus
n for each value of q. Thus, Eq. (5.8) can be presented as
log Fq ðnÞ ¼ HðqÞ logðnÞ þ log A1 ; ð5:9Þ
log Fqþ ðnÞ ¼ H þ ðqÞ logðnÞ þ log A2 ; ð5:10Þ
log Fq ðnÞ ¼ H ðqÞ logðnÞ þ log A3 : ð5:11Þ
The overall (resp. upward or downward) scaling exponents HðqÞ (resp. H þ ðqÞ
or H ðqÞ) can be obtained by observing the slope of the log–log plot of Fq ðnÞ
(resp. Fqþ ðnÞ or Fq ðnÞ) versus n through ordinary least squares method. The
scaling exponent HðqÞ is the generalized Hurst exponent. If Hð2Þ ¼ 0:5, then the
correlations in the time series are persistent, implying that an increment has more
chances of being followed by another increment. If Hð2Þ\0:5, then the correlations
in the time series are anti-persistent, indicating that an increment has more chances
of being followed by a decrement and vice versa. If Hð2Þ ¼ 0:5, only short-range
correlations or no correlations exist.
If H þ ðqÞ ¼ H ðqÞ, then the correlations in the time series are symmetrical. By
contrast, if H þ ðqÞ 6¼ H ðqÞ, then the correlations in the time series are asym-
metrical, which means that the correlations in the time series are different during
positive and negative trends. Moreover, if H þ ðqÞ 6¼ HðqÞ, which follows Greene
and Fieltz (1997), then H ðqÞ 6¼ HðqÞ and vice versa. Additionally, if
HðqÞ 6¼ H ðqÞ, then H þ ðqÞ 6¼ H ðqÞ holds, which means that the positive and
negative directions of the time series have different scaling exponents.
For q [ 0, the variables HðqÞ, H þ ðqÞ, and H ðqÞ respectively describe the
overall, upward, and downward scaling behavior of large fluctuations, which are
usually characterized by small scaling exponents for multifractal time series. By
contrast, for q\0, the variables HðqÞ, H þ ðqÞ, and H ðqÞ respectively describe the
overall, upward, and downward scaling behavior of small fluctuations. In addition,
if the scaling exponent HðqÞ (H þ ðqÞ or H ðqÞ) depends significantly on q, then the
correlations in the overall (resp. positive trending or negative trending) time series
are multifractal. By contrast, if HðqÞ (resp. H þ ðqÞ or H ðqÞ) does not depend
significantly on q, then the correlations in the overall (positive trending or negative
trending) time series is monofractal.
The Renyi exponent, sðqÞ, which is related to the general Hurst exponent HðqÞ
obtained from MF-DFA, can be expressed as
5.1 Methodology 83
sðqÞ ¼ qHðqÞ 1: ð5:12Þ
With the Legendre transform, the singularity strength a (H€

older exponent) and
its spectrum f ðaÞ can be calculated by

a ¼ HðqÞ þ qH 0 ðqÞ;
ð5:13Þ
f ðaÞ ¼ q½a HðqÞ þ 1;
where H 0 ðqÞ represents the derivative of HðqÞ with respect to q, and the singularity
spectrum f ðaÞ describes the singularity content of the time series. In Eqs. (5.12) and
(5.13), by replacing HðqÞ with H þðqÞ (resp. H ðqÞ), we will obtain the singularity
strength a and singularity spectrum f ðaÞ for the time series with uptrends
(resp. downtrends).
5.1.2 Asymmetric GARCH Model
Zakoian and Glosten et al. implemented an extension of the Generalized

AutoRegressive Conditional Heteroskedasticity (GARCH) model and proposed the
threshold ARCH (TARCH) model. The specific form of TARCH (p, q) is expressed
as follows:
X
q X
p
r2 ¼ a0 þ ai e2ti þ ce2t1 dt1 þ kj r2tj ð5:14Þ
i¼1 j¼1
(
1; et \0
where dt is a nominal variable dt ¼ . When dt is introduced, the effect of
0; et 0
price increase (et 0, good news) and price decrease (et \0, bad news) on the
conditional variance is different. If c 6¼ 0, then the news is asymmetrically
influential. If c [ 0, then bad news has a more significant effect on price fluctuation
than good news. If c\0, then good news has a more significant effect on price
fluctuation than bad news.
Nelson further analyzed and extended the GARCH model and proposed the
Exponential GARCH (EGARCH) model. This model intends to depict the asym-
metrical response of the conditional variance r2t to positive and negative interfer-
ence in the market. In this case, the conditional variance r2t is an asymmetric
function of the delayed interference term eti . The specific form of the EGARCH
(p, q) model is expressed as follows:
X X X
q
eti
q eti p
logðr2t Þ ¼ a0 þ ai þ
ci l þ kj logðr2tj Þ; ð5:15Þ
rti rti
i¼1 i¼1 j¼1
In the present study, the conditional variance in the model adopted the natural
logarithmic form, which indicates that the leverage effect is exponential. The
implication of ci is the same as that of parameter c in Eq. (5.14). The k1 is the
influence coefficient of r2tj .
5.2 Empirical Analysis on Shanghai-Shenzhen

Stock Market
In this section, we aims to achieve three objectives. First, we apply A-MFDFA to

explore the existence of asymmetric multifractal scaling behavior in the Shanghai
and Shenzhen stock markets. Second, we discuss the source of asymmetric mul-
tifractal scaling behavior by comparing the estimated results of the original data
with those of shuffled and surrogated data. Third, we apply rolling windows method
to investigate the time-varying feature of scaling asymmetries in the Chinese stock
market.
5.2.1 Data
Only two stock exchanges are present in China: the Shanghai Stock Exchange and
the Shenzhen Stock Exchange. Thus, we choose the daily closing price series of the
Shanghai Stock Exchange Composite Index (SSCI) and the Shenzhen Component
Index (SZCI) as original data to represent the Shanghai and Shenzhen stock mar-
kets, respectively. The Shanghai Stock Exchange started to operate in 19 December
1990, and the Shenzhen Stock Exchange started to operate in 3 April 1991.
Therefore, the sample period for SSCI is from 19 December 1990 to 27 April 2012,
and the sample period for SZCI is from 2 April 1991 to 27 April 2012. The original
sample data were obtained from FC Station 3.0 (http://icbc-zx1.finchina.com/
product-2.asp).
Traditionally, we select the logarithmic return rt ¼ logðPt Þ logðPt1 Þ as
empirical data, where Pt is the closing price index at time t, and logðÞ denotes the
nature logarithm. Thus, the return series of SSCI and SZCI is the empirical data
numbered 5229 and 5139, respectively.
The returns of SSCI and SZCI (also denoted as SSCI and SZCI) for the entire
sample period are shown in Fig. 5.1, whereas their descriptive statistics are pre-
sented in Table 5.1.
Figure 5.1 shows that fluctuation persists in the return series either for SSCI or
SZCI. The skewness and kurtosis values in Table 5.1 are respectively larger than 0
and 3, thus indicating that the distributions of SSCI and SZCI are all fat tailed and
peaked. Jarque–Bera (Bera and Jarque 1981) statistics for normality test suggest
that the normality assumption of returns can be rejected either for SSCI or SZCI.
5.2 Empirical Analysis on Shanghai-Shenzhen Stock Market 85
0.8
SSCI
0.6
0.4
r
0.2
-0.2
0 1000 2000 3000 4000 5000 6000
0.4
SZCI
0.2
0
r
-0.2
-0.4
0 1000 2000 3000 4000 5000 6000
Sample points
Fig. 5.1 Daily return series of SSCI and SZCI
The Jarque–Bera Q statistics with different lags are all significant at 1% significant
level. This result indicates that all series may present long-range correlations.
5.2.2 Empirical Results
Figure 5.2 presents two plots of the results estimated by A-MFDFA for SSCI. The
top figure in Fig. 5.2 shows the plot for log(F2(n)) versus log(n), whereas the bottom
figure provides the plot for F2(n) versus n. The top figure shows that the deviation
from symmetry for the time scales are higher than 200 days, particularly for SZCI.
However, the combined top and bottom figures in Fig. 5.2 show a symmetric pinch
for a time scale of about 600 days (i.e., n = 600 in the bottom figure and log
(n) = 2.778 in the top figure) after which a sustained deviation from symmetry is
evident. The same result can be obtained for SZCI. Figure 5.3 shows two plots of the
results estimated with A-MFDFA for SZCI. Figure 5.3 shows the time scale after a
sustained deviation from symmetry is evident, which is also about 600 days.
In addition, if Df ¼ log F2þ ðnÞ log F2 ðnÞ, then we can use Df to measure the
fluctuation function asymmetry. Figures 5.4 and 5.5 show the plots of Df versus
n for SSCI and SZCI, respectively. The values of Df fluctuate around 0 before
600 days but decrease sharply on the 600th day. Afterward, the values of Df
fluctuate between −0.1 and −0.2. These results agree with those of Figs. 5.2 and
5.3. Figures 5.4 and 5.5 also show that a cycle period of about 200 days exists for
both SSCI and SZCI. Thus, the Hurst exponent Hð2Þ changes approximately every
200 days.
86
Table 5.1 Descriptive statistics for the returns of SSCI and SZCI
Mean Max Min S.D Ske Kur J–B Q(3) Q(15) Q(30)
SSCI 0.0006 0.7192 −0.1791 0.0251 5.39 145.1 442479*** 20.68*** 60.77*** 94.04***
*** *** ***
SZCI 0.0005 0.2327 −0.2204 0.0230 0.42 16.47 38990 24.64 63.75 94.22***
Note *, ** and *** denotes 10%, 5% and 1% significant level respectively; “Max,” “Min,” “S.D,” “Ske,” and “Kur” denote “maximum,” “minimum,”
“standard deviation,” “skewness,” and “kurtosis,” respectively. J–B denotes Jarque–Bera statistics, and Q(i) denotes the value of Jarque–Bera Q statistics with
i lags
5 Asymmetric Multifractal Detrended Fluctuation Analysis …
(a)
0
overall
-0.5 upwards
log10(F2(n))
downwards
-1
-1.5 log10(n)=2.778
-2
1 1.5 2 2.5 3 3.5
log10(n)
(b)
0.4
0.3
n=205
F2(n)
0.2
0.1
0
0 200 400 600 800 1000 1200 1400
n
Fig. 5.2 Plots for SSCI: a log(F2(n)) versus log(n); b F2(n) versus n. Notes log 10ðÞ denotes the
common (base 10) logarithm. The common (base 10) logarithm is written as logðÞ in the text,
including figure captions, unless specified otherwise
(a)
0
overall
-0.5
upwards
log10(F2(n))
downwards
-1
-1.5 log10(n)=2.778
-2
1 1.5 2 2.5 3 3.5
log10(n)
(b)
0.5
0.4
n=600
0.3
F2(n)
0.2
0.1
0
0 200 400 600 800 1000 1200 1400
n
Fig. 5.3 Plots for SZCI: a Plots of log(F2(n) versus log(n); b plots of F2(n) versus n
Figure 5.6 plots the values of the generalized Hurst exponents HðqÞ, H þ ðqÞ,
and H ðqÞ versus q from −6 to 6 (interval of 0.1) to assess the multifractality of the
Chinese stock market with different trends. For all series, the values of HðqÞ,
H þ ðqÞ and H ðqÞ decrease with increases in q, a result that suggests gradually
Fig. 5.4 Plots of Df for SSCI
0.4
0.3
0.2
0.1
Df
-0.1
-0.2
-0.3
-0.4
0 200 400 600 800 1000 1200 1400
sample points
Fig. 5.5 Plots of Df for SZCI
SSCI SZCI
1 1
overall overall
0.9 upwards 0.9 upwards
downwards downwards
0.8 0.8
0.7 0.7
H(q)
H(q)
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
q q
Fig. 5.6 Plots of HðqÞ, H þ ðqÞ and H ðqÞ versus q (q ¼ −6, −5.9, …, 5.9, 6)
SSCI SZCI
1.5 1.5
1
1
0.5
0 0.5
-0.5
f(alpha)
f(alpha)
0
-1
-1.5 -0.5
-2
overall -1 overall
-2.5 upwards upwards
downwards downwards
-3 -1.5
0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
alpha alpha
Fig. 5.7 f ðaÞ versus a. (q ¼ −5.5, −5, …, 5, 5.5)
weakened correlations for both up and downtrends. Moreover, either for SSCI or
for SZCI, the departure degree of the general Hurst exponents for upward trends
and those for downward trends is larger for q [ 0 than for q\0. This finding
indicates that the correlation asymmetry is stronger for large fluctuations than for
small fluctuations in the Chinese stock market.
Figure 5.7 shows the multifractal spectra f ðaÞ versus a. The inverse parabolic shapes
of the spectra can be observed in Fig. 5.7, which confirms that multifractality exists in the
Chinese stock market, particularly both in SSCI and in SZCI. The negative generalized
fractal dimension (f ðaÞ\0) can be seen in the left parts of the spectra for both SSCI and
SZCI. This phenomenon, which can be interpreted mathematically in (Zhou and Yu
2001), is evident in many econophysics examples (Wang et al. 2011a, b). This result
implies that Chinese stock markets with large fluctuations are random (Mandelbrot
1990, 1991; Wang et al. 2011a, b). Furthermore, the singularity widths (the precise
values of Da are presented in Table 5.2) for uptrends for SSCI and SZCI are larger than
those for downtrends, particularly for SSCI. The multifractality of the Chinese stock
market with uptrends is stronger than those with downtrends; the asymmetry of
multifractality in the Shanghai stock market is stronger than that in the Shenzhen stock
market. This observation can be confirmed by Fig. 5.6 in which the difference in the
downward slopes of H þ ðqÞ −q and H ðqÞ −q is larger in SSCI than in SZCI.
5.2.3 Discussion
5.2.3.1 Origin of Multifractality with Different Trends
Two major sources of multifractality are traditionally present: (a) different long-range
correlations for small and large fluctuations, and (b) fat-tailed probability distributions
Table 5.2 DH of the original, shuffled, and surrogated series

SSCI SZCI
Original Shuffled Surrogated Original Shuffled Surrogated
series series series series series series
Overall 0.4035 0.1549 0.1117 0.2544 0.0861 0.0948
(61.61%) (72.32%) (66.16%) (62.74%)
Upward 0.5906 0.1548 0.2531 0.3211 0.0685 0.1339
(73.79%) (57.15%) (78.67%) (58.30%)
Downward 0.3413 0.1571 0.0657 0.2352 0.1542 0.0697
(53.97%) (80.75%) (34.44%) (70.37%)
Note The value in parentheses is the change (in percentage) in the DH value for the shuffled
(resp. surrogated) data to that of the original data, namely, ðDHorig DHshuf Þ=DHorig
(resp. ðDHorig DHsurr Þ=DHsurr ). q ¼ −5.5, −5, …, 5, 5.5
of variations (Norouzzadeh and Rahmani 2006). We can compare the multifractality

between original series and randomly shuffled series to understand the effect of
long-range correlations. The process of shuffling is described as follows:
Generate pairs (p, q) of random integer numbers with p; q N, where N is the
length of the
(I) Time series to be shuffled.
(II) Swap entries p and q.
(III) Repeat Steps (I) and (II) for 20 N times.
We can compare the multifractalities between the original series and the sur-
rogated series to investigate the contribution of the fat-tailed distribution. Fourier
phase randomization is widely used. However, the simulation results in
Norouzzadeh and Rahmani (2006) show that the long-range power-law correlations
in volatilities may completely disappear after phase randomization. Therefore, we
applied a simple method in Zhou (2009) to generate the surrogate data with
Gaussian distribution while maintaining the linear correlation of the original data
(Lim et al. 2007; Norouzzadeh and Rahmani 2006). The algorithm for this method
is described as follows.
For the Gaussian distribution, we generate a sequence of random numbers
fx0 ðtÞ : t ¼ 1; 2; . . .; Ng and then rearrange fx0 ðtÞ : t ¼ 1; 2; . . .; Ng to the rear-
ranged series fxðtÞ : t ¼ 1; 2; . . .; Ng, which has the same rank ordering as the
original series frðtÞ : t ¼ 1; 2; . . .; Ng. Thus, xðtÞ should have a rank n in the series
fxðtÞ : t ¼ 1; 2; . . .; Ng if and only if rðtÞ has a rank n in the original series frðtÞ :
t ¼ 1; 2; . . .; Ng (Bogachev et al. 2007; Zhou 2008). We employ the following
measure (Cao et al. 2012) to quantify the degree of multifractality:
DH ¼ maxðHðqÞÞ minðHðqÞÞ: ð5:16Þ

The greater the DH is, the stronger the degree of multifractality becomes. For
convenience of description, we use DHorig , DHshuf , and DHsurr to denote DH for the
original series, shuffled series, and surrogate series, respectively.
Table 5.2 provides the multifractality degrees of the original, shuffled, and
surrogated series for SSCI and SZCI. The values of DHshuf and DHsurr are smaller
than those of DHorig for SSCI and SZCI. This result means that the multifractal
scaling behavior in the Chinese stock market is caused not only by long-range
temporal correlation but also by fat-tailed distribution. However, the contributions
of these two factors are different in the multifractal scaling of each series with
different trends. On one hand, DHsurr is smaller than DHshuf for SSCI, whereas
DHsurr is larger than DHshuf for SZCI. This relationship indicates that the multi-
fractality in the Shanghai stock market is more attributed to fat-tailed distribution,
whereas the multifractality in the Shenzhen stock market is more attributed to
long-range correlation. On the other hand, DHshuf is smaller than DHsurr for SSCI
and SZCI with uptrends. Thus, the main source of multifractality in the Chinese
stock market with uptrends is long-range correlation. When the stock market is
going down either for SSCI or for SZCI, DHsurr is smaller than DHshuf . This
phenomenon implies that the main source of multifractality in the Chinese stock
market with downtrends is fat-tailed distribution.
5.2.3.2 Source of the Asymmetry
Alvarez-Ramirez et al. (2009) suggest that asymmetric scaling behavior may be

induced by either intrinsic correlation or fat-tailed distribution. Therefore, we use
the same method as that applied in Sect. 4.2 to investigate the source of asymmetry
scaling behavior.
Similar to that done in, the measure DH ðqÞ is defined as follows to quantify the
asymmetric degree of correlation:
DH ðqÞ ¼ jH þ ðqÞ H ðqÞj: ð5:17Þ
For a fixed q, the larger the DH ðqÞ is, the stronger the degree of asymmetry
becomes. If the value of DH ðqÞ is not far from zero, then the correlations are
symmetric for the time series with different trends.
Figures 5.8 and 5.9 present the values of DH ðqÞ for the original, shuffled, and
surrogated data for SSCI and SZCI, respectively. Either for SSCI or for SZCI, the
values of DH ðqÞ for the shuffled and surrogated data are smaller than those of the
original data. The only exception is the surrogated data for SSCI when q 2:5,
where the values of DH ðqÞ are larger than the values of DH ðqÞ for the original
SSCI data. Thus, fat-tailed distribution is not the source of the asymmetric scaling
behavior of the small fluctuations (q 2:5) in the Shanghai stock market,
although fat-tailed distribution is one of the sources of the asymmetric scaling in
larger fluctuations (q [ 2:5). However, long-range correlation is one of the
SSCI
0.25
original data
shuffled data
0.2 surrogated data
dalta(H+)
0.15
0.1
0.05
0
-6 -4 -2 0 2 4 6
q
Fig. 5.8 Plots of DH ðqÞ for the original, shuffled, and surrogated data for SSCI (q ¼ −5.5, −5,
…, 5, 5.5)
SZCI
0.18
original data
0.16
shuffled data
surrogated data
0.14
0.12
dalta(H+)
0.1
0.08
0.06
0.04
0.02
0
-6 -4 -2 0 2 4 6
q
Fig. 5.9 Plots of DH ðqÞ for the original, shuffled, and surrogated data for SZCI (q ¼ −5.5, −5,
…, 5, 5.5)
sources of asymmetry for SSCI and for SZCI. The values of DH ðqÞ for the shuffled
SSCI data are smaller than those for the surrogated SSCI data, whereas the values of
DH ðqÞ for the surrogated SZCI data are smaller than those for the shuffled SZCI
data. This observation suggests that the main source of asymmetric scaling behavior
in SSCI is more related to long-range correlation, whereas the main source of
asymmetric scaling behavior in SZCI is more related to fat-tailed distribution.
The above results on the source of asymmetry in the Chinese stock market
confirm the conclusions in Eldridge et al. (1993) that the source of asymmetric
scaling behavior may be induced by different factors, namely, long-range correla-
tion and fat-tailed distribution.
5.2.3.3 Time-Varying Feature of Asymmetry
Rolling windows method is often used to investigate the time-varying feature of

scaling behavior. Cristescu et al. (2012) suggest that structure functions and sin-
gular measure algorithms have possible applications with reliable results for time
series with window lengths of slightly more than 800 points. However, the results
are not reliable for shorter data sets with window lengths of less than 700 points.
This hypothesis is used explicitly or implicitly in relevant studies on financial time
series. The universal multifractal hypothesis may be misleading when short time
series are used. Therefore, we applied rolling windows method in data sets with a
window length of 1000, which corresponds to approximately four years. The
purpose is to investigate more precisely the time-varying feature of asymmetric
scaling behavior in the Chinese stock market.
Figure 5.10 provides the plots of the asymmetry degree DH ð2Þ with a slide
window of 10 days. Figure 5.10 shows that the scaling asymmetries in SSCI and
SZCI have time-varying features. Furthermore, the trend of evolution of the
asymmetry degree DH ð2Þ in SSCI is almost the same as that in SZCI. This finding
indicates that the time-varying features of the scaling asymmetries in SSCI and
SZCI are similar. Thus, external shocks exert almost the same effect on the
Shanghai stock market as on the Shenzhen stock market.
Figure 5.10 shows that the large fluctuation of the asymmetry degree is related to
major events, such as the Asian financial crisis (June 1997 to July 1998), the
introduction of the qualified foreign institutional investors (QFII) (July 2003), the
Chinese split-share structure reform (started on May 2005), and the global financial
crisis (October 2007 to November 2008). Major events may significantly enhance
the asymmetry of the scaling exponents in the Chinese stock markets (both for SSCI
and for SZCI). Hence, we suggest that scaling asymmetries in the Chinese stock
market (both SSCI and SZCI) may increase significantly because of the dramatic
changes in the external economic and financial environment.
5.2.4 Conclusions
We first applied A-MFDFA method in this chapter to investigate asymmetric

multifractal scaling behavior in the Shanghai and Shenzhen stock markets in China.
The sources of multifractality and asymmetry for these two markets are studied.
Results show that the price memory in the Chinese stock market is longer when the
market is going down than when it is going up. This finding may be different from
0.4
SSCI
dalta(H+) 0.3
0.2
0.1
0
0 50 100 150 200 250 300 350
1997 2003 2005 2007-2008

-1998 QFII Reoform Global
0.4 Asian of financial
financial split crisis SZCI
0.3 crisis share
structure
dalta(H+)
0.2
0.1
0
0 50 100 150 200 250 300 350
Sample points
Fig. 5.10 Time evolution of DH ð2Þ with a slide step of 10 days. Note The NaN results of
DH ð2Þ were removed from Fig. 5.10 because of the missing Hurst exponent. The Hurst exponent
is not calculated because of the insufficient data of uptrends in the SSCI (73 slide sample intervals)
and SZCI (66 slide sample intervals). In addition, the time-varying values of DH ð2Þ for SZCI
have an average lag of four days for SSCI because of the difference in sample length. This lag
exerts little effect on the results observed in Fig. 5.10
that of Zhou et al., who suggest that the bull market shows stronger long-term
correlation than the bear market. However, we obtain the same result as Bai and
Zhu (2010), who suggest that the correlation of the Shenzhen stock market is
stronger than that of the Shanghai stock market. The following is a summary of our
results:
(I) The Chinese stock market demonstrates multifractal scaling behavior, but the
multifractality degree for uptrends is stronger than that for downtrends. The
correlation asymmetries are stronger for large fluctuations than for small
fluctuations.
(II) Either long-range correlation or fat-tailed distribution induces scaling mul-
tifractality in the Chinese stock market. However, the main source of mul-
tifractal scaling behavior is more related to long-range correlation when the
market is going up, whereas the main source of multifractal scaling behavior
is more related to fat-tailed distribution when the market is going down.
(III) The main sources of scaling multifractality and scaling asymmetry in the
Shanghai stock market are related to fat-tailed distribution and long-range
correlations, respectively. The scaling multifractality and scaling asymmetry
in the Shenzhen stock market are induced by long-range correlation and
fat-tailed distribution, respectively.
(IV) The scaling asymmetries in SSCI and SZCI possess time-varying features,
and the trends of evolution of these scaling asymmetries are almost similar.
The scaling asymmetry in the Chinese stock market (both for SSCI and for
SZCI) may increase significantly because of dramatic changes in the external
economic and financial environment, such as financial crises, reforms in the
stock market, and QFII.
This section studies the nonlinear mechanism of the stock market from a new
perspective. The presence of scaling asymmetry suggests that the mechanisms
underlying a complex system act in a different way on the basis of the direction of
the dynamics. However, additional studies are needed to confirm or improve the
conclusions in our research. Barunik et al. (2012), Podobnik and Stanley (2008)
show that a finite-size effect in our empirical analysis may be present, and addi-
tional data are needed to obtain reliable results because a window length of 5000,
which is the length of the time series in our study, is not short. However, this
duration is not long enough to eliminate the finite-size effect. Moreover, asymmetric
MF-DCCA method can be proposed in further research as an extension of the
detrended cross-correlations method to study the relationship of multifractality and
long-range cross-correlations when the market is going up and down.
5.3 Empirical Analysis on International Gold Markets
In the present study, we applied the A-MFDFA method in the empirical analysis of
the asymmetric multifractal features of the Chinese and international gold markets.
Compared with existing studies, the aim of our work is threefold. First, we applied
the A-MFDFA method in the research field of gold market complexity for the first
time and presented the parameter (the asymmetric degree of multifractals) for the
A-MFDFA method. Second, the sliding window technology was applied to analyze
the multifractal features of the Chinese and international gold markets and the
time-varying feature of the asymmetry to reveal the evolution patterns of the
asymmetric multifractal features of the gold market. Third, we comprehensively
applied the threshold autoregressive conditional heteroskedasticity (TARCH) and
exponential generalized autoregressive conditional heteroskedasticity (EGARCH)
models in the empirical analysis of the asymmetric effects of good and bad news on
the Chinese and international gold markets to obtain consistent findings.
5.3.1 Descriptive Statistics Analysis of Gold Price
We selected the daily closing price of the Shanghai gold spot price (Au 99.95) in
the Shanghai Gold Exchange and the international gold spot price in the New York
Gold Exchange as representatives of the Chinese and international gold prices,
0.05 0.05
0.04 0.04
SHCP INCP
0.03 0.03
0.02 0.02
0.01 0.01
0 0
-0.01 -0.01
-0.02 -0.02
-0.03 -0.03
-0.04 -0.04
-0.05 -0.05
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
Fig. 5.11 Return series of the Shanghai and international gold prices (SHCP and INCP denote the
return rates of the Shanghai and international gold prices, respectively)
Table 5.3 Descriptive statistics for the returns of Shanghai and international gold prices
Mean Median S.D Ske Kur J–B ADF
SHCP 0.00057 0.00108 0.0117 −0.50 10.11 4880.66*** −50.11***
INCP 0.00057 0.00052 0.0127 −0.40 8.98 3382.57*** −48.49***
Note ***denotes 1% significant level; Symbols “S.D”, “Ske”, “Kur” denote Stv.Dev, Skewness
and Kurtosis respectively; The J–B denotes Jarque–Bera statistics 34. ADF denotes the
Augmented Dicky-Fuller (ADF) test 35 for unit root
respectively. The interval of the samples started from January 2, 2003 and ended on
April 27, 2012. A total of 2273 transaction days were included. The computational
equation of the daily return rate of gold price is rt ¼ logðpt Þ logðpt1 Þ, where pt
and pt−1 are the prices of the tth and (t − 1)th terms, respectively. The return rates
of the Shanghai and international gold prices are denoted by SHCP and INCP,
respectively. Thus, the actual sample size of the empirical calculation includes 2272
daily return rates. All data were obtained from the Shanghai Gold Exchange and
New York Gold Exchange.
The variation trends of the return rates of the Shanghai and international gold
prices are shown in Fig. 5.11 and their descriptive statistics are presented in
Table 5.3.
As shown in Fig. 5.11 and Table 5.3, the kurtosis of the Shanghai and inter-
national gold return series was 10.11 and 8.97, respectively, which are larger than 3.
These values indicate that a “sharp peak” existed. The skewness degrees of the
Shanghai and international gold return series were −0.50 and −0.40, respectively,
which are different from 0. These values indicate that a “fat rear” existed. Moreover,
the J–B statistical amounts of the return rates of the Shanghai and international gold
prices were 4880.66 and 3382.57, respectively. At the significance level of 5%, the
null hypothesis of the Gaussian distribution was denied, which indicates that the
return rate was not subject to Gaussian distribution. The Augmented Dickey-Fuller
(ADF) detection results showed that the t statistics of the return series of the
Chinese and international gold prices were −50.11 and −48.49, respectively. At the
5.3 Empirical Analysis on International Gold Markets 97
significance level of 1%, the null hypothesis of the “existence of unity roots” was
denied. This finding indicates that the daily return series of the Chinese and
international gold prices is stable. Thus, the Generalized Autoregressive
Conditional Heteroskedasticity model (GARCH) was used to explain the variation
of the return series.
5.3.2 Analysis of Asymmetric Scaling Behavior
5.3.2.1 Asymmetry of the Fluctuation Function
The changes in the estimated fluctuation function, namely, n and Fq(n), as well as
their double logarithmic diagrams, in the Shanghai and international gold markets
based on the A-MFDFA method are shown in Figs. 5.12 and 5.13, respectively.
Each figure is composed of two subgraphs (referred to as the upper and lower
figures). The upper figure is a double logarithmic diagram (the x coordinate as
log10(n) and the y coordinate as log10(F2(n); log10() is a logarithm with 10 as the
base). The lower figure shows changes in Fq(n) with n.
As shown in Figs. 5.12 and 5.13, the fluctuation functions of gold return rates in
the Shanghai and international gold markets share similar features. The upper
figures of Figs. 5.12 and 5.13 show that the timescale of the fluctuation function
deviating from the symmetry in the increasing and decreasing trends was over
200 days, i.e., log10(n) = 2.30. The asymmetry lasting approximately 200 days
had a certain regime switching feature. Between 200 and 400 days, i.e., log10
(n) = 2.60, the fluctuation function value in a decreasing trend was generally larger
than that in an increasing trend of gold price. After 400 days, the fluctuation
function value in an increasing trend was generally larger than that in a decreasing
(a)
-1
overall
log10 F (n))
-1.5 upwards
2
downwards
-2
log10(n)=2.6
-2.5
1.4 1.6 1.8 2 2.2 2.4 2.6 2.8
(b) log10 n
0.04
overall
upwards
downwards
F (n)
0.02
2
n=400
0
0 100 200 300 400 500 600
n
Fig. 5.12 Plots for SHCP: a Plots of log(F2(n)) versus log(n) and b plots of F2(n) versus n. Note
In the figure, log 10ðÞ denotes the common (base 10) logarithm.
(a)
-1
log10(F (n)) overall
-1.5 upwards
2
downwards
-2
log10(n)=2.6
-2.5
1.4 1.6 1.8 2 2.2 2.4 2.6 2.8
(b) log10(n)
0.04
overall
0.03 upwards
F (n)
0.02 downwards
2
0.01 n=400
0
0 100 200 300 400 500 600
n
Fig. 5.13 Plots for INCP: a Plots of log(F2(n)) versus log(n) and b plots of F2(n) versus n
0.2
0.15
0.1
0.05
0
Df
-0.05
-0.1
-0.15
-0.2
-0.25
0 100 200 300 400 500 600
n
Fig. 5.14 Plots of Df for SHCP
trend of gold price. This feature was also observed in the lower figures in Figs. 5.12
and 5.13. At approximately 400 days, regime switching of the asymmetric fluctu-
ation function occurred.
If the fluctuation function asymmetric indicator is defined as
Df ¼ log F2þ ðnÞ log F2 ðnÞ, then Df can precisely measure the fluctuation
function asymmetry obtained using the A-MFDFA method. The Df variation trends
of the Shanghai gold market (SHCP) and the international gold market (INCP) at
different timescales are shown in Figs. 5.14 and 5.15, respectively.
Based on Figs. 5.14 and 5.15, the fluctuation trends of the fluctuation function
asymmetric indicator are similar between the Chinese and international gold return
rates. In particular, the Df values fluctuated between 0 and 400 days. During this
period, the average value was negative. However, after 400 days, the level
0.3
0.2
0.1
0
Df
-0.1
-0.2
-0.3
-0.4
0 100 200 300 400 500 600
n
Fig. 5.15 Plots of Df for INCP
increased significantly and basically fluctuated above the straight line (Df ¼ 0).
These findings verified the analysis results shown in Figs. 5.12 and 5.13.
5.3.2.2 Estimating the Generalized Hurst Exponent H(q)
The estimated Hurst exponents HðqÞ of the Shanghai and international gold return
rates with q varying between −10 and 10 (the step size is 1) are presented in
Figs. 5.16 and 5.17.
Based on Figs. 5.16 and 5.17, the estimated HðqÞ, H þ ðqÞ, and H ðqÞ in the
overall, increasing, and decreasing trends exhibited a decreasing trend with the
increase in q whether in the Shanghai or international gold market. This finding
indicates that regardless of the trend, the Chinese and international gold markets
exhibited multifractal features. Moreover, large fluctuations persisted shorter than
small fluctuations because small fluctuations were more frequent and more affected
by political and economic fundamentals in the long term. Large fluctuations fre-
quently occur suddenly and have a significant influence, which can be detected
easily, such that people can handle it immediately and will not make it last long. In
Fig. 5.16 Plots of HðqÞ 0.8

versus q for SHCP overall
0.7 upwards
downwards
0.6
H (q)
0.5
0.4
0.3
0.2
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 5.17 Plots of HðqÞ 0.8

versus q for INCP overall
0.7 upwards
downwards
0.6
H (q)
0.5
0.4
0.3
0.2
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
addition, Figs. 5.16 and 5.17 show that, with a certain q value, H þ ðqÞ was greater
than H ðqÞ, which indicates that multifractal features were asymmetric between the
Chinese and international gold return series. Moreover, the persistence of the gold
return series was longer than that of gold return rate in an increasing trend. Also, as
q increased, the bandwidth (gap) between H þ ðqÞ and H ðqÞ gradually narrowed
(shrank), which indicates that the asymmetric degree of the multifractal charac-
teristics between the Chinese and international gold markets decreased with the
increase in fluctuation range. We assumed that this phenomenon may have resulted
from chasing increasing stocks and group psychology when the market is in an
increasing trend as well as the risk-averse mentality of investors in a decreasing
trend; hence, persistence is longer in an increasing trend. However, when large
fluctuations occur, regulars will take measures in time, and thus, the asymmetric
degree will decrease.
5.3.2.3 Analyzing the Multifractal Singularity Spectrum
The multifractal singularity spectra of the Shanghai and international gold return
rates, are plotted in Figs. 5.18 and 5.19, respectively, i.e., the relationship diagrams
of f ðaÞ and a.
Fig. 5.18 Multifractal 1.5

singularity spectrum for
SHCP 1
0.5
f(α )
-0.5 overall
upwards
-1 downwards
-1.5
0.2 0.3 0.4 0.5 0.6 0.7 0.8
α
Fig. 5.19 Multifractal 1.5

singularity spectrum for INCP
1
0.5
f(α )
0
-0.5
overall
upwards
-1
downwards
-1.5
0.2 0.3 0.4 0.5 0.6 0.7 0.8
α
Based on Figs. 5.18 and 5.19, the multifractal singularity spectra of the Shanghai
and international gold return rates presented an anti-parabolic fluctuation form
rather than a definite value. This finding further verified the multifractal features of
the Chinese and international gold markets. In addition, in the singularity spectra of
the Shanghai and international gold markets, the negative generalized fractal
dimension (f ðaÞ\0) fell on the left side of the singularity spectrum. This finding
indicates that the substantial fluctuations of the Chinese and international gold
markets were random (Mandelbrot 1990, 1991). Moreover, whether in the Shanghai
or international gold market, the singularity spectrum width in the increasing trend
was larger than that in the decreasing trend. This finding indicates that the
decreasing trend of the multifractal features of the Chinese and international gold
return rates was more intensive than the increasing trend. Moreover, based on the
width of the singularity spectrum, the return rate asymmetry of the Shanghai gold
market was slightly stronger than that of the international gold market, which,
however, was insignificant. We assumed that this phenomenon may be attributed to
the fact that the Chinese financial market still has to be perfected and has more
speculative arbitrage opportunities than the international market, which may result
in a strong asymmetry degree.
5.3.2.4 Time-Varying Analysis of Multifractal Asymmetry
The sliding window technology is frequently used in examining the time-varying

feature of time series. The research of Cristescu et al. (2012) showed that a time
window length that is slightly longer than 800 is appropriate. However, for a small
sample size, a time window length of less than 700 is frequently applied, which
leads to unreliable findings. This assumption was used in related studies on time
series in the financial field (Matteo 2007; Schmitt et al. 2000; Stanley and Plerou
2001; Grech and Mazur 2004). However, detecting multifractal features may result
in misleading conclusions when a small sample size is used. Thus, the sliding
window length used in our research was 1000 (close to 4 years), which was
0.6
2010 SHCP
2005 2007-2008 European
0.55 Share Subprime
Debt
INCP
Segregation Mortgage Crisis
2003 Reform Crisis
0.5 QDII
ΔΗ+-(2)
0.45
0.4
0.35
0.3
0.25
0 20 40 60 80 100 120 140
sample points
Fig. 5.20 Time evolution of DH ð2Þ for Shanghai gold return and international gold return
expected to depict the asymmetric time-varying trend of multifractal features

accurately. The slide step was 10 days. The values of the estimated DH ð2Þ of the
Shanghai and international gold returns are plotted in Fig. 5.20.
As shown in Fig. 5.20, the estimated DH ð2Þ of the Shanghai and international
gold markets changed with time (sample points), which indicates that the multi-
fractal asymmetry of the Chinese and international gold markets is characterized by
its time-varying feature. Further observations showed that the evolutionary patterns
of the asymmetric degree were basically identical in the Shanghai and international
gold markets. Moreover, considering the time points of some remarkably classic
events in the Chinese and international finance markets, the global financial crisis
(2007–2008 American Subprime Mortgage Crisis and 2010 European Debt Crisis)
seemingly increased the asymmetric degree of the multifractals of the Shanghai and
international gold markets to a significant extent. Thus, our research shows that
sudden changes in the external economic situation and financial environment may
significantly enhance the asymmetry of the multifractal features of the Chinese and
international gold markets.
5.3.3 Discussion
5.3.3.1 Statistical Tests
A statistical test should be made to show the results above are not spuriously reported.
Fourier phase randomization is widely used, but the long range power-law correlations
in volatilities may completely disappear after phase randomization according to the
simulation results (Podobnik et al. 2007). So we generate the surrogate data with
Gaussian distribution by applying a simple method in Zhou (2012), but maintain the
linear correlation of the original data (Lim et al. 2007; Norouzzadeh and Rahmani
2006). The algorithm for this method is described as follows.
We generate a sequence of random numbers fm0 ðtÞ : t ¼ 1; 2; . . .; N g with the
Gaussian distribution and rearrange it, then can get the rearranged series
fmðtÞ : t ¼ 1; 2; . . .; N g, which has the same rank ordering as the original series
fr ðtÞ : t ¼ 1; 2; . . .; N g. Thus, mðtÞ should have a rank n in the series
fmðtÞ : t ¼ 1; 2; . . .; N g if and only if rðtÞ has a rank n in the original series
frðtÞ : t ¼ 1; 2; . . .; N g.
Then we take the surrogated data to the MF-DFA model, and get corresponding
Hurst exponents, then repeat 100 times, so can get 100 pairs data. Each setting can
be analyzed with respect to the 2.5th and the 97.5th quartiles, that is, the 95%
confidence interval, which it the standard deviation of the estimate HðqÞ, H þ ðqÞ,
and H ðqÞ. We define H0:025 ðqÞ, H0:975 ðqÞ, DH0:025

ðqÞ, DH0:975 ðqÞ as the critical

values of the 95% confidence interval for HðqÞ and DH ðqÞ, respectively, where
DH ðqÞ ¼ jH þ ðqÞ H ðqÞj: The results are as follows.
Figures 5.21 and 5.22 show the results of statistical tests of HðqÞ for SHCP and
INCP. HðqÞ in most situations are not significant, but appears when q is greater
(like q = 6, 7, 8, 9, 10 in Fig. 5.21, and q = 6, 7, 9 in Fig. 5.22). And Figs. 5.23
Fig. 5.21 Statistical tests of 0.7

H(q)
HðqÞ for SHCP
H0.025(q)
0.6
H0.975(q)
0.5
H(q)
0.4
0.3
0.2
-10 -8 -6 -4 -2 0 2 4 6 8 10
q

HðqÞ for INCP H(q)
H0.025(q)
0.6
H0.975(q)
0.5
H(q)
0.4
0.3
0.2
-10 -8 -6 -4 -2 0 2 4 6 8 10
q

DH ðqÞ for SHCP ΔΗ±(q)
0.1 ΔΗ±0.025(q)
0.08 ΔΗ±0.975(q)
ΔΗ±(q)
0.06
0.04
0.02
0
-10 -8 -6 -4 -2 0 2 4 6 8 10
q

DH ðqÞ for INCP ΔΗ±(q)
0.1 ΔΗ±0.025(q)
0.08 ΔΗ± (q)
ΔΗ±(q)
0.975
0.06
0.04
0.02
0
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
and 5.24 show the results of statistical tests of DH ðqÞ for SHCP and INCP. We
find the values of DH ðqÞ are larger than the critical values except when q = 9 in
Fig. 5.24. So most of DH ðqÞ are significant, that is the asymmetry exist when the
market is in different trends.
5.3.3.2 Origin of Multifractality with Different Trends
According to Cao et al. (2013) discussion of origin of multifractality, there are two
major sources of multifractality: (a) different long-range correlations for small and
large fluctuations, and (b) fat-tailed probability distributions of variations. So we
should understand the effect of long range correlations by comparing the multi-
fractality between original series and randomly shuffled series or the surrogated
series.
The randomly shuffled series are constructed as follows: First, generate pairs
(p, q) of random integer numbers with p, q N, where N is the length of the
shuffled time series. Then swap entries p and q. At last, repeat the first and the
second step for 20 N times.
Table 5.4 DH of the original, shuffled, and surrogated series

SHCP INCP
series series series series series series
Overall 0.3481 0.2415 0.2294 0.3054 0.0899 0.3089
(30.62%) (34.10%) (70.56%) (−0.35%)
Upward 0.3892 0.2447 0.2646 0.3604 0.1518 0.3578
(37.13%) (32.01%) (57.88%) (0.72%)
Downward 0.3357 0.2723 0.2320 0.289 0.0902 0.2644
(18.89%) (30.89%) (68.79%) (8.51%)
Note The value in parentheses is the change (in percentage) in the DH value for the shuffled
(resp. surrogated) data to that of the original data, namely, ðDHorig DHshuf Þ=DHorig
(resp. ðDHorig DHsurr Þ=DHorig ). q = −10, −9, −8, …, 8, 9, 10
The surrogated series are constructed as the series in the statistical test.
According to Eq. (5.17), the degree of multifractality becomes stronger with the
greater DH, and DHorig , DHshuf and DHsurr represent DH for the ordinal series,
shuffled series and surrogate series respectively.
Table 5.4 shows the multifractality degree of the original, shuffled and surro-
gated series for SHCP and INCP. Overall, the values of DHshuf and DHsurr for
SHCP, and DHshuf for INCP are smaller than DHorig , which indicates that the
multifractal scaling behavior in the Chinese gold price return series is caused not
only by long-range temporal correlation but also by fat-tailed distribution, but
long-range temporal correlation plays a main effect on the international gold
market. And in different trends, DHshuf and DHsurr are smaller than DHorig in two
markets, indicating that both long-range temporal correlation and fat-tailed distri-
bution have an influence on the multifractal scaling behavior, and also the two
markets are more attributed to long-range correlation when they’re in an uptrend,
because DHshuf is smaller than DHsurr .
5.3.3.3 Source of the Asymmetry
According to Alvarez-Ramirez et al. study (2009), either intrinsic correlation or

fat-tailed distribution can induce the asymmetric scaling behavior. So we also use
the same method applied in Sect. 5.3.3.1 to study the source of asymmetry scaling
behavior. And the asymmetric degree of correlation is also quantified as
DH ðqÞ ¼ jH þ ðqÞ H ðqÞj.
Figures 5.25 and 5.26 show the values of DH±(q) for the original, shuffled and
surrogated data for SHCP and INCP respectively. In Figs. 5.21 and 5.22, the values
of DH±(q) for the shuffled and surrogated data are smaller than those of the original
data, except for the shuffled and surrogated data for SNCP when q 5, and the
surrogated data for INCP when q 3. So when q<5, both the long-range corre-
lation and fat-tailed distribution are the sources of the asymmetric scaling in
Fig. 5.25 Plots of DH±(q) 0.12

for the original, shuffled, and original data
surrogated data for SHCP 0.1 shuffled data
(q = −10, −9, −8,…8, 9, 10) surrogated data
0.08
ΔΗ+-(q)
0.06
0.04
0.02
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 5.26 Plots of DH±(q) 0.14
for the original, shuffled, and 0.12

surrogated data for INCP
(q = −10, −9, −8,…8, 9, 10) 0.1
ΔΗ+-(q)
0.08
0.06
0.04
original data
0.02 shuffled data
surrogated data
0
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Shanghai gold market. And the long-range correlation and fat-tailed distribution
(when q<3) are the sources of the asymmetric scaling in international gold market,
meanwhile, is more related to long-range correlation, because the values of DH±(q)
for the shuffled INCP data are smaller than those for the surrogated INCP data.
Above all, the asymmetric scaling behavior may be induced by different factors,
like long-range correlation and fat-tailed distribution.
5.3.4 Asymmetric Influences of Good and Bad News

on Gold Price Fluctuation
Based on the previous analysis of the Hurst exponent, we observed that the Hurst
exponent of the Chinese and international gold price return series, i.e., H(2), did not
significantly deviate from 0.5, which indicates that its long memory feature was
insignificant. Thus, the long memory parameter was not considered in establishing
the asymmetric GARCH model, namely, the fractional difference parameter (the
fractional difference parameter of the FIEGARCH model in sequence simulation
estimation was insignificant at the significance level of 10%). We only used the
Table 5.5 Estimated results of TARCH(1, 1)

Coefficient Estimated values Stv.Dev Z statistics P values
SHCP a0 4.18E-07 1.63E-07 2.555688 0.0106
a1 0.085270 0.007927 10.75749 0.0000
c −0.056327 0.008410 −6.697367 0.0000
k1 0.941984 0.004880 193.0209 0.0000
INCP a0 9.00E-07 2.58E-07 3.483747 0.0005
a1 0.055667 0.006075 9.163246 0.0000
c −0.038138 0.007811 −4.882536 0.0008
k1 0.958064 0.003893 246.1282 0.0000
Table 5.6 Estimated results of EGARCH (1, 1)

Coefficient Estimated values Stv.Dev Z statistics P values
SHCP a1 −0.192782 0.025269 −7.629119 0.0000
c 0.153823 0.010869 14.15192 0.0000
k1 0.991493 0.002333 424.9859 0.0000
INCP a1 −0.104761 0.016985 −6.167995 0.0000
c 0.083131 0.006652 12.49725 0.0000
k1 0.028177 0.005941 4.742557 0.0000
TARCH and EGARCH models to investigate the asymmetry of the return rate
fluctuations between the Chinese and international gold markets. Existing literature
showed that GARCH (1, 1) can be effectively used to evaluate conditional variance
for the fitting of the conditional variance equation. Thus, our research abided by this
rule of experience. For p and q in the TARCH and EGARCH models, p = 1 and
q = 1 were selected. The estimation results of the TARCH (1, 1) and EGARCH (1,
1) models of the Chinese and international gold price return series are shown in
Tables 5.5 and 5.6, respectively.
Based on Table 5.5, coefficient c of the asymmetric term of the SHCP and INCP
sequences was significantly not 0 at the significance level of 5%, which indicates
the asymmetry of gold price fluctuation between the Chinese and international gold
markets. The estimated asymmetric coefficients c of the Chinese and international
gold markets were −0.056327 and −0.038138, respectively. These values are
smaller than 0, which indicates that good news had a more significant effect on gold
price in the Chinese and international gold markets than bad news. Moreover, the
value of the estimated c also showed that good news had a more significant effect
on the Chinese gold market than on the international gold market.
Based on Table 5.6, coefficient c of the asymmetric term of the SHCP and INCP
sequences was significantly not 0 at the significance level of 5%, which indicates
the asymmetry of gold price fluctuation between the Chinese and international gold
markets. Using the EGARCH (1, 1) model, the estimated asymmetric coefficients c
of the Chinese and international gold markets were 0.153823 and 0.083131,
respectively. These values are larger than 0, which indicates that good news had a
more significant effect on gold market cyclical fluctuation in the Chinese and
international gold markets than bad news. Moreover, the value of the estimated c
also showed that good news had a more significant effect on the Chinese gold
market than on the international gold market. These results further verified the
empirical findings of the TARCH model.
5.3.5 Conclusions
Based on the daily price in the Shanghai and New York gold markets from January
2, 2003 to April 27, 2012, we conducted an empirical analysis of the asymmetric
multifractal feature of price fluctuations in the Shanghai and international gold
markets using the A-MFDFA method. The following conclusions were obtained.
(1) The Shanghai and international gold markets demonstrated multifractal scaling
behaviors. Large fluctuations persisted longer than small fluctuations.
(2) The multifractal features of the Shanghai and international gold markets in
different trends exhibited asymmetry. The persistence of the gold return series
in the increasing trend was longer than that in the decreasing trend. The
asymmetric degree of multifractals in the Chinese and international gold mar-
kets decreased with the increase in fluctuation range.
(3) Multifractal asymmetry in the Chinese and international gold markets is
characterized by its time-varying feature. The Shanghai and international gold
markets basically shared a similar asymmetric degree evolution pattern. The
2008 American Subprime Mortgage Crisis and 2010 European Debt Crisis
enhanced the asymmetric degree of the multifractal features of the Chinese and
international gold markets.
(4) Either long-range correlation or the fat-tailed distribution induces scaling
multifractality in the Shanghai gold markets, and only long-range correlation
induces scaling multifractality in the international gold markets. Meanwhile, the
multifractal scaling behavior is more related to long-range correlation when the
two markets are in an uptrend. Also, the two factors also induce the scaling
asymmetries, and the international gold market is more related to long-range
correlation.
(5) Good news had a more significant effect on gold market fluctuation in the
Chinese and international gold markets than bad news. Moreover, good news
had a more significant effect on the Chinese gold market than on the interna-
tional gold market.
The A-MFDFA method prevented the occurrence of multifractal features in a
partial trend of the time series and determined the asymmetry of multifractal fea-
tures in increasing and decreasing trends. Meanwhile, the asymmetric behaviors of
multifractal features in different trends are conducive to identifying financial risks
better, providing theoretical and empirical support for generating a portfolio, and
assisting in the macro-control of monetary markets with policy support.
Based on the study of Schmittbuhl et al. (1995) on measurement errors in the
Hurst exponent using different methods, we determined that errors may also exist in
the Hurst exponent used in this study. However, we only wished to formulate some
inherent laws for the gold markets using the Hurst exponent. Thereby, we did not
discuss the measurement errors of the Hurst exponent, similar to many other recent
studies that also used the DFA method to calculate the Hurst exponent (Abosedra
and Baghestani 2004; Du and Ning 2008; Wang et al. 2009). And for the results of
statistical test of the Hurst exponents, parts of the results are not significant, which
may be caused by the sample length or other factors. Therefore, they’re worth being
discussed in a future study, which may make our study extensive and exhaustive.
References
S. Abosedra, H. Baghestani. On the predictive accuracy of crude oil futures prices. Energ. Policy.
32(12), 1389–1393 (2004)
J. Alvarez-Ramirez, E. Rodriguez, J.C. Echeverria, A DFA approach for assessing asymmetric
A. Ang, G. Bekaert, International asset allocation with regime shifts. Rev. Financial Stud. 15(4),
1137–1187 (2002)
A. Ang, J. Chen, Asymmetric correlations of equity portfolios. J. Financ. Econ. 63, 443–494
(2002)
K.H. Bae, G.A. Karolyi, R.M. Stulz, A new approach to measuring financial market contagion.
Rev. Financial Stud. 16, 717–764 (2003)
M.Y. Bai, H.B. Zhu, Power law and multiscaling properties of the Chinese stock market. Phys.
A 389, 1883–1890 (2010)
J. Barunik, T. Aste, T.D. Matteo, R.-P. Liu, Understanding the source of multifractality in financial
markets. Phys. A 391, 4234–4251 (2012)
A. Bera, C. Jarque, Efficient tests for normality, heteroskedasticity and serial independence of
regression residuals: Monte Carlo evidence. Econ. Lett. 7, 313–318 (1981)
M.I. Bogachev, J.F. Eichner, A. Bunde, Effect of nonlinear correlations on the statistics of return
intervals in multifractal data sets. Phys. Rev. Lett. 99, 240601 (2007)
D.O. Cajueiro, B.M. Tabak, Ranking efficiency for emerging markets. Chaos Solitons Fractals 22,
349–352 (2004a)
D.O. Cajueiro, B.M. Tabak, The Hurst exponent over time: testing the assertion that emerging
markets are becoming more efficient. Phys. A 336, 521–537 (2004b)
D.O. Cajueiro, B.M. Tabak, Testing for time-varying long-range dependence in real state equity
returns. Chaos Solitons Fractals 38, 293–307 (2008)
D.O. Cajueiro, P. Gogas, B.M. Tabak, Does financial market liberalization increase the degree of
market efficiency? The case of the Athens stock exchange. Int. Rev. Financial Anal. 18, 50–57
(2009)
G.X. Cao, L.B. Xu, J.Cao, Multifractal detrended cross-correlations between the Chinese
exchange market and stock market. Phys. A, 4855–4866 (2012)
G. Cao, J. Cao, L. Xu, Asymmetric multifractal scaling behavior in the Chinese stock market:
Based on asymmetric MF-DFA. Phys. A: Statistical Mechanics and its Applications. 392(4),
797–807 (2013)
C.P. Cristescu, C. Stan, E.I. Scarlat, T. Minea, C.M. Cristescu, Parameter motivated mutual
correlation analysis: application to the study of currency exchange rates based on intermittency
parameter and Hurst exponent. Phys. A 391, 2623–2635 (2012)
L. Czarnecki, D. Grech, Multifractal dynamics of stock markets. Acta Phys. Pol. A 117, 623–629
(2010)
R. Demirer, in Asymmetric Correlation of Futures Markets and Optimal Hedging, (2003). http://
webradio.siue.edu/business/econfin/pdf/demirer-charnes.pdf
G.X. Du, X.X. Ning, Multifractal properties of Chinese stock market in Shanghai. Phys. A 387,
261–269 (2008)
M.R. Eldridge, C. Bernbarde, I. Mulvey, Evidence of Chaos in the S&P 500 cash index. Adv.
Futures Options Res. 6, 179–192 (1993)
D. Grech, Z. Mazur, Can one make any crash prediction in finance using the local hurst exponent
idea? Phys. A 336, 133–145 (2004)
D. Grech, G. Pamula, The local hurst exponent of the financial time series in the vicinity of crashes
on the polish stock exchange market. Phys. A 387, 4299–4308 (2008)
M.T. Greene, B.D. Fieltz, Long term dependence in common stock returns. J. Financ. Econ. 4,
249–339 (1997)
Z.Q. Jiang, W.X. Zhou, Multifractal analysis of Chinese stock volatilities based on the partition
function approach. Phys. A 387, 4881–4888 (2008)
Multifractal detrended fluctuation analysis of nonstationary time series. Phys. A 316, 87–114
(2002)
S. Kumar, N. Deo, Multifractal properties of the Indian financial market. Phys. A 388, 1593–1602
(2009)
G. Lim, S. Kim, H. Lee, K. Kim, D.-I. Lee, Multifractal detrended fluctuation analysis of
derivative and spot markets. Phys. A 386, 259–266 (2007)
F. Longin, B. Solnik, Extreme correlation of international equity markets. J. Finance 56, 649–676
(2001)
B.B. Mandelbrot, Negative fractal dimensions and multifractals. Phys. A 163, 306–315 (1990)
B.B. Mandelbrot, Random multifractals: negative dimensions and the resulting limitations of the
thermodynamic formalism. Proc. R. Soc. Lond. Ser. A 434, 79–88 (1991)
K. Matia, Y. Ashkenazy, H.E. Stanley, Multifractal properties of price fluctuations of stock and
T.D. Matteo, Multi-scaling in finance. Quant. Finance 7, 21–36 (2007)
P. Norouzzadeh, G.R. Jafari, Application of multifractal measures to Tehran price index. Phys.
A 356, 609–627 (2005)
P. Norouzzadeh, B. Rahmani, A multifractal detrended fluctuation description of Iranian rial-US
dollar exchange rate. Phys. A 367, 328–336 (2006)
C.K. Peng, S.V. Buldyrev, S. Havlin et al., Mosaic organization of DNA nucleotides. Phys. Rev.
E 49, 1685–1689 (1994)
nonstationary time series. Phys. Rev. Lett. 100, 084102 (2008)
B. Podobnik, D.F. Fu, H.E. Stanley, PCh. Ivanov, Power-law autocorrelated stochastic processes
with long-range cross-correlations. Eur. Phys. J. B 56, 47–52 (2007)
F. Schmitt, D. Schertzer, S. Lovejoy, Multifractal fluctuations in finance. Int. J. Theor. Appl.
Finance 3, 361–364 (2000)
J. Schmittbuhl, J.-P. Vilotte, S. Roux. Reliability of self-affine measurements. Phys. Rev. E. 51(1),
131 (1995)
H.E. Stanley, V. Plerou, Scaling and universality in economics: empirical results and theoretical
interpretation. Quant. Finance 1, 563–567 (2001)
B.M. Tabak, D.O. Cajueiro, Are the crude oil markets becoming weakly efficient over time? A test
for time-varying long-range dependence in prices and volatility. Energy Econ. 29, 28–36
(2007)
References 111
Y.D. Wang, L. Liu, R.B. Gu, Analysis of efficiency for Shenzhen stock market based on
multifractal detrended fluctuation analysis. Int. Rev. Financial Anal. 18, 271–276 (2009)
Y.D. Wang, L. Liu, R.B. Gu, J.J. Cao, H.Y. Wang, Analysis of market efficiency for the Shanghai
stock market over time. Phys. A 389, 1635–1642 (2010)
Y.D. Wang, C.F. Wu, Z.Y. Pan, Multifractal detrending moving average analysis on the US Dollar
exchange rates. Phys. A 390, 3512–3523 (2011a)
Y.D. Wang, Y. Wei, C.F. Wu, Detrended fluctuation analysis on spot and futures markets of West
Texas Intermediate crude oil. Phys. A 390, 864–875 (2011b)
Y. Wei, D.S. Huang, Multifractal analysis of SSEC in Chinese stock market: a different empirical
results from Heng Seng index. Phys. A 355, 497–508 (2005)
Y. Wei, P. Wang, Forecasting volatility of SSEC in Chinese stock market using multifractal
analysis. Phys. A 387, 1585–1592 (2008)
Y. Yuan, X.T. Zhuang, Measuring multifractality of stock price fluctuation using multifractal
detrended fluctuation analysis. Phys. A 388, 2189–2197 (2009)
Rev. E 77, 066211 (2008)
W.X. Zhou, The components of empirical multifractality in financial returns. Europhys. Lett. 88,
28004 (2009)
W.X. Zhou, Finite-size effect and the components of multifractality in financial volatility. Chaos
W.X. Zhou, Z.H. Yu, Multifractality of drop breakup in the air-blast nozzle atomization process.
Phys. Rev. E 63, 016302 (2001)
W.C. Zhou, H.C. Xu, Z.Y. Cai, J.R. Wei, X.Y. Zhu, W. Wang, L. Zhao, J.-P. Huang, Peculiar
statistical properties of Chinese stock indices in bull and bear market phases. Phys. A 388,
891–899 (2009)
Chapter 6
Asymmetric Multifractal Detrended
Cross-Correlation Analysis (MF-ADCCA)
Hedging crucially relies on the correlations between the assets hedged and the
financial instruments used. The presence of asymmetric correlations can potentially
cause problems to hedging effectiveness. Furthermore, standard mean-variance
investment theory advises portfolio diversification, but the value of this advice may
be questioned if all stocks tend to fall as the market falls (Hong et al. 2007).
Recently, several studies focus on asymmetric properties of financial markets or
asset returns (Ang and Bekaert 2002; Taamouti and Tsafack 2009; Ding et al.
2011). Login and Solnik (2001) find that international markets have greater cor-
relations with the US market when the latter is going down than when going
up. Ang and Chen (2002) detect strong asymmetric correlation between stock
portfolios and the US market.
Ang and Bekaert (2002) use the two-regime-switching model to determine the
connection between low returns and high correlation. Longin and Solnik (2001),
defining a new concept termed exceedance correlation, report a large correlation
between large negative returns, and zero correlation between large positive returns.
Ang and Chen (2002) use the exceedance correlation test to show that asymmetric
correlation exists in different types of domestic portfolios. Hong et al. (2007)
propose a model-free method and confirm such asymmetry in the US stock market.
Although these methods are able to detect the presence of asymmetric correlations,
they also depend on assumptions, such as the use of a model or the selection of the
threshold value. Therefore, the method of assessing asymmetric correlations is
worth studying further. Although A-MFDFA only detects asymmetries of the time
series itself, this method provides a new idea in measuring asymmetries of the
cross-correlation which can not be distinguished by MF-DCCA method proposed
by Podobnik and Stanley (2008) and Zhou (2008).
Therefore, the first contribution of this chapter is to propose a straightforward
modification of MF-DCCA to detect the asymmetric cross-correlation between two
non-stationary series. We combine MF-DCCA and A-MFDFA which are intro-
duced in Chaps. 4 and 5 respectively, and then propose the multifractal asymmetric
detrended cross-correlation analysis method (MF-ADCCA). MF-ADCCA has three
114 6 Asymmetric Multifractal Detrended Cross-Correlation …
appealing features. First, it is model free. Unlike the test of Ang and Chen (2002),
ours is computed without having to specify a statistical model for the data. Second,
MF-ADCCA is easy to implement. The MF-ADCCA scaling exponents are directly
computed and several properties are given. Third, MF-ADCCA can also measure
the multifractal characteristic of the different cross-correlations. The proposed
method can be directly applied to a variety of fields to provide insights in assessing
whether the asymmetric cross-correlation exists or not.
The second contribution of this chapter is to investigate the asymmetric
cross-correlation between Chinese stock returns with the exchange rate of the
Chinese Yuan (RMB) to different main foreign currencies and US stock returns. We
assess the asymmetric cross-correlations between Chinese stock returns with the
other financial returns in China when the Chinese market is rising and falling, and
when the other financial markets are going up and down. Furthermore, the multi-
fractal features of various asymmetric cross-correlation are also discussed. We find
that the asymmetries exist in the cross-correlation of the Chinese stock market and
the RMB exchange market, and the asymmetric cross-relations are multifractal.
Moreover, the cross-correlations between the Chinese stock market and the RMB/
USD exchange market are more persistent when any one of the markets is falling.
On the contrary, the cross-correlations between the Chinese stock market and the
RMB/EU, RMB/GBP, RMB/JPY exchange markets, and the US stock market are
more persistent when any one of the markets is rising. In addition, the asymmetries
of the cross-correlations between the Chinese stock market and the different RMB
exchange markets present different persistence for large and small price
fluctuations.
6.1 Methodology
We combine MF-DCCA and A-MFDFA, and propose the MF-ADCCA method,

which is described as follows. Assume that two time-series fxð1Þ ðtÞg and fxð2Þ ðtÞg
exist, t ¼ 1; 2; . . .; N, where N is the length of the series. The MF-ADCCA method
can then be summarized as follows.
Step 1: We construct the profile
X
j
yðiÞ ðjÞ ¼ ðxðiÞ ðtÞ xðiÞ Þ; j ¼ 1; 2; . . .; N; i ¼ 1; 2 ð6:1Þ
t¼1
P
N
where xðiÞ ¼ N1 xðiÞ ðtÞ.
t¼1
Step 2: The time series fxðiÞ ðtÞg and its profile fyðiÞ ðtÞg ði ¼ 1; 2Þ are divided
into Nn ¼ intðN=nÞ non-overlapping subtime series (i.e., boxes) of equal length n,
respectively. The record length N does not need to be a multiple of the considered
time scale n, and thus, a short part of the profile will remain in most cases. The same
6.1 Methodology 115
procedure is repeated starting from the other end of the record to account for the
ðiÞ
aforementioned part. Thus, 2Nn segments are obtained altogether. Let Sj ¼
ðiÞ ðiÞ ðiÞ
fsj;k ; k ¼ 1. . .; ng denote the jth subtime series of length n and Yj ¼ fyj;k ; k ¼
1; . . .; ng denote the according integrated time-series (i.e., profile) in the jth time
interval, j ¼ 1; 2; . . .; 2Nn . In the jth segment, k ¼ 1; 2; . . .; n, we have
ðiÞ ðiÞ
sj;k ¼ xðiÞ ððj 1Þn þ kÞ, yj;k ¼ yðiÞ ððj 1Þn þ kÞ, for j ¼ 1; 2; . . .; Nn , and
ðiÞ ðiÞ
sj;k ¼ xðiÞ ðN ðj Ns Þn þ kÞ; yj;k ¼ yðiÞ ðN ðj Ns Þn þ kÞ ð6:2Þ
for j ¼ Ns þ 1; . . .; 2Nn . Based on the recommendations of Peng et al. (1994),

5 n N=4 is traditionally selected.
ðiÞ ðiÞ
Step 3: For each subtime-series Sj ¼ fsj;k ; k ¼ 1. . .; ng and its profile time series,
ðiÞ ðiÞ
Yj ¼ fyj;k ; k ¼ 1; . . .; ng, we compute the corresponding local least-squares line fits
ðiÞ ðiÞ ðiÞ ðiÞ
LSðiÞ ðkÞ ¼ aSj þ bSj k and LY ðiÞ ðkÞ ¼ aYj þ bYj k, where k represents the horizontal
j j
coordinate and i ¼ 1; 2. The fits LSðiÞ ðkÞ and LY ðiÞ ðkÞ represent the linear trends for the
j j
ðiÞ ðiÞ
jth subtime series Sj and its integrated time-series Yj , respectively ði ¼ 1; 2Þ. The
linear fit LSðiÞ ðkÞ is used only to discriminate, via the slope bSðiÞ , whether the trend of the
j j
ðiÞ
subtime series Sj is positive or negative. The linear fit LY ðiÞ ðkÞ is used to detrend the
j
ðiÞ
integrated time series Yj . We then determine the fluctuation functions
n
1X
yð1Þ L ð1Þ ðkÞ yð2Þ L ð2Þ ðkÞ
Fj ðnÞ ¼ j;k Y j;k Y ð6:3Þ
n k¼1 j j
for each segment, j ¼ 1; 2; . . .; 2Nn .

Step 4: The average fluctuation functions for several cases when the time series
xðiÞ has piecewise positive and negative linear trends are considered to assess the
asymmetric cross-correlation scaling properties. This trend discrimination is made
by using the sign of the slope bSðiÞ , that is, bSðiÞ [ 0 (resp. bSðiÞ \0) indicates that the
j j j
ðiÞ ðiÞ
time series x has a positive (resp. negative) trend in the subtime series Sj .
Therefore, if we consider the asymmetric cross-correlation between two markets
due to different trends of one time-series (for example fxð1Þ ðtÞg), then the direc-
tional q-order average fluctuation functions are computed by
!1=q
2Nn signðb ð1Þ Þ þ 1
1 X Sj
Fqþ ðnÞ ¼ þ
½Fj ðnÞq=2 ð6:4Þ
M j¼1 2
!1=q
2Nn ½signðb ð1Þ Þ 1
1 X Sj
Fq ðnÞ ¼
½Fj ðnÞq=2 ð6:5Þ
M j¼1 2
Pn signðbSð1Þ Þ þ 1
2N Pn ½signðbSð1Þ Þ1
2N
where M þ ¼ j
2 and M ¼ 2
j
are the numbers of subtime
j¼1 j¼1
series with positive and negative trends, respectively. Assume that bSð1Þ 6¼ 0 for all
j
j ¼ 1; 2; . . .; 2Nn , such that M þ þ M ¼ 2Nn .

Similarly, the traditional MF-DCCA proposed by Zhou (2008) is implemented
by computing the average fluctuation function
!1=q
1 X2Nn
Fq ðnÞ ¼ ½Fj ðnÞq=2 ð6:6Þ
2Nn j¼1
Thereby, if power-law cross-correlations exist, then the scaling or power-law

relationship should satisfy
þ
Fq ðnÞ nH12 ðqÞ ; Fqþ ðnÞ nH12 ðqÞ ; Fq ðnÞ nH12 ðqÞ ð6:7Þ
þ
where H12 ðqÞ, H12 ðqÞ, and H12 ðqÞ, denote overall, upward (of time series
fx ðtÞg), and downward (of time series fxð1Þ ðtÞg) scaling exponents, respectively.
ð1Þ
These scaling behaviors of the fluctuations in Eq. (6.7) are determined by

analyzing the log-log plots of Fq ðnÞ, Fqþ ðnÞ, and Fq ðnÞ n for each value of q. Thus,
Eq. (6.7) can be presented as follows:
log Fq ðnÞ ¼ H12 ðqÞ logðnÞ þ log A1 ð6:8Þ
log Fqþ ðnÞ ¼ H12

þ
ðqÞ logðnÞ þ log A2 ð6:9Þ
log Fq ðnÞ ¼ H12

ðqÞ logðnÞ þ log A3 ð6:10Þ
The exponent H12 ðqÞ is known as the generalized cross-correlation scaling

exponent representing the power-law relationship between two temporally corre-
lated time series. If H12 ð2Þ [ 0:5, then the cross-correlations between the two time
series are persistent, which implies an increase of one price is likely to be followed
by an increase of the other price. If H12 ð2Þ\0:5, then the cross-correlations
between the two time series are anti-persistent, which indicates an increase of one
price is likely to be followed by a decrease of the other price. For H12 ð2Þ ¼ 0:5,
only short-range cross-correlations (or no correlations at all) are considered between
the two time series, which indicates the change of one price only affects the
behavior of the other price in a significantly short term (or not affect at all).
6.1 Methodology 117
þ
If H12 ðqÞ ¼ H12 ðqÞ, then the cross-correlations between the two time series are
þ
symmetric. By contrast, if H12 ðqÞ 6¼ H12 ðqÞ, then the cross-correlations between
the two time series are asymmetric, which means that the cross-correlations are
different when the trending of the time series fxð1Þ ðtÞg is positive than when it is
negative. Moreover, similar to the proof of Alvarez-Ramirez (2009), we can obtain
þ
that if H12 ðqÞ 6¼ H12 ðqÞ, then necessarily, H12 ðqÞ 6¼ H12 ðqÞ, and vice versa.
þ
Moreover, if H12 ðqÞ 6¼ H12 ðqÞ, then H12 ðqÞ 6¼ H12 ðqÞ must hold, which means that
the positive and negative directions of the time series fxð1Þ ðtÞg have different
scaling exponents.
þ
For q [ 0, H12 ðqÞ, H12 ðqÞ, and H12 ðqÞ respectively describe the overall,
upward, and downward (of time series fxð1Þ ðtÞg) scaling behaviors of large fluc-
tuations, which are often characterized by a smaller scaling exponent H12 ðqÞ for
þ
multifractal time series. On the contrary, for q\0, H12 ðqÞ, H12 ðqÞ, and H12 ðqÞ
ð1Þ
describe the respective overall, upward, and downward (of time series fx ðtÞg)
scaling behaviors of small fluctuations. If the scaling exponent H12 ðqÞ is signifi-
cantly dependent on q, then the cross-correlations between the two time series is
multifractal. By contrast, if H12 ðqÞ is not significantly dependent on q, the
cross-correlations between the two time series is monofractal. Similarly, if
þ
the scaling exponent H12 ðqÞ (resp. H12 ðqÞ) is significantly dependent on q, then the
cross-correlations between the two time series when the time series fxð1Þ ðtÞg has a
þ
positive (resp. negative) trending are multifractal. By contrast, if H12 ðqÞ

(resp. H12 ðqÞ) is not significantly dependent on q, then the cross-correlations
between the two time series when the time series fxð1Þ ðtÞg has a positive
(resp. negative) trending are monofractal.
In addition, to measure the asymmetric degree of the cross-correlation, we define
þ
DH12 ðqÞ ¼ H12 ðqÞ H12 ðqÞ ð6:11Þ
For a fixed q, the absolute of the DH12 ðqÞ is larger and the asymmetric degree is
stronger. If DH12 ðqÞ [ 0, then the persistence of the cross-correlation is stronger
when the time series fxð1Þ ðtÞg has a positive trending than when it has a negative
one. On the contrary, if DH12 ðqÞ\0, then the persistence of the cross-correlation is
weaker when the time series fxð1Þ ðtÞg has a positive trending than when it has a
negative one. If DH12 ðqÞ is not a departure than 0, then the cross-correlations is
symmetric for the different trends of the time series fxð1Þ ðtÞg.
6.2 Data Description
On July 21, 2005, the People’s Bank of China (PBC) reformed the exchange rate
regime by moving into a managed floating exchange rate regime based on market
supply and demand with reference to a basket of currencies. Up to the end of 2011,
Table 6.1 Descriptive statistics of the seven return series

SSCI RMB/USD RMB/EU RMB/JPY RMB/HK RMB/GBP SP500
Mean 0.00050 −0.00017 −0.00013 0.00008 −0.00017 −0.00032 0.00003
Median 0.00138 −0.00004 0.00012 −0.00027 −0.00010 −0.00018 0.00084
Maximum 0.0903 0.0036 0.0334 0.0344 0.0037 0.0409 0.1096
Minimum −0.0926 −0.0202 −0.0693 −0.0473 −0.0151 −0.0407 −0.1378
Std. Dev. 0.0189 0.0010 0.0069 0.0069 0.0010 0.0069 0.0156
Skewness −0.4246 −5.9053 −0.4707 −0.1056 −2.8598 −0.3882 −0.5183
Kurtosis 5.7681 118.4267 10.7758 7.0867 42.6551 7.6880 13.7862
Observations 1577 1577 1577 1577 1577 1329 1529
the top four trading partners of China are the European Union (EU), the US, Japan,
and Hong Kong. Along with further reforms and opening-up in China, the US
dollar (USD), euro (EU), British pound (GBP), and Japanese yen (JPY) act as the
invoicing and settlement currencies in the Chinese foreign trade, while the trade
contracts between China and non-dollar countries become significantly more fre-
quent (Wang et al. 2012). Therefore, we select the Chinese Yuan (RMB)/USD,
RMB/EU, RMB/GBP, RMB/JPY, and RMB/HK (Hong Kong dollar) exchange rate
as the foreign exchange rate in the Chinese foreign exchange market. We chose the
Shanghai Stock Comprise Index (SSCI) is chosen as the variable of the Chinese
stock market, which is one of the most important stock market indices that reflect
fluctuations in the latter. The empirical data in this chapter choose the daily median
price of various exchange rates and the daily closing price of SSCI. The sample
interval is from July 22, 2005 to January 13, 2012.1 Common price data were
adopted after removing the data from different trading dates in the RMB exchange
and stock markets. The original sample data were obtained from FC Station 3.0.
We analyze the return series of the Chinese financial markets in this chapter.
Returns are computed by rt ¼ logðPt Þ logðPt1 Þ, with Pt being the closing price
index at time t. Table 6.1 provides the descriptive statistics of the returns of RMB/
USD, RMB/EU, RMB/GBP, RMB/JPY, RMB/HK, SSCI, and SP500. The mean
values are close to zero. For the seven return series, the skewness is less than zero,
and the kurtosis is larger than three. This result implies that these seven turn series
are fat-tailed and peaked. The standard deviations of SSCI and SP500 are higher
than those of foreign exchange rates. This result indicates stock markets fluctuate
more fiercely than RMB exchange markets.
To motivate the application of the detrended fluctuation analysis for assessing
asymmetric cross-correlation in financial series mentioned in this study, Fig. 6.1
plots the SSCI return series and presents the linear fitting for non-overlapping boxes
of 50 observations in a time interval2, which show the SSCI return series containing
1
The daily data of RMB/GBP can only be obtained from August 1, 2006 in FC Station 3.0.
2
This time interval is selected arbitrarily.
6.2 Data Description 119
0.1
SSCI return
0.05
-0.05
-0.1
0 200 400 600 800 1000 1200 1400 1600
SSCI return
0.02
-0.02
700 800 900 1000 1100
Fig. 6.1 Plots of SSCI return time series and the well-defined trend for boxes with 50
observations
0.2
SSCI,USD(-i) lag
0.15 SSCI,USD(+i) lead
SSCI,EU(-i) lag
0.1 SSCI,EU(+i) lead
SSCI,JPY(-i) lag
0.05 SSCI,JPY(+i) lead
SSCI,HK(-i) lag
0
SSCI,HK(+i) lead
16 111621263136 SSCI,GBP(-i) lag
-0.05
SSCI,GBP(+i) lead
-0.1 SSCI,SP(-i) lag
SSCI,SP(+i) lead
-0.15
Fig. 6.2 Cross correlogram of six pairs of financial time series in China. The data are the
corresponding probability values of the statistic with 36 lags
both positive and negative trends. The other return series contains positive and
negative trends, although our study does not present the figures involved.
Figure 6.2 provides the cross correlogram of six pairs of financial time series in
China. We find that the statistics of the cross-correlation are almost significant at
5% significant level except for several probability values of the corresponding
statistics. As a result, we suggest that the Chinese stock market and exchange
market as well as the Chinese and US stock markets are all cross-related. However,
this result needs to be verified by further testing based on MF-ADCCA.
6.3 Empirical Cross-Correlation Analysis

on Chinese Stock Market
In this section, we apply MF-ADCCA to investigate the asymmetric

cross-correlation between Chinese stock returns with the RMB/USD, RMB/EU,
RMB/GBP, RMB/JPY, and RMB/HK exchange rates in the Chinese foreign
exchange market and with US stock returns. Three aspects of this problem have to be
addressed. First, we assess these asymmetric cross-correlations when the Chinese
market is rising and falling. Second, we assess these asymmetric cross-correlations
when foreign exchange markets (resp. US stock market) are going up and down.
Third, we discuss the multifractal features of the asymmetric cross-correlation in
Chinese financial markets. For convenience, we use SSCI-RMB/USD to denote the
SSCI pair return series and the RMB/USD return series. Similarly, the SSCI-RMB/
EU, SSCI-RMB/GBP, SSCI-RMB/JPY, SSCI-RMB/HK, and SSCI-SP500 have
their corresponding meanings.
6.3.1 Asymmetric Test for Different Trends

of the Chinese Stock Market
Figure 6.3 presents the log-log plots of F2 ðnÞ, F2þ ðnÞ, and F2 ðnÞ versus n when the
trends of the Chinese stock market vary. Regarding the SSCI-RMB/EU,
-1.5 -1
Overall
log(F2(n))
-2 Upwards -1.5
Downwards
-2.5 -2
SSCI-RMB/USD SSCI-RMB/EU
-3 -2.5
0.8 1 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
-1 -1
log(F2(n))
-1.5 -1.5
-2 -2
SSCI-RMB/GBP SSCI-RMB/JPY
-2.5 -2.5
-1.5 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
0
log(F2(n))
-2 -1
-2.5 -2
SSCI-RMB/HK SSCI-SP500
-3 -3
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
log(n) log(n)
Fig. 6.3 Log-log plots of F2 ðnÞ, F2þ ðnÞ, and F2 ðnÞ versus n (different trends of the Chinese stock
market)
6.3 Empirical Cross-Correlation Analysis on Chinese Stock Market 121
þ
Table 6.2 Estimated scaling exponents H12 ð2Þ, H12 ð2Þ, and H12 ð2Þ for different trends of the
Chinese stock market
SSCI-RMB/ SSCI-RMB/ SSCI-RMB/ SSCI-RMB/ SSCI-RMB/ SSCI-RMB/
USD EU GBP JPY HK SP500
H12 ð2Þ 0.65751 0.60294 0.58740 0.55073 0.66574 0.57851
þ
H12 ð2Þ 0.62746 0.60815 0.67752 0.57309 0.63445 0.58227

H12 ð2Þ 0.66998 0.56249 0.50402 0.51238 0.68130 0.53922
DH12 ð2Þ −0.04252 0.04566 0.17350 0.06071 −0.04685 0.04305
Note The bold denotes the minimum of scaling exponents H
SSCI-RMB/GBP, SSCI-RMB/JPY, and SSCI-SP500 pairs for time scales less than
110 days, no visible scaling differences are determined for F2þ ðnÞ and F2 ðnÞ,
which indicates the cross-correlations are symmetric within these time scales.
However, for higher time scales, a significant departure from symmetry is observed
and at the same time there are less cross-correlation when the Chinese stock return
is down than when the Chinese stock return is up. By contrast, for the SSCI-RMB/
USD and SSCI-RMB/HK pairs, no significant departure from symmetry is
observed for the time scales, with the time scales going up for the time longer than
300 days.
þ
Table 6.2 presents the estimated scaling exponents, H12 ð2Þ, H12 ð2Þ, H12 ð2Þ, and
DH12 ð2Þ. Based on Table 6.2, the overall scaling exponents are significantly larger
than 0.5, indicating that long-range persistence exists in the overall
cross-correlation. However, the cross-correlation between the Chinese stock market
and the different financial markets present different features when the former has
þ
different trends. The estimated values of H12 ð2Þ and H12 ð2Þ of the SSCI-RMB/
USD, SSCI-RMB/EU, and SSCI-RMB/HK pairs are all significantly larger than
0.5. These findings indicate the cross-correlations between the Chinese stock
market and RMB/USD, RMB/EU, and RMB/HK exchange markets are persistent

whether the former is up or down. By contrast, the estimated values of H12 ð2Þ for
the SSCI-RMB/GBP, SSCI-RMB/JPY, and SSCI-RMB/SP500 pairs are not sig-
þ
nificantly larger than 0.5, whereas the estimated values of H12 ð2Þ are significantly
larger than 0.5. These results illustrate that the cross-correlation between the
Chinese stock market and the RMB/GBP, RMB/JPY exchange markets and the US
stock market may be random when the Chinese stock market is falling.
Nevertheless, the latter exchanges are persistent when the Chinese stock market is
þ
growing. Furthermore, Table 6.2 shows that H12 ð2Þ is not equal to H12 ð2Þ for each
pair of the Chinese financial markets, which reveals the cross-correlations are
asymmetric when the Chinese stock market presents different trends (up or down).
In addition, from the fifth row in Table 6.2, we can determine that the estimated
values of DH12 ð2Þ are smaller than zero for the SSCI-RMB/USD and SSCI-RMB/
HK pairs, whereas they are larger than zero for the other pairs. These findings
illustrate that the cross-correlations between the Chinese stock market and the
RMB/USD exchange market as well as between the Chinese stock market and the
RMB/HK exchange market are more persistent when the Chinese stock market is
falling than when it is rising. However, the cross-correlations between the Chinese
stock market and the RMB/EU, RMB/GBP, and RMB/JPY exchange markets as
well as that between the Chinese and the US stock markets are more persistent
when the Chinese stock market is rising than when it is falling.
6.3.2 Asymmetric Test for Different Trends

of the Other Financial Markets
Figure 6.4 presents the log-log plots of F2 ðnÞ, F2þ ðnÞ, and F2 ðnÞ versus n for the
different trends of foreign exchange markets and the US stock market. For time
scales lower than approximately 100 days, no visible scaling differences are
determined for F2þ ðnÞ and F2 ðnÞ for the SSCI-RMB/EU and SSCI-RMB/GBP
pairs, respectively. This result indicates that the cross-correlations within these time
scales are symmetric. However, higher time scales show a significant departure
from symmetry with the RMB exchange rate return lowering dynamics being less
cross-correlated than the RMB exchange rate return up dynamics. Similar to the
SSCI-SP500 pair, for time scales higher than approximately 110 days, an important
difference for symmetry is determined with the US stock return lowering dynamics
being less cross-correlated than the US stock return up dynamics. By contrast,
regarding the SSCI-RMB/USD, SSCI-RMB/JPY, and SSCI-RMB/HK pairs, as far
as the log-log plots of the F2 ðnÞ, F2þ ðnÞ, and F2 ðnÞ versus n are concerned, little
significant departure from symmetry is observed for each time scale. However, this
-1.5 -1
Overall
log(F2(n))
-2 Upwards -1.5
Downwards
-2.5 -2
-3 -2.5
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
-1 -1
log(F2(n))
-1.5 -1.5
-2 -2
SSCI-RMB/GBP SSCI-RMB/JPY
-2.5 -2.5
-1 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 0 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
log(F2(n))
-1
-2
-2
-3 -3
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
log(n) log(n)
Fig. 6.4 Log-log plots of F2 ðnÞ, F2þ ðnÞ, and F2 ðnÞ versus n (different trends of the RMB
exchange market and the US stock market)
þ
Table 6.3 Estimated scaling exponents H12 ð2Þ, H12 ð2Þ, and H12 ð2Þ for the different trends of the
other financial markets
SSCI-RMB/ SSCI-RMB/ SSCI-RMB/ SSCI-RMB/ SSCI-RMB/ SSCI-RMB/
USD EU GBP JPY HK SP500
H12 ð2Þ 0.66751 0.60294 0.58740 0.55073 0.66574 0.57851
þ
H12 ð2Þ 0.62608 0.60826 0.69559 0.55987 0.66207 0.59287

H12 ð2Þ 0.68085 0.57259 0.50222 0.52590 0.66640 0.55734
DH12 ð2Þ −0.05477 0.03567 0.19337 0.03397 −0.00433 0.03553
Note The bold denotes the minimum of scaling exponents H
þ
case is only an intuitive judgment. The scaling exponents, H12 ð2Þ, H12 ð2Þ, and their
difference, DH12 ð2Þ, are estimated to obtain precise results (Table 6.2).
þ
Table 6.3 shows the estimated scaling exponents, H12 ð2Þ, H12 ð2Þ, H12 ð2Þ, and
DH12 ð2Þ, for each pair of the financial time series for different trends of the RMB
exchange market or the US stock market. The overall scaling exponents, H12 ð2Þ, for
each pair of the financial data are significantly larger than 0.5, which indicate that
overall long-range cross-correlations are persistent. In addition, although the esti-
mated DH12 ð2Þ of the SSCI-RMB/HK pair does not significantly depart from zero,
the other values of DH12 ð2Þ significantly depart from zero. This result indicates that
the cross-correlation between the Chinese stock market and the RMB/HK exchange
market is not significantly asymmetric when the RMB/HK exchange rate market is
going up or down. However, the cross-correlations between the Chinese stock
market and the RMB/USD, RMB/EU, RMB/GBP, RMB/JPY exchange rate mar-
kets, and the US stock market are significantly asymmetric when the aforementioned

markets have different trends. Moreover, with the exception of the estimated H12 ð2Þ
of the SSCI-GBP and SSCI-RMB/JPY pairs, the estimated scaling exponents of the
rest are all significantly larger than 0.5. This result implies that the cross-correlation
between the Chinese stock market and RMB/JPY exchange markets may be random
when the RMB/JPY exchange rate market is falling, whereas it is persistent when the
RMB/JPY exchange rate market is rising. Unlike the RMB/JPY exchange market,
the cross-correlations between the Chinese stock market and the RMB/USD, RMB/
EU, RMB/GBP, RMB/HK exchange markets, and the US stock market are all
persistent, despite the trends of the RMB exchange markets and the US stock market.
þ
Table 6.3 also shows that the estimated scaling exponents H12 ð2Þ are larger than

H12 ð2Þ DH12 ð2Þ [ 0 for the SSCI-RMB/EU, SSCI-RMB/GBP, SSCI-RMB/JPY,
and SSCI-RMB/SP500 pair-wise. This finding indicates that the cross-correlations
between the Chinese stock market and the RMB/EU, RMB/GBP, RMB/JPY
exchange markets (rept. the US stock market) present stronger persistence when the
exchange markets (resp. the US stock market) are going up than when going down.
þ
On the contrary, the estimated scaling exponents H12 ð2Þ are smaller than H12 ð2Þ for
the SSCI-RMB/US and SSCI-RMB/HK pairs. These observations imply the
cross-correlations between the Chinese stock market and the RMB/US and RMB/
HK exchange markets present stronger persistence when the exchange market is
falling than when it is rising.
6.3.3 Multifractal of the Asymmetric Cross-Correlation
þ
Figure 6.5 illustrates the plots of H12 ðqÞ and H12 ðqÞ versus q for each pair of the
financial return series to detect the multifractal of the asymmetric cross-correlation
of each pair of financial markets mentioned in our research.
þ
Figure 6.5 shows that both H12 ðqÞ and H12 ðqÞ are not constant. This relationship
indicates that the asymmetric cross-relation between the Chinese stock market and
the RMB exchange market is multifractal, which is the same as the asymmetric
cross-relation between the Chinese and the US stock markets. Furthermore, the
þ
difference of the estimated H12 ðqÞ and H12 ðqÞ seems to vary with increasing q. We
compute the values of DH12 ðqÞ when the Chinese stock market has different trends
and when the other financial markets have different trends to detect accurately
(Figs. 6.6 and 6.7).
Figures 6.6 and 6.7 show similar results when either the Chinese stock market or
the other financial markets have a different trend. The values of DH12 ðqÞ decrease
with q varying from q\0 to q [ 0 for the SSCI-RMB/USD and SSCI-RMB/HK
pairs, whereas the values of DH12 ðqÞ increase with q varying from q\0 to q [ 0
1.5 0.8
H12(q)
1 0.6
0.5 0.4
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
1 0.8
H12(q)
0.5 0.6
SSCI-RMB/GBP SSCI-RMB/JYP
0 0.4
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
1 1
0.8
H12(q)
0.5
0.6
0.4 0
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
q Upwards(S) Downwards(S) Upwards(R) Downwards(R) q
þ
Fig. 6.5 Plots of H12 ðqÞ and H12 ðqÞ versus q varying from −10 to 10. “Upwards(S)” and
“Downwards(S)” denote the up and down trends of the Chinese stock market; “Upwards (R)” and
“Downwards (R)” denote the up and down trends of the RMB exchange market or the US stock
market. For the pair of SSCI-RMB/HK, the n in MF-ADCCA method is selected such that
10 n N=4, instead of 5 n N=4, for q\0, because the non-normal value (NaN.) appears in
the computing process when n 9
0.25
0.2
0.15
0.1
DELTA(H12(q))
SSCI - RMB/USD
SSCI - RMB/EU
0.05 SSCI - RMB/GBP
SSCI - RMB/JPY
0 SSCI - RMB/HK
SSCI - RMB/SP500
-0.05
-0.1
-0.15
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 6.6 DH12 ðqÞ when the Chinese stock market has different trends
0.25
SSCI - RMB/USD
SSCI - RMB/EU
0.2
SSCI - RMB/GBP
SSCI - RMB/JPY
0.15 SSCI - RMB/HK
SSCI - SP500
0.1
DELTA(H12(q))
0.05
-0.05
-0.1
-0.15
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 6.7 DH12 ðqÞ when the RMB exchange market or the US stock market has different trends
for the SSCI-RMB/EU, SSCI-RMB/GBP, and SSCI-RMB/SP500 pairs. However,

for the SSCI-RMB/JPY pair, the values of DH12 ðqÞ decrease but change little with q
varying from q\0 to q [ 0. Hence, As a result, the asymmetries of the
cross-correlations between the Chinese stock market and the RMB/USD and the
RMB/HK exchange markets are less persistent for large price fluctuations than for
small price fluctuations. By contrast, the asymmetries of the cross-correlations
between the Chinese stock market and the RMB/EU and RMB/GBP exchange
markets are more persistent for large price fluctuations than for small ones. Besides,
the asymmetries of the cross-correlation between the Chinese and US stock markets
are stronger for larger price fluctuations than for small ones. In addition, for the
cross-correlation between the Chinese stock market and the RMB/USDRMB/UK
exchange market, the changes of the asymmetry persistence are larger when the
Chinese stock market has different trends than when the Chinese exchange market
has different trends.
6.4 Conclusions
We developed an MF-DCCA extension to explore the existence of asymmetries in

the cross-correlation scaling behavior of two time series. The MF-DCCA version
separates the positive and negative trends to in studying the individual contributions
to the overall scaling behavior. The empirical results on the Chinese financial
markets indicate that the asymmetries are scale-dependent, which means that the
cross-correlation scaling behavior for several scales is symmetric, but asymmetric
for other time scales. The empirical results clearly indicate the existence of different
cross-correlation properties when the trend of the financial market is going up or
down. The existence of asymmetric cross-correlation adds new clues into the
complex relationship of the two time series from real systems. This finding suggests
that the mechanisms underlying two complex systems act in a different way
depending on the directionality of the dynamics. For financial markets, these
scaling behavior features will help reveal the financial markets risks and to con-
struct a better portfolio. Furthermore, the presence of asymmetric cross-correlations
implies biased behavior of market agents. In turn, violation of the efficient market
hypothesis (EMH) might be also implied.
In addition, by means of this approach, our empirical study of the Chinese
financial markets presents the following results.
(1) Asymmetries exist in the cross-correlation of the Chinese stock market and the
RMB exchange markets (except the cross-correlation between the Chinese
stock market and the RMB/HK exchange market when the RMB/UK exchange
market has different trends), and the asymmetric cross-correlations are multi-
fractal. These findings might implies that the Chinese stock market and the
RMB exchange markets are not strong efficient.
(2) Cross-correlations between the Chinese stock market and the RMB/USD
exchange market are more persistent when any one of the markets is falling
than when it is rising. On the contrary, accounting for the Chinese stock market,
the RMB/EU, RMB/GBP, RMB/JPY exchange markets, and the US stock
market, the cross-correlations between the Chinese stock market and any one of
the other markets are more persistent when any one of these financial markets is
rising than when falling. However, the cross-correlation between the Chinese
stock market and the RMB/HK exchange market is more persistent when the
Chinese stock market is falling than when rising.
(3) Asymmetries of the cross-correlations between the Chinese stock market and
the RMB/USD and RMB/HK exchange markets are less persistent for large
price fluctuations than for small price fluctuations. However, the asymmetries
of the cross-correlations between the Chinese stock market and the RMB/EU as
well as the RMB/GBP exchange markets are more persistent for large price
fluctuations than for small ones. Besides, the asymmetries of the
cross-correlation between the Chinese stock market and the US stock market
are stronger for large price fluctuations than for small ones.
References 127
References
J. Alvarez-Ramirez, E. Rodriguez, J.C. Echeverria, A DFA approach for assessing asymmetric

A. Ang, G. Bekaert, International asset allocation with regime shifts. Rev. Financial Stud. 15,
1137–1187 (2002)
A. Ang, J. Chen, Asymmetric correlations of equity portfolios. J. Financial Econ. 63, 443–494
(2002)
L. Ding, H. Miyake, H. Zou, Asymmetric correlations in equity returns: a fundamental-based
explanation. Appl. Financial Econ. 21, 389–399 (2011)
Y. Hong, J. Tu, G. Zhou, Asymmetric in stock returns: statistical tests and economic evaluation.
Rev. Financial Stud. 20, 1547–1581 (2007)
F. Longin, B. Solnik, Extreme correlation of international equity markets. J. Financial 56, 649–676
(2001)
C.K. Peng, S.V. Buldyrev, S. Havlin et al., Mosaic organization of DNA nucleotides. Phys. Rev.
E 49, 1685–1689 (1994)
nonstationary time series. Phys. Rev. Lett. 100, 084102 (2008)
A. Taamouti, G. Tsafack, Asymmetric effects of return and volatility on correlation between
international equity markets (2009)
D.-H. Wang, X.-W. Yu, Y.-Y. Suo, Statistical properties of the yuan exchange rate index. Phys.
A 391, 3502–3512 (2012)
W.-X. Zhou, Multifractal detrended cross-correlation analysis for two nonstationary signals. Phys.
Rev. E 77, 066211-1–066211-4 (2008)
Chapter 7
Asymmetric DCCA Cross-Correlation
Coefficient
The explicit linkage between spot and futures prices is highlighted in financial
theory by both non-arbitrage and asset pricing theory (Nicolau 2012). Many studies
have considered the relationship between the spot and futures prices of the crude oil
market; however, most of these works used traditional methods such as cointe-
gration [as proposed by Engle and Granger (1987) and Johansen (1988)] or the
Granger test (1969). The generalized autoregressive conditional heteroskedasticity,
linear vector error correlation, and nonlinear threshold vector error correlation
models are also employed to investigate the relationship between oil spot and
futures markets (Wang and Wu 2013).
Complex systems with nonlinear and complex behavior, including self-affinity,
have become a hot research topic. Among the most popular methods is detrended
cross-correlation analysis (DCCA), which was proposed by Podobnik and Stanley
(2008). DCCA can quantify long-range cross-correlation between two
non-stationary time series. Wang et al. (2015) examined the cross-correlations of
the West Texas Intermediate (WTI) crude oil spot and futures markets using the
DCCA method. This method not only quantifies the quantitative cross-correlation
level, but it also evaluates the presence of cross-correlation and its fractal
characteristic.
When two time series are recorded simultaneously and are of the same length,
the Pearson correlation coefficient measurement method is most widely used.
However, this coefficient is weak (Wilcox 2012) and can be misleading if outliers
are present, as in real-world data that are characterized by a high degree of non-
stationarity (Devlin et al. 1975). Thus, a new cross-correlation coefficient called the
DCCA cross-correlation coefficient was proposed by Zebende (2011). It is based on
overlapping and nonoverlapping windows. The DCCA coefficient is then measured
in nonstationary time series analyses, such as financial markets, meteorology, and
traffic. And in this chapter, we discuss the asymmetry of the DCCA
cross-correlation coefficient when either of the two markets is either rising or
declining, as well as the differences between the uptrend/downtrend spot-futures
markets and carbon-energy markets.
130 7 Asymmetric DCCA Cross-Correlation Coefficient
7.1 Methodology
7.1.1 DCCA Cross-Correlation Coefficient
The DCCA method can be described as follows:

Given two time series fx1 ðiÞg and fx2 ðiÞg of the same length N, we compute two
P P
integrated signals R1 ðkÞ ki¼1 x1 ðiÞ and R2 ðk Þ ki¼1 x2 ðiÞ, where
k ¼ 1; 2; . . .N. Then, the entire time series is divided into N n overlapping
windows, each containing n þ 1 values. Each window ranges from i to i + n.
The local trends R ~ 2;i ðk Þ are evaluated by the least fit of data, where
~ 1;i ðkÞ and R
i k i þ n. Therefore, the covariance of the residuals can be calculated as
1 X iþn
~ 1;i ðk Þ R2 ðkÞ R
~ 2;i ðk Þ

2
fDCCA ðn; iÞ R 1 ðk Þ R ð7:1Þ
n þ 1 k¼i
The detrended covariance function is calculated by summing up all overlapping

N − n boxes of size n:
X
N n
2
FDCCA ðnÞ ðN nÞ1 2
fDCCA ðn; iÞ ð7:2Þ
i¼1
When R1 ðkÞ ¼ R2 ðkÞ; FDCCA

2
ðnÞ is extracted from the detrended variance
2
FDFA ð nÞusing the DFA method.
The DCCA cross-correlation coefficient proposed by Zebende (2011) is defined
as follows:
2
FDCCA ð nÞ
qDCCA ¼ ð7:3Þ
FDFA1 ðnÞFDFA2 ðnÞ
where FDFA1 and FDFA2 are the detrended variance functions FDFA of the two time
series.
Based on the multifractal asymmetric detrended cross-correlation analysis
method and the asymmetric DCCA cross-correlation coefficient proposed by Cao
et al. (2014), we set R ~ 1;i ðk Þ ¼ ai þ bi k (where 1 i N n); thus, R1 ðkÞ has a
positive (or negative) trend when bi [ 0 (or bi \0). The asymmetric
cross-correlation between fx1 ðiÞg and fx2 ðiÞg as a result of the different trends of
one series (e.g., fx1 ðiÞg) is expressed as follows:
7.1 Methodology 131
When fx1 ðiÞg is trending upward,
1 XNn
signðbi Þ þ 1 2
þ ð nÞ FDCCA ðnÞ ð7:4Þ
2
FDCCA þ
M i¼1 2
!1=2
1 XN n
signðbi Þ þ 1 2
FDFA1 þ ðnÞ ¼ þ
FDFA1 ðnÞ ð7:5Þ
M i¼1 2
Hence,
2
FDCCA þ
qup ¼ ð7:6Þ
FDFA1 þ ðnÞFDFA2 þ ðnÞ

where FDFA2 þ is calculated with the same divided boxes as FDFA1 þ .
When fx1 ðiÞg is trending downward,
1 XN n
½signðbi Þ 1 2
ð nÞ FDCCA ðnÞ ð7:7Þ
2
FDCCA
M i¼1 2
!1=2
1 XN n
½signðbi Þ 1 2
FDFA1 ðnÞ ¼ FDFA1 ðnÞ ð7:8Þ
M i¼1 2
Hence,
2
FDCCA
qdown ¼ ð7:9Þ

þ
where FDFA2 is calculated using the same divided boxes as FDFA1 :M ¼
PNn signðbi Þ þ 1
P Nn ½signðbi Þ1
i¼1 2 and M ¼ i¼1 2 represent the number of series with
positive and negative trends, respectively. i ¼ 1; 2; . . .; N n. If bi 6¼ 0, then
M þ þ M ¼ N n.
Meanwhile, the values of qDCCA ; qup and qdown range from −1 to 1. If the value
is equal to zero, then the two series are not cross-correlated. If the value is equal
to 1, then a perfect cross-correlation exists. If the value equals −1, then the
anti-cross-correlation is perfect.
7.1.2 Statistical Tests
We test the significance of qDCCA ; qup and qdown according to the methods proposed
by Podobnik et al. (2011), Kristoufek (2014), and Cao et al. (2014).
We simulate artificial time series with a length of N, which is similar to the

actual sample interval, following two ARFIMA (0, d, 0) processes: X1 ðiÞ ¼
P1 P1 Cðn þ d Þ
n¼0 an ðd1 Þuin and X2 ðiÞ ¼ n¼0 an ðd2 Þgin , where an ðd Þ ¼ Cðn þ 1ÞCðd Þ ; h/i i ¼
2 2
hgi i ¼ 0; /i ¼ gi ¼ 1 and h/i gi i ¼ q/g . Moreover, q/j value ranges from −1
to 1. d is related to the DFA exponent as indicated in a ¼ 0:5 þ d.
Then, we divide X1 ðiÞ into M ¼ ½N=v groups Z(t), where v is the length of each
group. We define the ordinate of the linear least square fit of each group as denoted
by ZðtÞ ¼ e þ fv, where 1 t v. When f > 0, the time series is in an uptrend;
otherwise, it is in a downtrend. If we assume that a probability p of the groups is in
an uptrend based on random samples of the original series, we randomly add
M p groups in an uptrend to the original series. Therefore, the probability
1 − p of the groups is in a random downtrend, and we add M (1 − p) groups in a
downtrend to the original series. The effect of the short part that is indivisible
(i − M N) is slight because the times are all detrended in qDCCA . Thus, the short
part is replaced with 0. In this manner, we obtain one pair of time series EðiÞ with
uptrend and downtrend, as well as X2 ðiÞ. If we simulate 200 repetitions and cal-
culate the qDCCA values among them, then each setting can be analyzed with respect
to the 2.5th and the 97.5th quintiles, that is, the 95% confidence interval. This
interval is the standard deviation of the estimated DCCA coefficient.
7.2 Empirical Analysis on Crude Oil Spot and Futures

Markets
In this section, we use the DCCA cross-correlation coefficient method to study the
quantitative cross-correlation level between the spot and futures market of crude oil.
We also discuss the asymmetry of the DCCA cross-correlation coefficient when
either of the two markets is either rising or declining, as well as the differences
between the uptrend/downtrend spot and futures markets. Our work aims to achieve
two objectives. Firstly, this study is the first to apply the DCCA cross-correlation
coefficient in the analysis on the crude oil spot and futures markets. Secondly, based
on the definition of asymmetric degree and significance statistical test, we also
discuss the asymmetric qDCCA ðnÞ value when either of the two markets is either
rising or declining, as well as the differences in the qDCCA ðnÞ values of the uptrend/
downtrend spot and futures markets.
7.2.1 Data and Descriptive Statistics
The daily closing spot prices of WTI and Europe Brent crude oil are considered
(denoted as S1 and S2, respectively). The daily closing prices of futures contracts
traded at the New York mercantile exchange with maturities of one, two, three, and
7.2 Empirical Analysis on Crude Oil Spot and Futures Markets 133
0.8
0.6
0.4
r
0.2
-0.2
-0.4 S1 S2 F1 F2 F3 F4
0 1000 2000 3000 4000 5000 6000 7000

sample points
Fig. 7.1 Return of crude oil spot and futures (The values of SEB, F1, F2, F3 and F4 in futures
have been plus 0.2, 0.4, 0.6, 0.8 and 1.0 respectively; The curve from bottom to above is of S1, S2,
F1, F2, F3 and F4, respectively.)
Table 7.1 Descriptive statistics

S1 S2 F1 F2 F3 F4
Mean 0.000232 0.000265 0.000238 0.000240 0.000241 0.000241
Median 0.000771 0.000386 0.000586 0.000553 0.000588 0.000591
Maximum 0.188677 0.181297 0.164097 0.137885 0.121150 0.114684
Minimum −0.406396 −0.361214 −0.400478 −0.384071 −0.328206 −0.284272
Std. dev. 0.024758 0.023092 0.024050 0.021201 0.019718 0.018836
Skewness −0.819869 −0.697470 −0.848652 −1.050608 −0.896773 −0.759841
Kurtosis 19.03825 17.86899 19.18173 22.32210 17.33226 14.13844
Jarque–Bera 71692.96*** 61519.88*** 73021.10*** 104198.4*** 57547.21*** 34858.23***
Note *** denotes 1% significant level
four months are selected as the futures prices (denoted as F1, F2, F3, and F4,
respectively). The sample period lasts from May 21, 1987 to January 13, 2014. The
original data originate from the Energy Information Administration.
Let Pt be the price of crude oil on day t. We choose logarithmic return rt ¼
logðPt Þ logðPt1 Þ as the empirical data, where logðÞ denotes the nature loga-
rithm. Thus, the length of all of the return series of the spot and the futures is 6620.
Figure 7.1 plots the returns of the crude oil spot and the futures, and Table 7.1
displays the descriptive statistics of these returns. The figure indicates that all series
exhibit similar characteristic fluctuations. For all of the time series presented in
Table 7.1, the skewness deviates from 0, whereas the kurtosis diverges from 3.
These results denote significant deviations from normality. The Jarque–Bera
statistics 29 are significant at a 1% significance level, thereby suggesting that the
normality assumption of returns and volatilities can be rejected for all series.
7.2.2 Estimation of the Cross-Correlation Coefficient
7.2.2.1 DCCA Cross-Correlation Coefficient of the Entire Interval
Figure 7.2 shows the DCCA cross-correlation coefficient qDCCA between the returns
of WTI and the futures with maturities. Figure 7.3 depicts the qDCCA between the
returns of Europe Brent crude oil and the futures with maturities. The qDCCA values
1 1
0.95 0.95
0.9 0.9
d d
0.85 0.85
S1F1 u
S1F2
u
0.8 0.8
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
1 1
0.95 0.95
0.9 0.9
d d
0.85 0.85
S1F3 S1F4
u u
0.8 0.8
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
Fig. 7.2 DCCA cross-correlation coefficient qDCCA between S1 and different futures
1 1
0.95 0.95
0.9 0.9
d d
0.85 0.85
S2F1 u S2F2 u
0.8 0.8
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
1 1
0.95 0.95
0.9 0.9
d d
0.85 0.85
S2F3 u S2F4 u
0.8 0.8
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
Fig. 7.3 DCCA cross-correlation coefficient qDCCA between S2 and different futures
between the spot and futures markets are all greater than 0.8 in different situations,
thus indicating a significant relationship between the two markets. The qDCCA value
in Fig. 7.3 is slightly lower than that in Fig. 7.2, that is, the qDCCA values between
WTI crude oil and the futures are slightly stronger than those between Europe Brent
crude oil and the futures. The origins and product qualities of the WTI and Europe
Brent crude oils differ, which may induce variations in the cross-correlation.
Meanwhile, the qDCCA is small and fluctuates considerably when the maturity
date is far from the spot. Given a long maturity period, the differences between qu
and qd are enhanced when the spot market is either in uptrend or downtrend.
Furthermore, the property is particularly significant between WTI crude oil and the
futures. The results are consistent with the general understanding that mispricing is
reduced with the shortening of the maturity period between the spot and the futures
markets.
Figures 7.4 and 7.5 display the qDCCA values between the returns of futures with
maturities and of WTI/Europe Brent crude oil, respectively. The overall results of
qDCCA are similar to those in Figs. 7.2 and 7.3. The differences between qu and qd
are also highlighted when the futures market is in uptrend or downtrend with a long
maturity period. Furthermore, the property is increasingly significant between the
futures and WTI crude oil. qu and qd vary as well when spot market trends are
different from those of futures markets.
We must then conduct a statistical test to develop a criterion with which we can
determine whether the qDCCA , qu , and qd in Figs. 7.2, 7.3, 7.4, and 7.5 are sig-
nificant or not. For the statistical test, we select a qDCCA value equal to 0 from the
simulated artificial time series. The length of each group v is 10, and the probability
p for the uptrend is 0.5. Given the small difference in the DFA exponents with
1 1
0.95 0.95
0.9 0.9
d d
0.85 0.85
F1S1 u F2S1 u
0.8 0.8
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
1 1
0.95 0.95
0.9 0.9
d d
0.85 0.85
F3S1 u F4S1 u
0.8 0.8
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
Fig. 7.4 DCCA cross-correlation coefficient qDCCA between futures and different S1 values
1 1
0.95 0.95
0.9 0.9
0.85 0.85
0.8 d 0.8 d
0.75 u 0.75 u
F1S2 F2S2
0.7 0.7
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
1 1
0.95 0.95
0.9 0.9
0.85 0.85
0.8 d
0.8 d
0.75 u 0.75 u
F3S2 F4S2
0.7 0.7
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
Fig. 7.5 DCCA cross-correlation coefficient qDCCA between futures and different S2 values
1.2
025
1
u025
0.8 d025
975
0.6
u975
0.4 d975
0.2
-0.2
-0.4
-0.6
0 200 400 600 800 1000 1200 1400 1600 1800
n
Fig. 7.6 Confidence intervals at 5 and 2.5% significance levels
regard to the returns of spot and the futures markets, we consider the mean DFA
exponents of all markets (a = 0.461382).
Figure 7.6 exhibits the criteria for qDCCA , qu , and qd . The approximate range is
[−0.6, 0.6]. Therefore, the DCCA cross-correlation coefficients are all significant
given that the values of the qDCCA , qu , and qd in Figs. 7.2, 7.3, 7.4 and 7.5 are all
>0.8.
7.2.2.2 Asymmetric Characteristic of the DCCA Cross-Correlation

Coefficient
To measure the asymmetric degree when one market is either in uptrend or

downtrend, we define
Dq ¼ jqu qd j ð7:10Þ
and
(
DqSc Fr ¼ qu Sc Fr qd Sc Fr
; c ¼ 1; 2; r ¼ 1; 2; 3; 4 ð7:11Þ
DqF S ¼ qu Fr Sc qd Fr Sc
r c
To compute the difference [denoted as DðDqÞ] of the asymmetric degree for

futures with different maturities, we establish
(
DðDqSc Fr Þ ¼ DqSc Fr DqSc F1
; c ¼ 1; 2; r ¼ 2; 3; 4 ð7:12Þ
DðDqFr Sc Þ ¼ DqFr Sc DqF1 Sc
Figure 7.7 shows the results of the asymmetric characteristic between the spot
and futures markets when the former is in either uptrend or downtrend. As per
Figs. 7.7a, c, the cross-correlation coefficients are asymmetrical when the trends of
(a) (b)
0.08 0.06
Δρ Δ(Δρ )
S1F1
S1F1
Δρ Δ(Δρ )
0.06 S1F2 0.04 S1F2
Δρ Δ(Δρ )
S1F3 S1F1
Δρ
S1F4
Δ(Δρ)
Δρ
0.04 0.02
0.02 0
0 -0.02
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
(c) (d)
0.1 0.06
Δρ Δ(Δρ )
S1F1 S2F1
Δ(Δρ )
0.08 Δρ 0.04 S2F2
S1F2 Δ(Δρ )
Δρ
Δρ S2F3
0.06 S1F3 0.02
Δρ
Δ(Δρ)
S1F4
0.04 0
0.02 -0.02
0 -0.04
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
Fig. 7.7 Asymmetric characteristic between the spot and the futures markets
(a) (b)
0.06 0.06
Δρ
F1S1
0.05 Δρ 0.04
F2S1
0.04 Δρ
F3S1
0.02
Δ(Δρ)
Δρ
0.03
0
0.02 Δ(Δρ )
F2S1
-0.02 Δ(Δρ )
0.01 F3S1
Δ(Δρ )
F4S1
0 -0.04
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
(c) (d)
0.12 0.06
Δρ
0.1 F1S2 0.04
Δρ
F2S2
0.08 Δρ 0.02
F3S2
Δρ
F4S2
Δ(Δρ)
Δρ
0.06 0
0.04 -0.02 Δ(Δρ )

F2S2
Δ(Δρ )
0.02 -0.04 F3S2
Δ(Δρ )
F4S2
0 -0.06
0 500 1000 1500 2000 0 500 1000 1500 2000
n n
Fig. 7.8 Asymmetric characteristic between the futures and spot markets
the spot market vary. However, the differences between qu and qd are slight.
Figure 7.7b shows that when the trends of the WTI market are different, the degree
of asymmetry increases with a long maturity period. However, this property varies
slightly when the trends of the market of Europe Brent crude oil are different, as
indicated in Fig. 7.7d. DðDqÞ is negative in some situations, that is, when the
asymmetry weakens as the maturity period is lengthened, because both the contracts
of WTI and of the selected futures are drawn up in the American market. Hence,
their relationship is close and unchanging. Market uptrend or downtrend is also
affected by the economy or the policies of Europe, as well as other accidental
factors. By contrast, Europe Brent crude oil is traded in London. Thus, the regu-
larity of the asymmetric characteristic is vague.
Figure 7.8 displays the results for the asymmetric characteristic between the
futures and the spot markets when the futures market is in either uptrend or
downtrend. In particular, Figs. 7.8a, c depict a slight asymmetry between qu and qd
when the trends of futures markets vary. Figure 7.8b suggests that the degree of
asymmetry between WTI and the futures market increases as the maturity period is
lengthened when the trends of futures markets differ. However, the characteristic
between Europe Brent crude oil and the futures market remains unclear.
7.2.2.3 Differences Between the Spot and the Futures Market Under
Various Trends
We define Dqur as the differences in qDCCA between the uptrend spot market and
the futures markets. Dqdr denotes the differences in qDCCA between the downtrend
spot and the futures markets.
(
Dqur ¼ qu ðSc Fr Þ qu ðFr Sc Þ
c ¼ 1; 2 r ¼ 1; 2; 3; 4: ð7:13Þ
Dqdr ¼ qd ðSc Fr Þ qd ðFr Sc Þ
Figure 7.9 presents the differences in qDCCA between the uptrending spot and the
futures market. Specifically, Fig. 7.9a shows that most Dqur values are >0, thus
implying that qDCCA increases more when the WTI market is in an uptrend than
when the futures market is.
Although the Dqur value between the Europe Brent crude oil and the futures
markets does not reflect the property, this value is extremely volatile and still
represents the presence of differences in qDCCA value between the uptrending spot
and futures markets.
Figure 7.10 exhibits the differences in qDCCA between the downtrending spot and
futures markets. In particular, Fig. 7.10a suggests that most Dqdr values are <0;
thus, qDCCA is greater when the futures market is downtrending than when the WTI
market is.
Furthermore, the Dqdr between the Europe Brent crude oil and futures markets
does not reflect the property. However, this value is extremely volatile and still
(a)
0.04
0.02
uσ
0
Δρ
-0.02
ρu(S1F1)-ρ u(F1S1) ρu(S1F2)-ρ u(F2S1) ρu(S1F3)-ρ u(F3S1) ρu(S1F4)-ρ u(F4S1)
-0.04
0 200 400 600 800 1000 1200 1400 1600 1800
ν
(b)
0.06
ρu(S2F1)-ρ u(F1S2) ρu(S2F2)-ρ u(F2S2) ρu(S2F3)-ρ u(F3S2) ρu(S2F4)-ρ u(F4S2)
0.04
uσ
0.02
Δρ
-0.02
-0.04
0 200 400 600 800 1000 1200 1400 1600 1800
ν
Fig. 7.9 Differences in qDCCA between the uptrending spot and futures markets
(a)
0.04
ρd(S1F1)-ρ d(F1S1) ρd(S1F2)-ρ d(F2S1) ρd(S1F3)-ρ d(F3S1) ρd(S1F4)-ρ d(F4S1)
0.02
dσ
0
Δρ
-0.02
-0.04
0 200 400 600 800 1000 1200 1400 1600 1800
n
(b)
0.1
ρd(S2F1)-ρ d(F1S2) ρd(S2F2)-ρ d(F2S2) ρd(S2F3)-ρ d(F3S2) ρd(S2F4)-ρ d(F4S2)
0.05
dσ
0
Δρ
-0.05
-0.1
0 200 400 600 800 1000 1200 1400 1600 1800
n
Fig. 7.10 Differences in qDCCA between the downtrending spot and futures markets
indicates that qDCCA values are generated between the downtrending spot and
futures markets.
qDCCA value is high in an uptrending WTI market and in a downtrending futures

market potentially because the futures market is based on the spot market and
utilizes the price discovery function. Therefore, the rise in the futures market when
the spot market is in an uptrend strengthens the relationship between the two
markets more than the uptrending futures market. However, most investors are
risk-averse; hence, the declining market fluctuates with increased volatility. The
downtrend of the futures market rapidly affects the spot market. Thus, qDCCA is
increased in a downtrend futures market.
7.2.3 Conclusions
This study investigates the relationship between crude oil spot (WTI and Europe
Brent crude oil) and futures markets using the qDCCA method.
First, the crude oil spot and the futures market are significantly cross-correlated.
Moreover, qDCCA ðnÞ is small and it fluctuates strongly when the maturity period
from the spot is lengthened.
Second, we discuss the asymmetric qDCCA when either of the two markets is
either rising or declining. qDCCA displays an asymmetric characteristic. The degree
of asymmetry also increases with a long maturity date period when the trends of
WTI and of futures markets differ.
Third, we study the differences in qDCCA between the uptrending/downtrending

spot and futures markets. The qDCCA differentiates these downtrending markets.
Moreover, the variation in qDCCA increases when the WTI market is in an uptrend
and the futures market is in a downtrend.
7.3 Empirical Analysis on Carbon and Energy Market
In this section, we will investigate the asymmetric structure between the carbon and
energy markets from two aspects of different trends (up or down) and
volatility-transmission direction using asymmetric detrended cross-correlation
analysis (DCCA) cross-correlation coefficient test, multifractal asymmetric
DCCA (MF-ADCCA) method, asymmetric volatility-constrained correlation met-
ric, and time rate of information-flow approach.
7.3.1 Data
The EU ETS for carbon emission trading products are artificially composed of three
separate stages. However, different periods are significantly different in terms of the
distribution mechanism of carbon quota and the policy of global climate change.
Hence, we chose the relatively “mature period” from January 1, 2008 to December
31, 2012. For carbon market, we use EU emission allowance (EUA) daily future
price series provided by the ECX, which accounts for about 90% of the total daily
future market transaction volume. Data are derived from the Wind database. As for
energy markets, we chose daily future prices series of ICE Brent crude oil (Brent),
Richards bay coal (Coal), UK base electricity (Electricity), and UK natural gas
(Gas), which are traded in the similar platform, as representatives. Energy future
market data are obtained from the Intercontinental Exchange (www.theice.com).
Both pairs of carbon and energy market daily price series include 1283 data points.
Returns are computed by, with being the daily closing price. Moreover, we have
converted the units of all different markets into dollars per ton.
To unveil detailed statistical properties, we display the basic statistics of these
time series in Table 7.2. We find that the returns of the carbon and energy markets
possess peakedness and fat-tails. Moreover, Jarque–Bera test results reject the null
hypothesis at 1% significance level. Therefore, the logarithmic returns series of the
carbon and energy markets does not follow the efficient market hypothesis, which
assumes a normal distribution.
Table 7.2 Statistics of the series

Series Mean Std. Dev. Skewness Kurtosis Jarque–Bera Probability
EUA −0.0009 0.0284 0.1788 6.7692 765.7060 0.0000
Brent 0.0001 0.0240 −0.0847 6.6106 697.8926 0.0000
Coal −0.0001 0.0199 0.3577 36.3409 59406.1500 0.0000
Electricity −0.0001 0.0194 0.4804 18.7372 13278.4800 0.0000
Gas 0.0007 0.0313 2.7611 30.6810 42558.6700 0.0000
7.3.2 Asymmetric Cross-Correlation Coefficient Test
We employ the asymmetric DCCA cross-correlation coefficient method proposed in

Sect. 7.1 to quantify the significance level of the cross-correlation between the
carbon and energy markets in this subsection.
As shown in Fig. 7.11, the red curves represent qDCCA critical value qc between
carbon and brent, coal, electricity and gas, respectively. The blue curves qcp and
qcm in Fig. 7.11 denote 95% confidence interval for qDCCA . In addition, Dqc in
Fig. 7.12 represents the asymmetric characteristic of qDCCA when energy markets
are in an uptrend or a downtrend. Most qDCCA are significant at 95% confidence
interval, except for a small part of the results in Fig. 7.11b (50 < n < 100) and
(a) EUA-Brent
(b) EUA-Coal
0.6 0.6
0.4 0.4
0.2 0.2
ρ
0 0
-0.2 -0.2
-0.4 -0.4
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
n n
(c) EUA-Electricity
(d) EUA-Gas
1 0.6
0.8
0.4
0.6
0.4 0.2
ρ
0.2 0
0
-0.2
-0.2
-0.4 -0.4
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
n n
ρ ρ ρ
c c-p c-m
Fig. 7.11 qDCCA critical value qc between carbon and energy markets
7.3 Empirical Analysis on Carbon and Energy Market 143
(a) EUA-Brent
(b) EUA-Coal
1 0.5
0.5 0
ρ
ρ
0 -0.5
-0.5 -1
-1 -1.5
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
n n
(c) (d)
EUA-Electricity EUA-Gas
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
ρ
ρ
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
-1 -1
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
n n
Δρ Δρ Δρ
c c-up c-down
Fig. 7.12 Dqc between carbon and energy markets
Fig. 7.11d (150 < n < 200). This result indicates a positive correlation between
carbon and returns in the energy markets, especially carbon and electricity markets
and a negative correlation between carbon and volatilities in the energy markets in
certain circumstances. Meanwhile, most Dqc are significant at 5% significance
level, which signifies an asymmetric characteristic from the cross-correlation
between carbon and returns in the energy markets.
We also study the qDCCA between carbon and returns in the energy markets in an
uptrend or a downtrend, as shown in Fig. 7.13, where the pink curves represent
qDCCA , red and green curves represent uptrend value qcup , and downtrend value
qcdown between carbon and brent, coal, electricity and gas, respectively. The blue
curves, qupp and qupm , black curves, qdownp and qdownm , in Fig. 7.13 denote
95% confidence interval for qcup and qcdown . We also find that most qup and qdown
are significant at 5% significance level, except for a small part of some time scales
(50 < n < 150) for qup and qdown between carbon and coal, as well as qup between
carbon and gas. This finding also indicates that a positive cross-correlation may
exist between carbon and energy markets not only for returns but also for
volatilities. This outcome may be a result of energy prices as the most important
predictors of carbon prices (Mansanet et al. 2007; Bunn and Fezzi 2007; Convery
and Redmond 2007). Moreover, power plants can switch between their fuel inputs,
such as oil, coal, and natural gas. Thus, the price change between carbon trading
and energy markets is closely related.
EUA-Brent EUA-Coal
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
ρ
ρ
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
n n
EUA-Electricity EUA-Gas
1 0.8
0.6
0.5
0.4
0.2
0
ρ
-0.2
-0.5
-0.4
-1 -0.6
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
n n
ρ ρ ρ ρ ρ ρ ρ
c-up up-p up-m c-down down-p down-m c
Fig. 7.13 qDCCA uptrend and downtrend value between carbon and energy markets, respectively
7.3.3 MF-ADCCA Analysis
In this section, we use MF-ADCCA proposed by our research team to analyze the
asymmetric characteristics of cross-correlation between carbon and returns in the
energy markets. We can quantify the asymmetric cross-correlation when returns in
the energy markets is increasing or decreasing.
We present in Fig. 7.14 the H þ ðqÞ and H ðqÞ plot versus q of each pair of series.
On the basis of this figure, we can detect the asymmetric cross-correlation of each pair
of series in carbon and energy markets. We find that H þ ðqÞ and H ðqÞ for all q values
are not equal; thus, the cross-correlations for each pair of series are asymmetric. In
addition, we find that both H þ ðqÞ and H ðqÞ vary with different q values. Thus, the
asymmetric cross-correlations for four pairs of series are multifractal.
Table 7.3 shows the estimated scaling exponents, Hqð2Þ, Hq þ ð2Þ, Hq ð2Þ, and
DHqð2Þ among four pairs of series for different trends between the carbon and
energy markets, i.e., Brent, coal, electricity, and gas, respectively. Based on
Table 7.3, the overall scaling exponents are significantly >0.5; thus, long-range
persistence exists in the overall cross-correlation. Furthermore, the estimated values
of Hq þ ð2Þ and Hq ð2Þ of the EUA-Brent, EUA-Coal, EUA-Electricity, and
EUA-Gas pairs are all significantly >0.5. Thus, all the cross-correlations between
carbon and energy markets are persistent regardless of the trend of the returns in the
EUA-Brent EUA-Electricity
0.8
0.7
0.7
0.65
0.6 0.6
Hq
Hq
0.55 0.5
0.5
0.4
0.45
0.4
-10 -5 0 5 10 -10 -5 0 5 10
q q
EUA-Coal EUA-Gas
0.8
0.7
0.7
0.65
0.6 0.6
Hq
Hq
0.55 0.5
0.5
0.4
0.45
0.4
-10 -5 0 5 10 -10 -5 0 5 10
q q
upwards downwards
Fig. 7.14 H þ ðqÞ and H ðqÞ plots versus q values varying from −10 to 10 between carbon and
energy markets
Table 7.3 Estimated scaling EUA-Brent EUA-Coal EUA-Electricity EUA-Gas

exponents Hqð2Þ, Hq þ ð2Þ
0.5953 0.5992 0.6030 0.5040
and Hq ð2Þ for the different
trends of the other energy 0.5329 0.5012 0.5585 0.4922
markets 0.6234 0.6627 0.6238 0.5195
−0.0905 −0.1615 −0.0653 −0.0273
energy markets. Moreover, we find that Hq þ ð2Þ is not equal to Hq ð2Þ for each
pair of series. Therefore, the cross-correlations are asymmetric regardless of the
returns in the energy markets.
In addition, on the basis of the last row of Table 7.3, we can determine that
DHqð2Þ are smaller than zero. These findings indicate that the cross-correlations for
EUA-Brent, EUA-Coal, EUA-Electricity, and EUA-Gas are more persistent when
returns in the energy markets are decreasing and not increasing.
7.3.4 Asymmetric Volatility-Constrained Correlation

and Volatility-Transmission Direction
We often observe that two assets are correlated with each other in financial markets.
In particular, the correlation tends to increase when volatility increases. However, in
general, the influence strength from asset one to asset two can be different from the
opposite direction (from asset two to asset one). On the basis of this concept, Ochiai
and Nacher (2014) proposed the volatility-constrained correlation method, which
can quantify which asset is more influential to the other.
The expectation value, standard deviation,
and correlation
coefficient where the
data points are constrained to the C ¼ tti t tf subset are defined as follows
(Ochiai and Nacher 2014).
1 X
E ðRðtÞ; CÞ ¼ RðtÞ ð7:14Þ
#C t2C
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 X
rðRðtÞ; CÞ ¼ ðRðtÞ EðRðtÞ; CÞÞ2 ð7:15Þ
#C t2C
1 X ðR1 ðtÞ E ðR1 ðt; CÞÞÞ ðR2 ðtÞ E ðR2 ðt; CÞÞÞ
C ðR1 ðtÞ; R2 ðtÞ; CÞ ¼ ð7:16Þ
#C t2C rðR1 ðtÞ; CÞ rðR2 ðtÞ; CÞ
where #C denotes the number of elements of C. ti , and tf denote the initial and final
time points for datasets, respectively.
In particular, for C,

C½t1 ; t2 ; a; b ¼ t 2 ti :tf jt1 \t\t2 anda rðR1 ðtÞÞ jR1 ðtÞj b rðR1 ðtÞÞ
ð7:17Þ
where ti t1 t2 tf . Moreover, Ochiai and Nacher (2014) define the different types
of correlations as follows:

F ½a; bðtÞ ¼ C R1 ðtÞ; R2 ðtÞ; C½t;t þ Dt;a;b ð7:18Þ

CðtÞ ¼ C R1 ðtÞ; R2 ðtÞ; C½t;t þ Dt;0;1 ð7:19Þ
where F ½a; bðtÞ represents a correlation between R1 ðtÞ and R2 ðtÞ for each period in
which the dataset ðR1 ðtÞ; R2 ðtÞÞ is constrained, such that jR1 ðtÞj is limited to a
specific range [i.e., a rðR1 ðtÞÞ jR1 ðtÞj b rðR1 ðtÞÞ]. C ðtÞ is the standard cor-
relation coefficient; however, the period is limited from t to t þ Dt. We set Dt as a
half year in the present work. The value of Dt, however, which depends on your
choice. As long as the different types of correlations defined by Ochiai and Nacher
(2014). unchanged, the analysis of the results will not change. We assign R1 ðtÞ as
the base asset. Furthermore, if we exchange R1 ðtÞ for R2 ðtÞ in the Eq. (7.18), then
we obtain different values for F ½a; bðtÞ.
We plot the time evolution of the two types of constrained correlations,F ½0; 1ðtÞ
and F ½1; 1ðtÞ, and the standard correlation coefficient,C ðtÞ, for the base asset is
R1 ðtÞ and R2 ðtÞ in Fig. 7.15. Furthermore, we see that the C ðtÞ varies with time.
Figure 7.15a shows that the C ðtÞ is high from 2008 to 2011, whereas this value is
significantly lower in the second half of 2012. Moreover, when the C ðtÞ is high (for
example, from 2008 to 2011), the constrained correlation F ½1; 1ðtÞ is higher than
the C ðtÞ. This finding implies that the correlation becomes stronger than C ðtÞ when
the EUA (or Brent oil) returns are higher than its standard deviation, and vice versa.
Therefore, higher volatility results in a higher correlation, and lower volatility leads
to a lower correlation.
As previously mentioned, the standard correlation coefficient C ðtÞ for paired
data, R1 ðtÞ and R2 ðtÞ, does not vary when we exchange R1 ðtÞ and R2 ðtÞ. However,
the constrained correlation F ½a; bðtÞ varies. Hence, we focus on this asymmetric
feature of the constrained correlation to reveal which asset is more influential than
the other asset in this subsection.
Figure 7.15a shows the constrained correlation F ½1; 1ðtÞ between EUA and
Brent for both cases of base assets. This figure is constructed using data from
Fig. 7.15a. To see the net effect of volatility to strengthen the correlation, we
subtract the constrained correlation F ½1; 1ðtÞ by standard correlation CðtÞ after
Fisher’s Z transformation defined as Z ð xÞ ¼ 1=2 ln½ð1 þ xÞ=ð1 xÞ. In addition,
Z1 ; ZC denote the Fisher’s Z transformation of F ½1; 1ðtÞ and CðtÞ, respectively.
Subsequently, we can interpret Z1 ZC as the net volatility effect on correlation
(Ochiai and Nacher 2014). In Fig. 7.16, we plot these data to determine the
asymmetric effect between both base assets. We see that the net volatility effect on
correlation Z1 ZC with the base asset Brent oil is generally stronger than that with
the base asset EUA (refer to Table 7.4). Thus, the volatility of the Brent oil return
strengthens the correlation more than the volatility of EUA does. Hence, Brent Oil
is more influential to EUA than the opposite case. This result is consistent with the
conclusion that Brent prices are the main predictor of natural gas prices, which then
affects power prices and ultimately carbon prices (Kanen 2006). Furthermore, the
energy price (i.e., for Brent prices) is the most important factor for the short-term
price of EUA (Alberola et al. 2008).
In addition, we provide the same analysis for the asset pair of EUA and coal,
EUA, and electricity, as well EUA and Gas, respectively. We show the constrained
correlation F ½1; 1ðtÞ for both base assets in Fig. 7.15. Furthermore, Fig. 7.16
shows the net volatility effect on correlation Z1 ZC over the whole period.
Therefore, on average, the net volatility effect on correlation Z1 ZC for the energy
base asset is larger than that for the carbon base asset (see Table 7.4). This finding
implies that these energy base assets are more influential to carbon base asset than
the opposite case. This outcome is also consistent with the result that electricity and
gas prices can be regarded as significant determinants of the future EUA carbon
price in Phase II of the EU ETS (Boersen and Scholtens 2014). Moreover, this
finding is in line with the findings of Alberola et al. (2008, 2009) and
(a)
EUA and Brent (constrained by EUA) EUA and Brent (constrained by Brent)
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
2008 2009 2010 2011 2012 2013 2008 2009 2010 2011 2012 2013
t (year) t (year)
EUA and Coal (constrained by EUA) EUA and Brent (constrained by Coal)
0.6 1
0.4
0.5
0.2
0
0
-0.2 -0.5
2008 2009 2010 2011 2012 2013 2008 2009 2010 2011 2012 2013
t (year) t (year)
F[1,∞ ] F[0,1] C
(b)
EUA and Electricity (constrained by EUA) EUA and Electricity (constrained by Electricity)
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
2008 2009 2010 2011 2012 2013 2008 2009 2010 2011 2012 2013
t (year) t (year)
EUA and Gas (constrained by EUA) EUA and Gas (constrained by Gas)
0.8 0.6
0.6
0.4
0.4
0.2
0.2
0
0
-0.2 -0.2
2008 2009 2010 2011 2012 2013 2008 2009 2010 2011 2012 2013
t (year) t (year)
F[1, ∞] F[0,1] C
Fig. 7.15 Constrained correlation between the carbon and energy markets
Net volatility effect on correlation between EUA and Brent Net volatility effect on correlation between EUA and Electricity
0.6 0.4
EUA base asset EUA base asset
Brent base asset 0.3 Electricity base asset
0.4
0.2
C
C
Z -Z
Z -Z
0.2
1
1
0.1
0
0
-0.2 -0.1
2008 2009 2010 2011 2012 2013 2008 2009 2010 2011 2012 2013
t (year) t (year)
Net volatility effect on correlation between EUA and Coal Net volatility effect on correlation between EUA and Gas
0.6 0.6
EUA base asset EUA base asset
Coal base asset Gas base asset
0.4 0.4
C
C
Z -Z
Z -Z
0.2 0.2
1
0 0
-0.2 -0.2
2008 2009 2010 2011 2012 2013 2008 2009 2010 2011 2012 2013
t (year) t (year)
Fig. 7.16 Net volatility effect on the correlation between carbon and energy markets
Table 7.4 Average of net volatility effect on correlation

Average of net volatility EUA-Brent EUA-Electricity EUA-Coal EUA-Gas
effect on correlation
EUA base asset 0.1170 0.1288 0.0707 0.1180
Energy base asset 0.1178 0.1352 0.0790 0.1425
Chevalier (2009). This conclusion is based on the fact that electricity price is greatly
determined by the cost of the fuel inputs. In addition, these costs are affected by the
CO2 allowance price. The positive coefficients of gas, coal, and oil confirm the
finding that high and low fuel prices contribute to an increase and decrease in
carbon prices, respectively. The negative impact of coal is not as significant as in
most previous studies; however, gas–coal switch variable reveals a significant
indirect impact of coal on carbon prices (Boersen and Scholtens 2014). In general,
hydropower resources are common scarce in the EU, and power generation facil-
ities are limited. The EU Member States, especially in Germany, propose an
extensive use of coal for power generation to meet electricity demand. An increase
of demand in coal results in a steady increase of coal prices. Meanwhile, a large
number of coal consumption leads to increased carbon emissions, which then
increases carbon quota demand and carbon price.
To confirm our empirical results, we further investigate the asymmetric causality
(or transmission direction) between the two markets by employing the time rate of
information flow method proposed by Liang (2014). For the two series, the rate of
information flow from the latter to the former is computed as follows:
C11 C12 C2;d1 C12

2
C1;d1
T2!1 ¼ 2 C C C2
ð7:20Þ
C11 22 11 12
where Cij is the sample covariance between Xi :Xj and Ci;dj are the covariances
between Xi and X_ j . X_ j is the difference approximation based on Euler forward
scheme:
Xj;n þ k Xj;n
X_ j;n ¼ ð7:21Þ
kDt
where k 1; and this parameter should not be large to ensure precision.

Furthermore, T2!1 could be zero or nonzero. If T2!1 ¼ 0, then X2 does not cause
X1 ; otherwise, this value is causal. A positive T2!1 means that X2 functions to result
in a more uncertain X1 , whereas a negative value indicates that X2 tends to stabilize
X1 . T1!2 can be obtained by switching indices 1 and 2. The interpretation of T1!2 is
the same as that of T2!1 .
For comparison, in Fig. 7.17, we plot their absolute values jT2!1 j and jT1!2 j,
which measure the strength of the underlying asymmetric causality. We can con-
clude that the empirical results are similar to the previous empirical analysis. During
EUA-Brent EUA-Electricity
0.04 0.04
T2 1 T2 1
Rate of information flow
0.03 T1 2 0.03 T1 2
0.02 0.02
0.01 0.01
0 0
2008 2009 2010 2011 2012 2008 2009 2010 2011 2012
t (year) t (year)
EUA-Coal EUA-Gas
0.025 0.04
T2 1 T2 1
0.02
T1 2 0.03 T1 2
0.015
0.02
0.01
0.01
0.005
0 0
2008 2009 2010 2011 2012 2008 2009 2010 2011 2012
t (year) t (year)
Fig. 7.17 Plot of the absolute values jT2!1 j and jT1!2 j between carbon and energy markets
Table 7.5 Estimated the rate of information flow for the overall period
EUA-Brent EUA-Electricity EUA-Coal EUA-Gas
T2!1 −0.0060 −0.0119 −0.0047 −0.0037
T1!2 0.0041 −0.0026 −0.0012 −0.0026
most period, jT2!1 j is >jT1!2 j. On the one hand, the causality between the two
markets is asymmetric, i.e., the influence between the two markets is asymmetric.
On the other hand, the energy market base asset is more influential than the carbon
markets (see Table 7.5). According to empirical analysis, we can conclude that the
structure between the two markets is nonlinear. Moreover, for nonlinear systems,
we use the absolute values to analyze the causality (Liang 2014).
7.3.5 Implications and Conclusions
The current study uses asymmetric DCCA cross-correlation coefficient method,

MF-ADCCA, asymmetric volatility-constrained correlation metric, and the infor-
mation flow time rate approach to determine the significance level of the
cross-correlation on different trends (up or down) as well as the asymmetric mul-
tifractal detrended cross-correlations between the carbon and energy markets. In
addition, we investigate the asymmetric volatility-transmission direction between
the two markets.
First, the use of asymmetric DCCA cross-correlation coefficient method quan-
tifies significance level testing of the cross-correlation on different trends (up or
down) between the carbon and energy markets. The empirical results show that
most qDCCA , Dqc , qup , and qdown are significant at 95% confidence interval, which
shows an asymmetric characteristic from the cross-correlation between carbon and
returns in the energy markets. Moreover, a positive cross-correlation may exist
between carbon and energy markets not only for returns but also for volatilities.
Second, our evidence indicates that the asymmetric cross-correlation between
carbon and energy market price returns is persistent and multifractral. Furthermore,
cross-correlations for EUA-Brent, EUA-Coal, EUA-Electricity, and EUA-Gas are
more persistent when returns in the energy markets are decreasing and not
increasing.
Lastly, higher volatility results in a higher correlation, and lower volatility leads
to a lower correlation. In addition, the volatility of the energy market returns base
assets are more influential to carbon market base asset than the opposite case.
To some extent, the impact of the energy market is on the carbon market. In
addition, regulators need to strengthen fluctuation monitoring in the energy market.
In terms of the crude oil market volatility, regulators should be more focused on.
Since the effect of crude oil market on the carbon price is stable, we should consider
the volatility effect of crude oil price on coal prices, natural gas prices, and other
energy prices. In addition, to respond to the volatility risk caused by carbon market,
the relevant regulatory authorities should assess the specific impact of the crude oil
market volatility risk on the carbon market in advance. Moreover, we are required
to construct a feasible plan to deal with the volatility risk to obtain timely appro-
priate risk controls to hedge risk or to improve the stability of the carbon market.
Future work may need further tracking study of the interaction mechanism
between the energy market and carbon trading market, especially their interaction
mechanism changes in post Kyoto era. Furthermore, we need to further explore the
effect of an increase in the amount of renewable energy on carbon market
development.
References
E. Alberola, J. Chevallier, B. Chèze, Price drivers and structural breaks in European carbon price
2005–2007. Energy Policy 36(2), 787–797 (2008)
E. Alberola, J. Chevallier, B. Cheze, Emissions compliances and carbon prices under the EU ETS:
a country specific analysis of industrial sectors. J. Policy Model. 31, 446–462 (2009)
A. Boersen, B. Scholtens, The relationship between European electricity markets and emission
allowance futures prices in phase II of the EU (European Union) emission trading scheme.
Phys. A 74(1), 585–594 (2014)
D.W. Bunn, C. Fezzi, Interaction of European carbon trading and energy prices (Social Science
Electronic Publishing, 2007)
G.X. Cao, J. Cao, L.B. Xu, L.Y. He, Multifractal detrended cross-correlation between the Chinese
domestic and international gold markets based on DCCA and DMCA methods. Modern Phys.
Lett. B 28(11), 243 (2014)
J. Chevallier, Carbon futures and macroeconomic risk factors: a view from the EU ETS. Energy
Econ. 31(4), 614–625 (2009)
F.J. Convery, L. Redmond, Market and price developments in the European Union emission-
strading scheme. Rev. Environ. Econ. Policy 1(1), 88–111 (2007)
S.J. Devlin, R. Gnanadesikan, J.R. Kettenring, Robust estimation and outlier detection with
correlation coefficients. Biometrika 62(3), 531–545 (1975)
R.F. Engle, C.W. Granger, Co-integration and error correction: representation, estimation, and
testing. Econometrica 55(2), 251–276 (1987)
C.W.J. Granger, Econometrica 37, 424–438 (1969)
S. Johansen, Statistical analysis of cointegration vectors. J. Econ. Dyn. Control 12(2), 231–254
(1988)
J.L. Kanen, Carbon trading and pricing (Environmental Finance Publications, London, 2006)
L. Kristoufek, Measuring correlations between non-stationary series with DCCA coefficient. Phys.
A 402, 291–298 (2014)
M. Mansanet-Bataller, A. Pardo, E. Valor, CO2 prices, energy and weather. Energy J. 28(3), 67–86
(2007)
M. Nicolau, Do spot prices move towards futures prices? A study on crude oil market. Acta Univ.
Danub. Oecon. 8(5), 166–176 (2012)
T. Ochiai, J.C. Nacher, Volatility-constrained correlation identifies the directionality of the
influence between Japan’s Nikkei225 and other financial markets. Phys. A 393(1), 364–375
(2014)
B. Podobnik, Z.Q. Jiang, W.X. Zhou, H.E. Stanley, Statistical tests for power-law cross-correlated
processes. Phys. Rev. E: Stat. Nonlin. Soft Matter Phys. 84(6 Pt 2), 066118 (2011)
References 153
nonstationary time series. Phys. Rev. Lett. 100(8), 084102 (2008)
X. San Liang, Unraveling the cause-effect relation between time series, Phys. Rev. E 90(5),
052150 (2014)
Y.D. Wang, C.F. Wu, Prevention of infant and childhood injury through an active pediatrician
participated model, in APHA Meeting and Exposition (2013)
Y.D. Wang, Y. Wei, C.F. Wu, Stock market network’s topological stability: evidence from planar
maximally filtered graph and minimal spanning tree. Int. J. Mod. Phys. B 29(22), 1550161
(2015)
R.R. Wilcox, Robust regression (Chapter 10), in Introduction to Robust Estimation & Hypothesis
Testing, vol. 18, 3rd edn (18) (2012), pp. 471–532
G.F. Zebende, DCCA cross-correlation coefficient: quantifying level of cross-correlation. Phys.
A Stat. Mech. Its Appl. 390(4), 614–618 (2011)
Chapter 8
Simulation—Taking DMCA as an Example
8.1 ARFIMA Process
We employ a periodic two-component fractionally autoregressive integrated mov-

ing average (ARFIMA) process [26, 38, 34–35, 47] to generate two cross-correlated
time series xðtÞ and yðtÞ:
" #
X
þ1 X
þ1
2p
xð t Þ ¼ W ai ðq1 Þxij þ ð1 W Þ ai ðq2 Þyij þ A1 sin i þ ni ð8:1Þ
i¼1 i¼1
T1
" #
X
þ1 X
þ1
2p
yðtÞ ¼ ð1 W Þ ai ðq1 Þxij þ ð1 W Þ ai ðq2 Þyij þ A2 sin i þ fi
i¼1 i¼1
T2
ð8:2Þ
Cði qÞ
ai ðqÞ ¼ q ð8:3Þ
Cð1 qÞCð1 þ jÞ
where qi ði ¼ 1; 2Þ is a parameter, ranging from 0.5 to 1, Ak ðk ¼ 1; 2Þ and

Tk ðk ¼ 1; 2Þ denote sinusoidal amplitude and sinusoidal period, respectively. n and
f represent Gaussian noise, Cð xÞ denotes Gamma function, W 2 ½0:5; 1 quantifies
the coupling strength of power-law cross-correlation between xðtÞ and yðtÞ. If
W ¼ 1 and A1 ¼ A2 ¼ 0, the two correlated time series (Eqs. 8.1 and 8.2) are fully
decoupled and become two separate ARFIMA series, which Hurst Exponent sub-
ject to H ¼ 0:5 þ q [47]. Especially, it has been proven that the cross-correlation
exponent is equal to the average of individual Hurst exponents for two time series
generated by separate ARFIMA process, namely HxyðqÞ ¼ HxxðqÞ þ HyyðqÞ (Zhou
2008; He et al. 2009).

156 8 Simulation—Taking DMCA as an Example
8.2 DMCA Method
8.2.1 Detrended Moving-Average Cross-Correlation

Analysis (DMCA)
In this part, we combine the DCCA and DMA, and then propose a new method,
Detrended Moving-average Cross-correlation Analysis (DMCA), trying to incor-
porate the advantages of both DCCA and DMA.
Suppose that there are two simultaneously recorded time series fxðiÞg and
fyðiÞg; i ¼ 1; 2; . . .; N, where N is the length of each time series. We calculate two
integrated signals:
X
i X
i
XðiÞ ¼ ðxðtÞ xÞ; YðiÞ ¼ ðyðtÞ yÞ; i ¼ 1; 2; . . .; N ð8:4Þ
t¼1 t¼1
For a window of size l, the moving average is given by:
1 X
½ðl1Þð1hÞ
1 X
½ðl1Þð1hÞ
Xl ðiÞ ¼ Xði tÞ and Yl ðiÞ ¼ Yði tÞ ð8:5Þ
l t¼ððl1ÞhÞÞ l t¼ððl1ÞhÞÞ
where h is a position parameter ranging from 0 to 1. In this formulation of moving

average (Eq. 8.5), three special cases are considered, namely h ¼ 0 (backward
moving average, in which the filter is obtained by the past data points), h ¼ 0:5
(centered moving average, in which the filter is obtained by the present data points)
and h ¼ 1 (forward moving average, in which the filter is obtained by the future
data points) [31, 33]. Then the detrended covariance can be defined as:
½Nhðn1Þ
X
1
2
FDMCA ðnÞ ¼ ðXðiÞ X n ðiÞÞðYðiÞ Y n ðiÞÞ: ð8:6Þ
N n þ 1 i¼ðnhðn1ÞÞ
If the power-law cross-correlation exists, the following scaling relationship can

be observed:
FDMCA ðn; iÞ / nHDMCA : ð8:7Þ
The Exponent HDMCA can describe the power-law cross-correlation relationship

between the two related time series. If xðiÞ is identical to yðiÞ, this method
degenerates into DMA.
As a comparison, let us briefly introduce the algorithm of DCCA [23]: First, two
integrated signals xðiÞ and yðiÞ are calculated by Eq. (8.4). Then we divide both
time series into N n overlapping boxes, each containing n þ 1 values. For each
8.2 DMCA Method 157
box that starts at i and ends at n þ i, the local trends are estimated by linear
~i ðkÞ and Y~i ðkÞ. Then, the covariance of residuals can be given by
least-squares fits X
1 X iþn
~i ðkÞÞðYðkÞ Y~i ðkÞÞ:
2
fDMCA ðn; iÞ ¼ ðXðkÞ X ð8:8Þ
n þ 1 t¼i
Then calculate the detrended covariance by summing over all overlapping N n

boxes of size n,
1
2
FDMCA ðn; iÞ ¼ f2 ðn; iÞ: ð8:9Þ
N n DMCA
If power-law cross-correlations do exist, the square root of the detrended

covariance that grows with time window n will satisfy
FDCCA ðn; iÞ / nHDCCA : ð8:10Þ
The power-law correlation can be quantified by the exponent HDCCA .
8.2.2 Comparative Studies
In order to compare the performance of DMCA and DCCA in the detection of

cross-correlation, and to estimate the influence of periodic trend, we generate two
cross-correlated time series xðiÞ and yðiÞ by a periodic two-component fractionally
autoregressive integrated moving average (ARFIMA) process (Podobnik et al.
2007, 2009a, b, 2005a, b, 2008; Hosking 1981; Granger and Joyeux 1980).
Our intention is to estimate the cross-correlation exponents of the generated time
series whose theoretical values are already known, using both DMCA (under dif-
ferent parameters h ¼ 0; 1; 0:5) and DCCA. By doing so, we can compare the
actual performance of these two methods. As the real world data are often of
different sizes, we generate four groups of time series by separate ARFIMA pro-
cesses with same noise. The groups comprise of 2000, 4000, 10,000, and 20,000
data points (lengths); for each length of time series we generated 10 different pairs.
Figure 8.1 shows the results of DMCA versus DCCA for all those different
lengths when q1 ¼ 0:1 and q2 ¼ 0:4, from which one can find that power-law
cross-correlations can be clearly identified by both methods. We also calculated the
averages of cross-correlation exponents (see Table 8.1), and the means of standard
errors (see the results in parentheses) for each group. Compared with theoretical
values of cross-correlation exponents between the ARFIMA series, it is interesting
to note that the cross-correlation exponents are somewhat underestimated by both
methods for most cases, and that the unfavorable deviation tends to increase for the
results estimated by the DMCA if the theoretical exponent is large (e.g. when the
(a) 8 (b) 10
H = 0:7797 § 0:0013
6 8 H = 0:7682 § 0:0019
6
4 H=0:7231 § 0:0077
ln F (n)
ln F (n)
H=0:7282 § 0:0053
H = 0:7802 § 0:0008
4 H=0:7640 § 0:0019
2
2
DMCA, θ =0 DMCA, θ =0
0 DMCA, θ =1 DMCA, θ =1
H=0:7388 § 0:0068
0
DMCA, θ =0.5 DMCA, θ =0.5
H = 0:7458 § 0:0054
DCCA DCCA
-2 -2
2.5 3.0 3.5 4.0 4.5 5.0 5.5 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5
ln n ln n
(c) 10 (d) 12
H=0:7380 § 0:0043
8 10
H=0:7395 § 0:0025
8
6
H=0:7540 § 0:0018
6
ln F (n)
ln F (n)
H=0:7305 § 0:0040
4 H=0:7388 § 0:0040
4 H=0:7367 § 0:0035
2
DMCA, θ =0 2 DMCA, θ =0
0
DMCA, θ =0.5 0
H=0:7654 § 0:0028 H=0:7388 § 0:0033 DMCA, θ =0.5
DCCA DCCA
-2 -2
2 3 4 5 6 7 8 2 3 4 5 6 7 8
ln n ln n
Fig. 8.1 The log-log plot for the square root of the detrended covariance versus time scale n. Each
time series is generated by separate ARFIMA process (namely W ¼ 0) with same Gaussian noise,
whose Hurst Exponent is equal to q þ 0:5 and the cross-correlation exponent between each series
Hxy ¼ ðHxx þ HyyÞ ¼ 2. In order to make clearer contrast among the different curves, some
constants are subtracted from the original results. It is unnecessary to plot the results from all ten
pairs; therefore we choose the results from one pair out of the ten when q1 ¼ 0:1 and q2 ¼ 0:4. For
each panel, the lengths of data points are a L ¼ 2000; b L ¼ 4000; c L ¼ 10; 000; d L ¼ 20; 000,
from which one can find that power-law cross-correlations are identified by both methods
theoretical value is 0.85); meanwhile, compared with results estimated by DCCA,

for most case the exponents estimated by DMCA are slightly smaller. For example,
H DCCA ¼ 0:7471; H DMCA ¼ 0:7232, (for h ¼ 0), 0.7328 (for h ¼ 0:5) and 0.7241
(for h ¼ 1) when q1 ¼ 0:2; q2 ¼ 0:3 and L ¼ 2000.
Then, what are the advantages of this new approach compared with DCCA?
In the real world, influenced by economic, seasonal, or some other factors, actual
time series usually possess trends and are non-stationary; therefore, detrended
approaches are usually applied to analyze those series. Polynomial fitting is thereby
widely applied, but this approach sometimes can not completely efface the influence
Table 8.1 The results of cross-correlation exponents estimated by DMCA and DCCA for time series generated by separate ARFIMA processes with same
noise
Theoretical value q1 ¼ 0:1; q2 ¼ 0:4 q1 ¼ 0:2; q2 ¼ 0:3 q1 ¼ 0:1; q2 ¼ 0:2 q1 ¼ 0:3; q2 ¼ 0:4
8.2 DMCA Method
0.75 0.75 0.65 0.85

L = 2000 DCCA 0.7578(0.0058) 0.7396(0.0067) 0.6566(0.0048) 0.8039(0.0082)
DMCA(l ¼ 0) 0.7519(0.0041) 0.7299(0.0047) 0.6331(0.0057) 0.7768(0.0046)
DMCA(l ¼ 0:5) 0.7368(0.0073) 0.7188(0.0082) 0.6369(0.0063) 0.7794(0.0104)
DMCA(l ¼ 1) 0.7559(0.0038) 0.7260(0.0045) 0.6404(0.0041) 0.7631(0.0064)
L = 4000 DCCA 0.7106(0.0053) 0.7149(0.0052) 0.6373(0.0051) 0.8309(0.0067)
DMCA(l ¼ 0) 0.7206(0.0047) 0.7163(0.0048) 0.6480(0.0051) 0.8205(0.0043)
DMCA(l ¼ 0:5) 0.6936(0.0060) 0.6979(0.0060) 0.6217(0.0059) 0.8148(0.0070)
DMCA(l ¼ 1) 0.7188(0.0043) 0.7147(0.0046) 0.6508(0.0045) 0.8196(0.0043)
L = 10,000 DCCA 0.7396(0.0037) 0.7458(0.0041) 0.6312(0.0047) 0.8447(0.0038)
DMCA(l ¼ 0) 0.7190(0.0052) 0.7268(0.0047) 0.6222(0.0049) 0.7981(0.0061)
DMCA(l ¼ 0:5) 0.7279(0.0041) 0.7339(0.0044) 0.6191(0.0054) 0.8324(0.0043)
DMCA(l ¼ 1) 0.7191(0.0055) 0.7283(0.0041) 0.6198(0.0057) 0.7976(0.0064)
L = 20,000 DCCA 0.7387(0.0035) 0.7471(0.0038) 0.6458(0.0042) 0.8419(0.0037)
DMCA(l ¼ 0) 0.7264(0.0050) 0.7232(0.0052) 0.6222(0.0067) 0.8045(0.0057)
DMCA(l ¼ 0:5) 0.7294(0.0040) 0.7382(0.0042) 0.6369(0.0051) 0.8320(0.0040)
DMCA(l ¼ 1) 0.7292(0.0049) 0.7241(0.0052) 0.6215(0.0070) 0.8067(0.0055)
Note For each length of time series, we generate 10 pairs. The results show the average of 10 cross-correlation exponents of those generated pairs; in
parentheses there are the means of standard errors
159
of trends, which may lead to crossovers (Hu et al. 2001). In our DMCA method,
moving average is used to estimate trends. If the time series are long- or short-range
correlated, the moving average for large (or small) window size n may contain
long-term (or short-term) trend.
To further discuss this problem, we generated the time series by a periodic
two-component ARFIMA process (see Eq. 8.1). We chose W ¼ 0; A1 ¼ A2 ¼
0:3; q1 ¼ q2 ¼ 0:4 and T1 ¼ T2 ¼ 1000 or 500. As a comparison, we also gen-
erated a group of time series with no trend, namely A1 ¼ A2 ¼ 0. Each group
consists of 2000, 4000, 10,000, or 20,000 data points; and for each length of time
series we generated 10 pairs. The weight W controls the strength of power-law
cross-correlations between two correlated time series; that is, when W 6¼ 0, each
variable depends not only on its own past, but also on the historical values of the
other variables. When W ¼ 0:5, each time series is equally dependent on the past of
the other; and the two time series are almost similar. The cross-correlation exponent
approximates to the Hurst exponent, namely, Hxy H ¼ q þ 0:5 ¼ 0:9.
In Fig. 8.2, in order to improve the readability and make clearer contrast among
the different curves, we subtract some constants, i.e., 3 is subtracted from the
original results when L ¼ 10;000, 6 when L ¼ 4000, and 9 when L ¼ 2000. From
this figure, one can find that for both DMCA and DCCA methods, there are
crossovers. The moving average approach can not completely eliminate the effect of
trend.
From Table 8.2, interestingly, we find that the results of DCCA and DMCA (for
h ¼ 0:5; namely, centered moving average) are similar, while the exponents esti-
mated by DMCA for h ¼ 0 and h ¼ 1 (i.e., backward and forward moving aver-
ages) are similar, too. Compared with DMCA (for backward and forward moving
average cases), DCCA may have a disadvantage in dealing with the unfavorable
influence of trend. For example, when T ¼ 1000 and L ¼ 10; 000,
H DCCA ¼ 1:4569; H DMCA ¼ 0:8304, (for h ¼ 0) and 0.8396 (for h ¼ 1), while the
theoretical value is only 0.9 (see Table 2). Because of the existence of crossovers,
the cross-correlation exponents may be significantly overestimated by DCCA and
DMCA (for h ¼ 0:5), but slightly underestimated by DMCA (for h ¼ 0 and h ¼ 1).
Meanwhile, please note that if the length of time series is short, DMCA (for h ¼ 0
and h ¼ 1) outperform DCCA and DMCA (for h ¼ 0:5). But if the length is
relatively longer, the results tend to be underestimated by DMCA (for h ¼ 0 and
h ¼ 1), especially when T ¼ 500 and L ¼ 20; 000, H DMCA ¼ 0:6639 (for h ¼ 0)
and 0.6666 (for h ¼ 1). At this time, H DCCA ¼ 1:1861 and H DMCA ¼ 1:1574 (for
h ¼ 0:5) seem to be closer to the theoretical value 0.9. In short, if the length of time
series is long enough, DCCA and DMCA (for h ¼ 0:5) may perform better; but if
the length is short, DMCA (for l¼0 and h ¼ 1) may be a better choice.
Furthermore, since the empirical time series are usually short in length, DMCA (for
h ¼ 0 and h ¼ 1) may be more accurate, and thereby a better solution in practice,
because for short time series, its estimations are closer to theoretical values.
Podobnik et al. propose a global detrending approach, which can effectively
eliminate the influence of trend. But one might criticize that they have known the
8.2 DMCA Method 161
Fig. 8.2 The log-log plot for the square root of the detrended covariance versus time scale n It is
unnecessary to plot the results from all ten pairs; therefore we choose the results from one pair out
of the ten when the parameters A ¼ A1 ¼ A2 ¼ 0:3; q1 ¼ q2 ¼ 0:4; W ¼ 0:5; with the
sinusoidal periods T ¼ 1000 and T ¼ 500, and for the DMCA method, we only show the results
for the case h ¼ 0 (the others, namely h ¼ 0:5 (overestimated) and h ¼ 1 (underestimated), are
similar). In order to make clearer contrast among the different curves, some constants are
subtracted from the results, namely, 3 is subtracted when L ¼ 10; 000, 6 is subtracted when
L ¼ 4000 and 9 when L ¼ 20; 000. For each panel, from top to bottom, the lengths of each series
are 2000, 4000, 10,000 and 20,000 respectively. One may find the crossovers, which cause the
exponents underestimated by DMCA (for l ¼ 0) and overestimated by DCCA
format of the trend in advance. In real world, there are seasonal periods, economic
periods, non-periodic cycles, or even complex trends, which we do not know the
format or function before we analyzed the time series. Although the DMCA we
propose can not complete eliminate all influence of trend, it can reduce the effect,
which can identify the power-law cross-correlation given the presence of unfa-
vorable trends in the real world.
Therefore, compared with the DCCA, this new approach may possess the fol-
lowing advantages: First, it can reduce the amount of calculation. Compared with
the polynomial fitting applied in the DCCA, the latter only needs simple moving
average filtering. In addition, unlike the DCCA, there is no need to divide the boxes
for the DMCA; while as for the DCCA, to make sure the accuracy, overlapping
boxes are used, which may increase a huge amount of additional calculations.
Table 8.2 The results of DMCA and DCCA methods for the parameters q ¼ 0:4 and W ¼ 0:5
Theoretical value T = 1000 T = 500 No trend
0.9 0.9 0.9
L = 2000 DCCA 1.2707(0.0188) 1.3802(0.0262) 0.9808(0.0154)
DMCA(l ¼ 0) 0.9892(0.0037) 0.9349(0.0102) 0.8225(0.0052)
DMCA(l ¼ 0:5) 1.1796(0.0296) 1.3135(0.0307) 0.9447(0.0175)
DMCA(l ¼ 1) 0.9830(0.0041) 0.9312(0.0097) 0.8227(0.0052)
L = 4000 DCCA 1.3771(0.0297) 1.4495(0.0221) 0.9985(0.0101)
DMCA(l ¼ 0) 0.9557(0.0094) 0.8348(0.0278) 0.8754(0.0045)
DMCA(l ¼ 0:5) 1.3414(0.0296) 1.4250(0.0230) 0.9780(0.0112)
DMCA(l ¼ 1) 0.9525(0.0091) 0.8419(0.0256) 0.8778(0.0044)
L = 10,000 DCCA 1.4569(0.0278) 1.3503(0.0372) 0.9473(0.0095)
DMCA(l ¼ 0) 0.8304(0.0345) 0.6816(0.0404) 0.8609(0.0049)
DMCA(l ¼ 0:5) 1.4246(0.0332) 1.3296(0.0410) 0.9283(0.0099)
DMCA(l ¼ 1) 0.8396(0.0324) 0.6857(0.0401) 0.8616(0.0047)
L = 20,000 DCCA 1.3767(0.0343) 1.1861(0.0575) 0.9491(0.0088)
DMCA(l ¼ 0) 0.7292(0.0423) 0.6639(0.0322) 0.8522(0.0058)
DMCA(l ¼ 0:5) 1.3616(0.0379) 1.1574(0.0637) 0.9391(0.0089)
DMCA(l ¼ 1) 0.7319(0.0420) 0.6666(0.0320) 0.8524(0.0057)
Note For each length of time series, we generate 10 pairs. The results show the average of 10
cross-correlation exponents of those generated pairs; in parentheses there are the means of standard
errors
Second, continuous adjusting filter is applied. For the DCCA, the polynomial
trend between each box is discontinuous; but all of the filters are continuous in our
method, for the moving average filter adjusts the fitting curve dynamically, which
may increase the accuracy.
Third, the influence of trend is reduced. Although DMCA (for backward and
forward moving average cases) can not completely eliminate the influence of the
trend, and for a long period time series, the results may be underestimated; but for a
short period, DMCA for backward and forward moving average cases can perform
better than DCCA.
Four, it is a practical choice for real world data with short lengths. As we
discussed earlier, DMCA using for backward and forward moving average filters
outperforms DCCA in more accurate estimation when the analyzed times series are
short in length.
8.2 DMCA Method 163
8.2.3 Empirical Analysis on the International Crude Oil

Spot Markets
As an example, we apply DMCA to investigate the cross-correlation between some

commodity markets. We proceed an empirical study on international crude oil spot
markets, which according to our previous empirical findings, are found to be fractal
and multifractal (He et al. 2007, 2010), and choose the daily price of WTI Spot
Price FOB (Dollars per Barrel) and Europe Brent Spot Price FOB (Dollars per
Barrel), from May 20th, 1987 to June 29th, 2010, which consist of 5775 simulta-
neously recorded data points (data source: US Energy Information Administration,
http://tonto.eia.doe.gov/dnav/pet/pet_pri_spt_s1_d.htm).
Let pðtÞ to be the price of crude oil. Then the daily price return rðtÞ is calculated
as its logarithmic difference, rðtÞ ¼ lnðpðt þ DtÞ=pðtÞÞ, where rDt ¼ 1. To get
better understanding of the data, we perform summary statistics of the logarithmic
returns (see Table 8.3), from which we can see a large skewness and kurtosis. To
better describe the time series, we also plot the integrated profiles IðtÞ ¼
Pt
i¼1 ðjrðiÞj hjrðiÞjiÞ in Fig. 8.3, from which one can find that they are highly
correlated.
In Table 8.3 the asymmetry and large kurtosis can be clearly seen in crude oil
returns, which imply existence of power-law tails in probability density function
(PDF). Podobnik et al. (2005a, b) analyze the asymmetry in presence of long-range
correlations by a new process. We investigate the tails in PDF, following the
method suggested by Podobnik et al. (2009a, b), which is also proposed by Ren and
Zhou (2010a, b) independently. Several studies investigate the returns intervals s
between consecutive price fluctuations above a volatility threshold q, especially in
Chinese stock market (Ren et al. 2009a, b). The PDF of returns intervals Pq ðsÞ
scales with the mean of s as in (Yamasaki et al. 2005):
1 s
Pq ðsÞ ¼ f ð Þ ð8:11Þ
s s
where f ðxÞ is a stretched exponential function. Then Podobnik et al. propose a new
estimate for the power-law exponent by
sq / qa ð8:12Þ
By means of Eqs. (8.11) and (8.12), we obtain the relationships between average
return interval s and threshold q (in units of standard deviation) (see Fig. 8.4). From
Table 8.3 Mean, standard deviation, skewness and kurtosis of crude oil returns
Mean Std. dev. Skewness Kurtosis
WTI 0.0177 0.0186 4.0157 46.274
Brent 0.0168 0.0172 3.9382 43.757
WTI
2 Brent
0
I (t)
-2
-4
-6
-8
0 1000 2000 3000 4000 5000 6000 7000
time
Fig. 8.3 The integrated profiles of absolute logarithmic returns for WTI and Brent. It is clear that
the two markets are highly correlated
this log-log plot, the linear relationships can be clearly identified, which implies that
there exist power law relationships between the intervals and thresholds for both
WTI and Brent prices; and the estimated exponents, that is, ~ a ¼ 2:7655 0:0775
(WTI) and ~a ¼ 2:6864 0:0419 (Brent), are slightly less than the inverse cubic law
reported in (Gopikrishnan et al. 1998).
To provide a practical example, we further applied DMCA and DCCA methods
to quantify the cross-correlation relationships between WTI and Brent crude oil spot
markets respectively (see Fig. 8.5; Table 8.4). From Fig. 8.5 and Table 8.4, the
power-law cross-correlation relationships can be found for both markets. The
results measured by DMCA and DCCA (see Table 8.4) tell us that the exponents
are around 0.5 for original returns, but approximate to 1 for the absolute returns. It
suggests that price fluctuations can propagate and then transmit to the other; and
that one large volatility of price change in one market is likely to cause another
large volatility in the correlated market.
8.2.4 Conclusions
We proposed a new method Detrended Moving-average Cross-correlation Analysis

(DMCA) by combining DCCA and DMA. Many comparisons were made between
this new method and DCCA. To make a practical example, we also applied DMCA
to investigate empirically the cross-correlation between real world data, i.e. WTI
8.2 DMCA Method 165
7 6
6 WTI 5 Brent
5 4
ln τ
ln τ
4 3
® e = 2:6864 § 0:0419
3 ® e = 2:7655 § 0:0775
2
2 1
.6 .8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 .6 .8 1.0 1.2 1.4 1.6 1.8 2.0 2.2
ln q ln q
Fig. 8.4 The log-log plot for average return interval s versus threshold q (in units of standard
deviation). One can find distinct power law relationships, whose exponents are 2:7655 0:0775
(WTI) and 2:6864 0:0419 (Brent), which are slightly less than the inverse cubic law
(a) 4 (b)
8
H = 0:4699 § 0:0151
6 H = 0:9850 § 0:0031
2
4
0 2
H = 0:5594 § 0:0083
ln F (n)
ln F (n)
H = 0:4583 § 0:0182 H = 0:9756 § 0:0248

0
H = 0:9810 § 0:0040
-2
-2
-4
-4 DMCA, θ =1 DMCA, θ =1
H = 0:5648 § 0:0071 DMCA, θ =0.5 -6 H = 0:9933 § 0:0227 DMCA, θ =0.5
DCCA DCCA
-6 -8
2 3 4 5 6 7 2 3 4 5 6 7
ln n ln n
Fig. 8.5 The log-log plot for the square root of the detrended covariance versus time scale n. The
results are measured by DMCA and DCCA, using a the original returns and b absolute returns
series. In order to make clearer contrast among the different curves, some constants are subtracted
from the original results. One can find that there is power-law cross-correlation between WTI and
Brent crude oil spot markets, where the exponents are around 0.5 (for original returns) and
approximate to 1 (for absolute returns or volatility)
Table 8.4 The Original returns Absolute returns

cross-correlation exponents
between WTI and Brent crude DCCA 0.5648 ± 0.0071 0.9933 ± 0.0227
oil spot markets DMCA(l ¼ 0) 0.4699 ± 0.0151 0.9850 ± 0.0031
DMCA(l ¼ 0:5) 0.5594 ± 0.0083 0.9756 ± 0.0248
DMCA(l ¼ 1) 0.4583 ± 0.0182 0.9810 ± 0.0040
and Brent crude oil spot markets in this paper. Our conclusions can be summarized
as follows.
First, we proposed a new approach—DMCA, which can efficiently quantify the
power-law correlation between two non-stationary time series.
Second, for long time series with trend, DCCA and DMCA (for centered moving
average case) outperform DMCA (for background and forward moving average
case); but if time series are short, the latter method seems to be a better choice.
Since the real world data are often short in length, DMCA (for background and
forward moving average cases) may be a more practical choice compared with
DCCA.
Third, this new approach can outperform the present method by significantly
reducing the amount of calculations and the effect of trend, although it can not
completely eliminate the effect.
Fourth, the empirical study also shows that by means of this method, some
commodity markets (in this paper, WTI and Brent crude oil spot markets) are found
to be power-law cross-correlated. We thereby present an example that our method
has potential application to real world problems.
In all, this method provides another practical choice to identify the
cross-correlation between two non-stationary time series, especially of short
periods.
References
P. Gopikrishnan, M. Meyer, L.A.N. Amaral, H.E. Stanley, Inverse cubic law for the distribution of
stock price variations. Eur. Phys. J. B—Condens. Matter Complex Syst. 3(2), 139–140 (1998)
C.W.J. Granger, R. Joyeux, An introduction to long-memory time series models and fractional
differencing. J. Time Ser. Anal. 1(1), 15–29 (1980)
L.Y. He, Y. Fan, Y.M. Wei, The empirical analysis for fractal features and long-run memory
mechanism in petroleum pricing systems Int. J. Glob. Energy Issues 27(4), 492–502 (2007)
L.Y. He, Y. Fan, Y.M. Wei, Impact of speculator’s expectations of returns and time scales of
investment on crude oil price behaviors. Energy Econ. 31(1), 77–84 (2009)
J.R.M. Hosking, Fractional differencing. Biometrika 68(1), 165–176 (1981)
K. Hu, P.C. Ivanov, Z. Chen, P. Carpena, H.E. Stanley, Effect of trends on detrended fluctuation
analysis. Phys. Rev. E 64(1), 011114 (2001)
Multifractal detrended fluctuation analysis of nonstationary time series. Phys. A 316(1–4),
87–114 (2002)
References 167
nonstationary time series. Phys. Rev. Lett. 100(8), 084102 (2008)
B. Podobnik, P.C. Ivanov, K. Biljakovic, D. Horvatic, H.E. Stanley, I. Grosse, Fractionally
integrated process with power-law correlations in variables and magnitudes. Phys. Rev. E 72(2),
026121 (2005a)
B. Podobnik, P.C. Ivanov, V. Jazbinsek, Z. Trontelj, H.E. Stanley, I. Grosse, Power-law correlated
processes with asymmetric distributions. Phys. Rev. E 71(2), 025104 (2005b)
B. Podobnik, D.F. Fu, H.E. Stanley, P.C. Ivanov, Power-law autocorrelated stochastic processes
with long-range cross-correlations. Eur. Phys. J. B 56(1), 47–52 (2007)
B. Podobnik, D. Horvatic, A.L. Ng, H.E. Stanley, P.C. Ivanov, Modeling long-range
cross-correlations in two-component ARFIMA and FIARCH processes. Phys. A 387(15),
3954–3959 (2008)
B. Podobnik, D. Horvatić, A.M. Petersen, H.E. Stanley, Cross-correlations between volume
change and price change. Proc. Natl. Acad. Sci. U.S.A. 106(52), 22079–22084 (2009a)
B. Podobnik, I. Grosse, D. Horvati, S. Ilic, P.C. Ivanov, H.E. Stanley, Quantifying cross-correlations
using local and global detrending approaches. Eur. Phys. J. B 71(2), 243–250 (2009b)
F. Ren, W.X. Zhou, Recurrence interval analysis of high-frequency financial returns and its
application to risk estimation. New J. Phys. 12(7), 075030 (2010a)
F. Ren, W.X. Zhou, Recurrence interval analysis of trading volumes. Phys. Rev. E 81(6), 066107
(2010b)
F. Ren, G.F. Gu, W.X. Zhou, Scaling and memory in the return intervals of realized volatility.
Phys. A 388(22), 4787–4796 (2009a)
F. Ren, L. Guo, W.X. Zhou, Statistical properties of volatility return intervals of Chinese stocks.
Phys. A 388(6), 881–890 (2009b)
K. Yamasaki, L. Muchnik, S. Havlin, A. Bunde, H.E. Stanley, Scaling and memory in volatility
return intervals in financial markets. Proc. Natl. Acad. Sci. U.S.A. 102, 9424–9428 (2005)
Rev. E 77(6), 066211 (2008)
Chapter 9
Multifractal Detrend Method
with Different Filtering
To improve this new method, we need to clarify whether the new method can
improve the efficiency, such as in some aspects compared to the original method
has outstanding performance, here we use the ARFIMA method for simulation test.
9.1 Nonlinear Structure Analysis of Carbon and Energy

Markets: MFDCCA-MODWT
Global warming and other climate change problems caused by CO2 and other
greenhouse gases from burning fossil fuels and other energy sources have attracted
increasing attention (Chen and He 2014). According to previous literature (Convery
and Redmond 2007; Bunn and Fezzi 2007; Mansanet-Bataller 2007), energy prices
are the most important drivers of carbon prices because power plants can switch
between their fuel inputs, such as oil, coal, and natural gas. Such switch leads to the
presence of an intrinsic conduction mechanism between carbon and energy markets.
Thus, the price change between carbon trading and energy markets is closely
related. As the main driver of natural gas prices, Brent prices affect power prices,
which ultimately influence carbon prices according to Kanen (2006). Afterwards,
Chevallier (2011) developed a carbon pricing model to further confirm Brent price
as the leader in price formation among energy markets. By establishing empirical
multiple regression models, Alberola et al. (2008) determined that energy prices
were one of the main significant factors that influenced EU Emissions Trading
Scheme (EU ETS) carbon price changes during 2005–2007. Koch (2014) adopted a
multivariate GARCH model to explore the process through which conditional
correlation between carbon, energy, and financial prices in the EU ETS varies over
time. Empirical results show that correlations depend on market uncertainty con-
ditions. However, these studies failed to reflect the dynamics or time-varying
correlations between carbon and energy markets. Fortunately, the Baba–Engle–

170 9 Multifractal Detrend Method with Different Filtering
Kraft–Kroner, constant conditional correlation, and dynamic conditional correlation

MGARCH (DCC-MGARCH) models can be comprehensively used to reflect the
dynamics of the correlations between the oil, gas, and CO2 variables over time
according to Chevallier (2012).
In addition, a number of studies focused on the relationship between electricity
and carbon prices. Smale et al. (2006) concluded that EU ETS can significantly
induce price hikes because emission quotas are direct production cost factors
among power producers. By using a structural co-integrated VAR model, Fezzi and
Bunn (2009) empirically showed the joint influence of carbon and gas prices on the
equilibrium price of electricity. An autoregressive distributed lag model is adopted
to confirm the hypothesis that rising prices of emission allowances have a stronger
effect on wholesale electricity prices compared with falling prices (asymmetric cost
pass-through) for the German market (Zachmann, and Von Hirschhausen 2008).
Daskalakis and Markellos (2009) speculated that a positive relationship exists
between emission allowance spot returns and electricity. Their empirical results
indicate that the allocation of free allowances and their unrestricted trading enables
electricity producers to achieve windfall profits in the derivatives market at the
expense of other market participants. From these studies, we can determine the
close relationship between carbon and electricity markets.
When analyzing the correlation between carbon and energy markets, most
scholars adopt econometric models such as GARCH and DCC-MGARCH. These
models generally consider the limited range of the data interval without considering
its long-term correlation with the two markets. This paper uses maximum overlap
wavelet transform (MODWT), which has been widely used in time series analysis
to investigate the cross-correlations between the two markets that ranked on dif-
ferent time scales (Nason and Von Sachs 1999). Based on an autoregressive model,
Soltani et al. (2000) and Renaud et al. (2003) discussed wavelet analysis for the
time series prediction. From then on, more researchers developed a model for
analyzing and forecasting non-stationary time series (Lineesh and John 2010;
Suhartono and Subanar 2009; Minu et al. 2010). This phenomenon can be attrib-
uted to the multi-resolution analysis characteristic of the wavelet transformation,
which can be analyzed from two different angles: time-domain and
frequency-domain analyses. A set of band-pass filters based on wavelet transform
that filter data can properly select scale and translation coefficients. Afterwards, the
length of time series data based on economic cycle to different frequency bands are
divided to effectively eliminate the influence of irregular and random factors, as
well as long-term trends.
The DMCA algorithm was extended to multifractal detrending cross-correlation
moving average analysis (MF-X-DMA) (Jiang and Zhou 2011). The MF-X-DMA
method which includes backward MF-X-DMA, centered MF-X-DMA, and forward
MF-X-DMA and can effectively eliminate trend by means of a continuous moving
average, despite the fact that polynomial adjustment cannot fully eliminate the
impact of trends in the time series. Moreover, the three algorithms should be all
used and compared to make a better trade-off. As a new method for detection the
long-range cross-correlations and multifractality, based on Barabasi and Vicsek’s
9.1 Nonlinear Structure Analysis of Carbon and Energy … 171
height-height correlation analysis, the MF-HXA method was proposed by

Kristoufek (2012). However, MF-HXA is based on qth order covariance, where
q satisfies the condition of q [ 0.
Although MFDCCA approach can effectively describe the multifractal and
long-term memory characteristics of non-stationary time-series, it also presents a
number of limitations. First, MFDCCA requires polynomial selection in detrended
processing. However, polynomial selection exhibits a variable mode that ranges
from 1 to k orders. Moreover, segmentation of the entire partition sequence is
intermittent, which leads to intermittent polynomial fitting adjacent segmentation on
the interval. Thus, new errors of the pseudo wave are produced and wave function
deviation is caused, which results in distortion of scaling exponents. However,
MODWT method can decompose time series into different detail coefficients that
capture higher frequency oscillations and represent increasingly finer-scale devia-
tions from the smooth trend, whereas smooth coefficients capture trends. This
transformation can effectively avoid the defect. Thus, we propose the MFDCCA
based on MODWT to quantitatively measure the cross-correlations between carbon
and energy markets.
9.1.1 MODWT Methodology
MODWT refers to the maximal overlap discrete wavelet transform (Percival and
Mofjeld 1997). This transformation is deemed as an improved method over the
discrete wavelet transform (DWT) and others ordinary DWTs. Furthermore, the
advantages discussed in Sect. 9.1 includes its smoothness and capacity to handle a
non-dyadic (size refers to a size that is a multiple of 2) sample size. In recent years,
increasing studies have chosen this method to conduct economic research
(Khalfaoui et al. 2015; He and Xie 2015).
Based on MODWT, a multi-scale decomposition of a time series xðtÞ can be
achieved by a sequence of projections onto the father and mother wavelets
X X X X
xð t Þ ¼ SJ;i /J;i ðtÞ þ DJ;i uJ;i ðtÞ þ DJ1;i uJ1;i ðtÞ þ þ D1;i u1;i ðtÞ;
i i i i
ð9:1Þ
where J and i are integers that represent the maximum level of scale s and the
number of coefficients in the specified component, respectively. DJ;i ; . . .D1;i denote
the detail or wavelet coefficients that capture higher frequency oscillations and
represent increasingly finer-scale deviations from the smooth trend, whereas SJ;i
represents the smooth coefficients that capture the trend. These coefficients are
approximated by the integrals according to Bruce and Gao (1996).
Z
SJ;i /J;i ðtÞxðtÞdt; ð9:2Þ
Z
DJ;i uJ; iðtÞxðtÞdt ðJ ¼ 1; 2; . . .J Þ: ð9:3Þ
Generally speaking, Eq. (9.1) can be rewritten as
xðtÞ ¼ SJ;i þ DJ;i þ DJ1;i þ D1;i ð9:4Þ
where
X
SJ;i ¼ SJ;i /J;i ðtÞ ð9:5Þ
i
X
DJ;i ¼ SJ;i /J;i ðtÞ ði ¼ 1; 2; . . .J Þ ð9:6Þ
i
Equations (9.4), (9.5) and (9.6) show the original time series xðtÞ decomposed
into a set of scale components with an increasing order of fineness. J represents the
maximum level of scales used to decompose the data. The size of J or how detailed
one wants to analyze data depends on the research question. In this paper, we chose
the LA(8) wavelet filter, which allows the most accurate alignment in time between
wavelet coefficients and the original time series, and is therefore one of the most
favored approaches in the literature.
We can obtain a set of wavelet coefficients as defined in Eqs. (9.2) and (9.3) by
applying MODWT with the LA(8) filter to a stochastic time series. Using these
wavelet and smooth coefficients, we can achieve numerous statistics for a time
series. Afterwards, the wavelet correlation is defined as follows:
covXY ðkJ Þ
qXY ðkJ Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð9:7Þ
tX ðkJ ÞtY ðkJ Þ
where kJ ¼ 2J1 , and jqXY ðkJ Þj 1 as the usual correlation coefficients between
two random variables. To obtain the details on covXY ðkJ Þ and tðkJ Þ refer to Liu
et al. (2016), Tsai (2013) for review.
9.1.2 Comparative Analysis on the Performance

of MFDCCA-MODWT and MF-X-DMA
As mentioned before, the MF-HXA method was proposed by Kristoufek. However,

MF-HXA is based on qth order covariance, where q satisfies the condition of q [ 0
(Liu et al. 2016). Furthermore, we apply the MFDCCA-MODWT method to
investigate the cross-correlation between carbon and energy markets for q varies
from negative values to positive. Hence, we will not to compare this method with
MFDCCA-MODWT and MF-X-DMA again.
To compare the performance of MFDCCA-MODWT and MF-X-DMA in the
detection of cross-correlation between two non-stationary time series, and further to
investigate the influence of periodic trend, we employ a periodic two-component
fractionally autoregressive integrated moving average (ARFIMA) process to gen-
erate two cross-correlated time series.
We chose W = 0.5, q1 ¼ q2 ¼ 0:4, T1 ¼ T2 ¼ 500 or 1000, A1 ¼ A2 ¼ 0:2 in
this paper. And for comparison, we also generated a group of time series with no
trend, namely A1 ¼ A2 ¼ 0. Each group consists of 2000 or 10,000 data points.
Figures 9.1, 9.2, 9.3 and 9.4 show the log–log plots of log FxyðqÞ versus logðsÞ
for the two time series generated by separate ARFIMA process for q = −10, −5, 0,
5 and 10 with q1 ¼ q2 ¼ 0:4. As shown in Figs. 9.1, 9.2, 9.3 and 9.4 we can
observe that there are exist crossovers for MF-X-DMA method, especially for
MF-X-DMA method with h ¼ 0:5. The MF-X-DMA method which includes
backward MF-X-DMA(h ¼ 0), centered MF-X-DMA(h ¼ 0:5), and forward
MF-X-DMA(h ¼ 1) and can effectively eliminate trend by means of a continuous
moving average, despite the fact that polynomial adjustment cannot fully eliminate
the impact of trends in the time series. However, nice power-law relations are
observed for MFDCCA-MODWT method. Moreover, we find that the exponents
estimated by MF-X-DMA for h ¼ 0 and h ¼ 1 are similar, which is consistent with
the results of Cao et al. (2012).
As shown in Table 9.1, on the one hand, if the length of time series is short,
MF-X-DMA for h ¼ 0 and h ¼ 1 slightly better than MFDCCA-MODWT and
outperform MF-X-DMA for h ¼ 0:5. For example, when T = 1000 and L = 2000,
MF-X-DMA(θ=0) MF-X-DMA(θ=0.5)
8 6
q=-10
q=-5
6 4
q=0
logFxy(q)
logFxy(q)
q=5
4 2 q=10
2 0
0 -2
2 3 4 5 6 7 2 3 4 5 6 7
log(s) log(s)
MF-X-DMA(θ=1) MFDCCA-MODWT
6 6
5
4
4
logFxy(q)
logFxy(q)
2
3
0
2
1 -2
0 -4
2 3 4 5 6 7 1 2 3 4 5 6 7
log(s) log(s)
Fig. 9.1 The linear relationship between log FxyqðsÞ and logðsÞ (L = 2000, T = 500)
6 5
5 4 q=-10
q=-5
4 3 q=0
logFxy(q)
logFxy(q)
q=5
3 2
q=10
2 1
1 0
0 -1
2 3 4 5 6 7 2 3 4 5 6 7
log(s) log(s)
6 4
5 2
4
logFxy(q)
logFxy(q)
0
3
-2
2
1 -4
0 -6
2 3 4 5 6 7 1 2 3 4 5 6
log(s) log(s)
Fig. 9.2 The linear relationship between log FxyqðsÞ and logðsÞ (L = 10,000, T = 500)
6 5
5 4 q=-10
q=-5
4 3 q=0
logFxy(q)
logFxy(q)
q=5
3 2
q=10
2 1
1 0
0 -1
2 3 4 5 6 7 2 3 4 5 6 7
log(s) log(s)
6 4
5 2
4
logFxy(q)
logFxy(q)
0
3
-2
2
1 -4
0 -6
2 3 4 5 6 7 1 2 3 4 5 6
log(s) log(s)
Fig. 9.3 The linear relationship between log FxyqðsÞ and logðsÞ (L = 2000, T = 1000)
HMFXDMA ¼ 0:9618 (h ¼ 0) and 0.9382 (h ¼ 1). However, HMFXDMA ¼

1:2088 (h ¼ 0:5), the cross-correlation exponents may be significantly overesti-
mated. Meanwhile, HMFDCCAMODWT ¼ 0:8552, suggesting that the
cross-correlation exponents may be slightly underestimated. But if the length is
relatively longer, the results tend to be underestimated by MF-X-DMA for h ¼ 0
and h ¼ 1, for example, when T = 500 and L = 10,000, HMFXDMA ¼ 0:6981
(h ¼ 0) and 0.6956 (h ¼ 1). However, HMFXDMA ¼ 1:2491 (h ¼ 0:5), the
6 5
q=-10
5 4
q=-5
4 3 q=0
logFxy(q)
logFxy(q)
q=5
3 2 q=10
2 1
1 0
0 -1
2 3 4 5 6 7 2 3 4 5 6 7
log(s) log(s)
6 4
5
2
4
logFxy(q)
logFxy(q)
3 0
2
-2
1
0 -4
2 3 4 5 6 7 1 2 3 4 5 6
log(s) log(s)
Fig. 9.4 The linear relationship between log FxyqðsÞ and logðsÞ (L = 10,000, T = 1000)
Table 9.1 The results of MFDCCA-MODWT and MF-X-DMA methods for the parameters q ¼
0:4 and W = 0.5
No trend T = 500 T = 1000
Theoretical value 0.9 0.9 0.9
L = 2000 MFDCCA-MODWT 0.8171(0.0221) 0.9436(0.0127) 0.8552(0.0015)
MF-X-DMA(h ¼ 0) 0.7588(0.0207) 0.7641(0.0045) 0.9618(0.0010)
MF-X-DMA 0.7431(0.0446) 1.2705(0.0569) 1.2088(0.0470)
(h ¼ 0:5)
MF-X-DMA(h ¼ 1) 0.8251(0.0153) 0.7335(0.0044) 0.9382(0.0023)
L = 10,000 MFDCCA-MODWT 0.8703(0.0135) 1.0480(0.0398) 1.0436(0.0061)
MF-X-DMA(h ¼ 0) 0.8943(0.0121) 0.6981(0.0034) 0.8419(0.0042)
MF-X-DMA 0.9456(0.0466) 1.2491(0.0493) 1.2509(0.0307)
(h ¼ 0:5)
MF-X-DMA(h ¼ 1) 0.8943(0.0120) 0.6956(0.0035) 0.8407(0.0042)
Note Values in parentheses are the mean of standard errors
cross-correlation exponents may be significantly overestimated. Meanwhile,

HMFDCCAMODWT ¼ 1:0480 seem to be closer to the theoretical value 0.9.
On the other hand, if the length of time series is long, MF-X-DMA for h ¼ 0 and
h ¼ 1 also outperform MF-X-DMA for h ¼ 0:5. At this time, the cross-correlation
exponent estimated by MFDCCA-MODWT method is more accurate. For example,
when T = 1000 and L = 10,000, HMFDCCAMODWT ¼ 1:0436, HMFXDMA ¼
0:8419 (for h ¼ 0), 1.2509 (for h ¼ 0:5) and 0.8407 (for h ¼ 1), while the theo-
retical value is only 0.9 (see Table 9.1). This maybe due to the existence of
crossovers and then led to the cross-correlation exponents significantly
overestimated by MF-X-DMA with h ¼ 0:5, but slightly underestimated by

MF-X-DMA for h ¼ 0 and h ¼ 1.
In short, if the length of time series is short, MFDCCA-MODWT and
MF-X-DMA for h ¼ 0 and h ¼ 1 may be a better choice. However, if the length is
long enough, MFDCCA-MODWT may perform better than MF-X-DMA.
Moreover, since the empirical time series are usually short in length,
MFDCCA-MODWT and MF-X-DMA for h ¼ 0 and h ¼ 1 may be more accurate,
because for short time series, its estimations are closer to theoretical values.
Therefore, compared with the MF-X-DMA, MFDCCA-MODWT approach may
possess the following advantages.
First, unlike MF-X-DMA, it does not require choosing different moving average
filter and also can be ensure the continuity of empirical time series.
Unlike MFDCCA, segmentation of the entire partition sequence is intermittent,
which leads to intermittent polynomial fitting adjacent segmentation on the interval.
Thus, new errors of the pseudo wave are produced and wave function deviation is
caused, which results in distortion of scaling exponents. However, MODWT
method can decompose time series into different detail coefficients that capture
higher frequency oscillations and represent increasingly finer-scale deviations from
the smooth trend, whereas smooth coefficients capture trends. This transformation
can effectively avoid the defect.
Second, regardless of the sample length is relatively large or short, it can be
better applied to analyze the multifractal characteristics of two cross-correlation
non-stationary time series. As discussed before, if the length of time series is long,
the cross-correlation exponent estimated by MFDCCA-MODWT method is more
reasonable. Meanwhile, if the empirical time series length is short,
MFDCCA-MODWT and MF-X-DMA for h ¼ 0 and h ¼ 1 both may be a better
choice.
9.1.3 Empirical Results and Analysis
9.1.3.1 Data
For the carbon market, we use the EU Emission Allowance (EUA) daily future
price series provided by the ECX, which makes up about 90% of the total daily
futures market transaction volume. The EU ETS for carbon emission trading
products are artificially composed of three separate stages. Given that different
periods are significantly different in the distribution mechanism of carbon quota and
the policy of global climate change, we chose the second phase or relatively
“mature period” from January 1, 2008 to December 31, 2012 for this paper. Data
are derived from the Wind database. As for the energy market, we chose daily
future price series of ICE Brent Crude Oil (Oil), Richards Bay Coal (Coal), UK
Base Electricity (Electricity), and UK Natural Gas (Gas) that trade on the same
platform as representatives. Energy future market data are obtained from the
(a) (b)
0.3 0.3
EUA Brent Coal
0.2 0.2
0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
-0.3 -0.3
-0.4 -0.4
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
(c) (d)
0.2 0.5
Electricity Gas
0.4
0.1
0.3
0.2
0
0.1
0
-0.1
-0.1
-0.2 -0.2
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
Fig. 9.5 Daily returns of carbon and energy markets. Note We place the Brent down translation as
0.2 units in Fig. 9.2a to distinguish it from EUA

Series Mean Std. Skewness Kurtosis Jarque–Bera Probability
Dev.
EUA −0.0009 0.0284 0.1788 6.7692 765.7060 0.0000
Brent 0.0001 0.0240 −0.0847 6.6106 697.8926 0.0000
Coal −0.0001 0.0199 0.3577 36.3409 59,406.1500 0.0000
Electricity −0.0001 0.0194 0.4804 18.7372 13,278.4800 0.0000
Gas 0.0002 0.0313 2.7611 30.6810 42,558.6700 0.0000
Intercontinental Exchange (www.theice.com). Both pairs of carbon and energy

markets daily price series include 1283 data points. Returns are computed by
rt ¼ log pt1 , with Pt as the daily closing price. Moreover, we convert the units of
Pt
all different markets into dollars per ton.

Figure 9.5 presents the curve of daily carbon and energy market returns. To
reveal the detailed statistical properties of the returns for carbon and energy mar-
kets, we display the basic statistics of these time series in Table 9.2. We find that
the carbon and energy market returns possess peakedness and fat tails. The Jarque–
Bera test results reject the null hypothesis at 1% significance level, which indicate
that the logarithmic returns series of the carbon and energy markets does not follow
the efficient market hypothesis, which assumes a normal distribution.
9.1.3.2 MODWT Multiresolution Analysis
Given sample size N = 1283, we conduct the analysis with seven scale levels. As
mentioned earlier, we use the LA(8) wavelet filter. Under this filter, the wavelet
coefficients of scale level J ¼ 1; 2; . . .J are associated with time interval of
½2J1 ; 2J . The use of daily data allows for the ease of determining that scale level
J = 1 corresponds to 1–2 day periods, and the other scale levels J = 2–7 correspond
to day periods of 2–4, 4–8, 8–16, 16–32, 32–64, and 64–128, respectively
(Fig. 9.6).
After processing time series data for seven MODWT transformations, the seven
wavelet coefficient series D1. . .D7 and the seventh layer’s scale coefficient S7 are
shown in Fig. 9.5. Scale wavelet coefficients at all levels are shown in Fig. 9.7,
which indicates that the first layer of wavelet coefficients are the most dramatic
fluctuations that belong to typical high frequency information. When we consider
these small fluctuations non-essential, we can consider them as noise. Larger scales
and flatter fluctuations indicate definite trends.
The overall wavelet correlations are weaker at lower scales although they
become stronger at higher ones, which is consistent with previous research that
argued that correlations are less correlated on the intra-day time scales and become
stronger when moving towards the larger scale (Zhou 2012a). Thus, the daily
timescale indicates the lowest correlations on the scale dimension. Table 9.3 further
demonstrates that although the correlation coefficients of the pairs of time series
reflect its overall relevance, correlations under various scales are different.
9.1.3.3 Cross-Correlation Test
We employ the cross-correlation test proposed by Podobnik et al. to quantify the

cross-correlation between the carbon and energy markets in this subsection. Qcc ðmÞ
is approximately v2 ðmÞ distributed with m degrees of freedom. If the
cross-correlation statistic exceeds the critical value of v2 ðmÞ distribution, then the
cross-correlations are significant. Figure 9.8 shows the cross-correlation statistics
and the critical value of v2 ðmÞ distribution at 5% significance level for the returns
between carbon and energy markets. The degrees of freedom vary from 100 to 103.
The Qcc ðmÞ values that are larger or close to the v2 ðmÞ suggests the existence of
relatively significant cross-correlation between the return series of the carbon and
energy markets.
In order to confirm our results above more carefully, then we employ another
method also proposed by Podobnik et al. (2009a, b, c, 2011), the DCCA
cross-correlation coefficient (qDCCA ) is defined as the ratio between the detrended
2
covariance function FDCCA and two detrended variance functions FDFA (denoted as
FDFA1 and FDFA2 , respectively) expressed as follows:
(a) -3
x 10
2 0.02 0.1
D5 D2
0 0.01 0.05
-2 0 0
-4 -0.01 -0.05
S7
-6 -0.02 -0.1
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
-3
x 10
5 0.02 0.3
D7 D4 D1
0.01 0.2
0 0 0.1
-0.01 0
-5 -0.02 -0.1
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
0.01 0.05 0.6

D6 D3 original series
0.005 0.4
0 0 0.2
-0.005 0
-0.01 -0.05 -0.2

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
(b) -3
x 10
5 0.01 0.04
S7 D5 D2
0.005 0.02
0
0 0
-5
-0.005 -0.02
-10 -0.01 -0.04

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
-3
x 10
5 0.02 0.1
D7 D4 D1
0.01 0.05
0
0 0
-5
-0.01 -0.05
-10 -0.02 -0.1

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
-3
x 10
5 0.05 0.2
0.1
0 0 0
-0.1
-5 -0.05 -0.2
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
Fig. 9.6 MODWT multiresolution analysis for the returns between carbon and energy markets.
a EUA; b Brent; c Coal; d Electricity; and e Gas
2
FDCCA ð nÞ
qDCCA ¼ ð9:8Þ
The value of qDCCA ranges from −1 to 1. If the value equals 1, then a perfect
cross-correlation exists. If the value equals −1, then a perfect anti-cross-correlation
exists. If the value is equal to zero, which means the two series have no
(c) -3
x 10
5 0.01 0.1
S7 D5 D2
0.05
0 0
0
-5 -0.01
-0.05
-10 -0.02 -0.1

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
-3
x 10
10 0.02 0.3
D7 D4 D1
0.01 0.2
5
0 0.1
0
-0.01 0
-5 -0.02 -0.1
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
-3
x 10
5 0.04 0.4
0.02 0.2
0 0 0
-0.02 -0.2
-5 -0.04 -0.4
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
(d) -3
x 10
5 0.01 0.1
S7 D5 D2
0.005 0.05
0
0 0
-5
-0.005 -0.05
-10 -0.01 -0.1

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
-3
x 10
4 0.02 0.1
D7 D4 D1
2 0.01 0.05
0 0 0
-2 -0.01 -0.05
-4 -0.02 -0.1
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
0.01 0.04 0.2

0.005 0.02 0.1
0 0 0
-0.005 -0.02 -0.1
-0.01 -0.04 -0.2

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
Fig. 9.6 (continued)
cross-correlation. We calculate the value of qDCCA based on different values of

window size n (n = 16, 32, 64, 128, 256). As shown in Table 9.4, we can conclude
that the results are consistent with the cross-correlation test as above.
(e) -3
x 10
5 0.02 0.3
S7 D5 D2
0.01 0.2
0
0 0.1
-5
-0.01 0
-10 -0.02 -0.1

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
-3
x 10
4 0.04 0.2
D7 D4 D1
2 0.02 0.1
0 0 0
-2 -0.02 -0.1
-4 -0.04 -0.2
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
0.01 0.15 0.6

0.005 0.1 0.4
0 0.05 0.2
-0.005 0 0
-0.01 -0.05 -0.2

0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
Fig. 9.6 (continued)
EUA&Brent EUA&Coal
1 0.6
lower bound at 95% confidence interval
0.8
cross wavelet correlations 0.4
wavelet correlations
0.6 upper bound at 95% confidence interval
0.4 0.2
0.2 0
0
-0.2
-0.2
-0.4 -0.4
1 2 3 4 5 6 7 1 2 3 4 5 6 7
level level
EUA&Electricity EUA&Gas
0.6
0.8
0.4
0.6
0.2
0.4
0
0.2 -0.2
0 -0.4
1 2 3 4 5 6 7 1 2 3 4 5 6 7
level level
Fig. 9.7 Wavelet correlations for the returns between carbon and energy markets
9.1.3.4 Transmission Direction of the Cross-Correlation
Although the MODWT methodology and the method proposed by Podobnik et al.
can detect the significance level of the cross-correlation in the time series between
Table 9.3 Estimated correlations for the returns between carbon and energy markets ranking on
different time scales (95% confidence interval)
Wavelet correlations
Level EUA and brent EUA and coal EUA and electricity EUA and gas
D1 0.2542 0.1238 0.2383 0.1599
D2 0.3162 0.2160 0.2522 0.2076
D3 0.2178 0.0821 0.1453 0.0857
D4 0.1114 0.0578 0.1691 −0.0204
D5 0.0952 0.0338 0.4057 0.1345
D6 0.0362 0.0335 0.4814 0.3323
D7 0.5258 0.1500 0.7268 0.1373
EUA&Brent EUA&Coal
8 8
Qcc (m)
6
χ2 (m) 6
4
logQ (m)
logQ (m)
4
cc
cc
2
2
0
0
-2
-4 -2
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
log(m) log(m)
EUA&Electricity EUA&Gas
8 8
6
6
4
logQ (m)
logQ (m)
cc
cc
2 4
0
2
-2
-4 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
log(m) log(m)
Fig. 9.8 Cross-correlation statistic for carbon and energy markets
Table 9.4 qDCCA based on different values of window size n

Size 16 32 64 128 256
EUA and brent 0.2151 0.1894 0.1725 0.2393 0.4552
EUA and coal 0.0899 0.0726 0.1091 0.0944 0.3241
EUA and electricity 0.2123 0.2672 0.4294 0.5044 0.5034
EUA and gas 0.0789 0.0865 0.2613 0.1863 0.0715
Table 9.5 Granger causality test in wavelet domain

Series D1 D2 D3 D4 D5 D6 D7
EUA ! Oil 0.2583 0.7729 1.3377 1.8503 1.7051 1.3913 9.0790*
(0.9788) (0.6267) (0.2204) (0.0642) (0.0928) (0.1957) (0.0000)
Oil ! EUA 2.0662* 2.5653* 1.4848* 3.8525* 3.5605* 1.3938 15.6649*
(0.0362) (0.0089) (0.1580) (0.0002) (0.0004) (0.1946) (0.0000)
EUA ! Coal 1.8920 1.5921 1.3509 2.3045* 5.0862* 5.4359* 6.6288*
(0.0576) (0.0089) (0.2226) (0.0188) (0.0000) (0.0000) (0.0000)
Coal ! EUA 2.8194* 4.7897* 2.8577* 4.2798* 5.8795* 3.7310* 4.2895*
(0.0042) (0.0000) (0.0058) (0.0000) (0.0000) (0.0003) (0.0000)
EUA ! Electricity 3.4792* 3.2930* 2.2618* 3.1331* 3.6162* 5.3836* 8.6932*
(0.0006) (0.0010) (0.0212) (0.0016) (0.0004) (0.0000) (0.0000)
Electricity ! EUA 3.4304* 4.3315* 3.8864* 3.2714* 2.0430* 2.0781* 3.6951*
(0.0006) (0.0000) (0.0002) (0.0011) (0.0386) (0.0351) (0.0003)
EUA ! Gas 2.8104* 2.0588* 1.5286 0.9812 1.9003 3.3486* 3.9424*
(0.0043) (0.0370) (0.1534) (0.4488) (0.0563) (0.0008) (0.0003)
Gas ! EUA 0.6100 2.4326* 2.8847* 3.4464* 2.8957* 3.4551* 4.4168*
(0.7701) (0.0131) (0.0054) (0.0006) (0.0034) (0.0006) (0.0000)
Note The original data of time series between carbon and energy markets are transformed by the wavelet filter
(LA(8)) up to time scale 7. The significance levels are in parentheses. The first detail (wavelet coefficient) D1
captures the oscillations with a period length of two to four days. The last detail D7 captures oscillations with a
period length of 64–128 days
*Denotes that the value is significant at 5% level
carbon and energy markets, they cannot distinguish transmission direction.

Therefore, in this subsection, we use the Granger test method which have been
widely used to analysis the transmission direction between two markets, to inves-
tigate the transmission direction of the cross-correlation between carbon and energy
markets.
Table 9.5 indicates that bidirectional Granger causality relationships exist
between carbon and electricity future markets, which suggest that carbon price
changes will cause changes in the price of electricity, and vice versa. This situation
may be attributed to the usually long life cycle of power plants, as well as the high
cost of technical updates. Furthermore, plants tend to buy more emission quotas,
which drive carbon prices up. Moreover, power plants transfer part of the additional
costs due to purchasing quotas, which results in increased electricity prices.
Short-term unidirectional Granger causality relationships between carbon and
coal future markets exist such that future carbon prices are determined by future
coal prices. However, our finding suggests that bidirectional Granger causality
relationships between carbon and coal future markets persist in the long term.
Furthermore, a considerable long term co-integration relationship exists between
the two. The interaction between the carbon and coal markets has been highlighted
to a certain extent. Result indicates that unidirectional Granger causality relation-
ships exist between carbon and natural gas future markets. This condition arises
primarily because the rise in the price of natural gas will drive people to use other
EUA&Brent EUA&Coal
-2 -2
-4 -4
-6 -6
logFq(s)
logFq(s)
-8 -8 q=-10
q=-5
q=0
-10 -10 q=5
q=10
-12 -12
0 1 2 3 4 5 0 1 2 3 4 5
log(s) log(s)
EUA&Electrictiy EUA&Gas
-2 -2
-4 -4
-6 -6
logFq(s)
logFq(s)
-8 -8
-10 -10
-12 -12
0 1 2 3 4 5 0 1 2 3 4 5
log(s) log(s)
Fig. 9.9 The linear relationship between log FqðsÞ and logðsÞ for carbon and the energy markets
cheaper energy sources. When carbon emissions increase, demand for carbon
emissions will increase, which will increase carbon price as well.
However, the Granger causality between carbon and oil future markets remains
insignificant. On the one hand, volatility in carbon prices failed to affect the oil
market during the study period. On the other hand, carbon price trend is insufficient
to affect oil price.
9.1.3.5 MFDCCA-MODWT Analysis
In this paper, we use the MFDCCA-MODWT algorithm, which is an extended

method of MFDCCA, to quantitatively measure cross-correlations and to analyze
cross-correlations for each pair of time series between carbon and energy markets.
Figure 9.9 shows the log–log plots of log FqðsÞ versus logðsÞ between carbon
and energy markets for q = −10, −5, 0, 5, 10. Excellent power-law scaling is
observed in the fluctuation functions for the MFDCCA based on MODWT algo-
rithms. For different q, all of the curves are linear, which suggests that there exist
power-law cross-correlations in each pairs of carbon and energy markets. For
example, EUA is power-law cross-correlated with Brent Oil (see Fig. 9.9). The
power-law cross-correlation relationship indicates that a large increment of price
change in a futures market may be more likely to be followed by a large increment
of price change in the other spatially or temporally correlated futures market.
Using the MFDCCA-MODWT algorithm, we calculate the generalized Hurst
exponent and deduce the singularity spectrum afterwards. Figure 9.10a shows the
plotting of HxyðqÞ of each pair of series. We can conclude that the
cross-correlations for each pair of time series between carbon and energy markets
(a)
1 1
Hxx(q)-EUA Hxx(q)-EUA
Hyy(q)-Brent Hyy(q)-Coal
0.8 Hxy(q) 0.8 Hxy(q)
[Hxx(q)+Hyy(q)]/2 [Hxx(q)+Hyy(q)]/2
H(q)
H(q)
0.6 0.6
0.4 0.4
0.2 0.2
-10 -5 0 5 10 -10 -5 0 5 10
q q
1 1
Hxx(q)-EUA Hxx(q)-EUA
Hyy(q)-Electricity Hyy(q)-Gas
0.8 Hxy(q) 0.8 Hxy(q)
[Hxx(q)+Hyy(q)]/2 [Hxx(q)+Hyy(q)]/2
H(q)
H(q)
0.6 0.6
0.4 0.4
0.2 0.2
-10 -5 0 5 10 -10 -5 0 5 10
q q
(b) EUA EUA&Brent Brent EUA EUA&Coal Coal

1.2 1.2
1 1
0.8 0.8
0.6 0.6
Dq
Dq
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
0.2 0.4 0.6 0.8 1 1.2 0.2 0.4 0.6 0.8 1 1.2
hq hq
EUA EUA&Electricity Electricity EUA EUA&Gas Gas

1.2 1.2
1 1
0.8 0.8
0.6 0.6
Dq
Dq
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
hq hq
Fig. 9.10 Generalized Hurst exponent and spectrum singularity for different q values for carbon
and energy markets
are multifractal because HxyðqÞ is not constant when q varies from −10 to 10,
which is consistent with the results of Junior and Franca (2012), He and Xie (2015).
Moreover, the scaling exponents Hq for q\0 are larger than those for q [ 0, which
implies that the cross-correlated behavior of small fluctuations is more persistent
than that of large fluctuations. For the case of Hxyð2Þ, values that are on the brink of
0.6 reflect a persistent cross-correlation across all series in the long term.
To explore the other multifractal cross-correlation features, we present the sin-
gularity spectrum in Fig. 9.10b to deduce easily which return series exposes rich
multifractal cross-correlation.
Table 9.6 Scaling exponents and multifractality degrees of carbon and energy markets
DHq Dhq
Carbon Energy Cross Carbon Energy Cross
EUA and oil 0.6346 0.4439 0.3201 0.8372 0.5907 0.4822
EUA and coal 0.6346 0.6861 0.4706 0.8372 0.8375 0.6674
EUA and electricity 0.6346 0.6041 0.3284 0.8372 0.7966 0.4869
EUA and gas 0.6346 0.5944 0.3547 0.8372 0.7883 0.5047
The width of the multifractal spectrum can be regarded as an estimate of mul-

tifractal strength. As shown in Table 9.6, the values of spectrum width Dhq are
close to 0.5, which shows that multifractal cross-correlations are important in these
markets. Multifractal cross-correlations are slightly more important between EUA and
coal and between EUA and gas. Moreover, the multifractality degrees of four future
series of the carbon market are larger than those of energy markets except for the coal
market, which implies that the volatilities of carbon markets are more violent than
energy markets, whereas carbon markets are more inefficient than energy markets.
This finding may be due to the relative immaturity of the coal futures market,
which leads to less active market trading and small trading volume. Furthermore,
the trend of coal prices does not fully depict the situation of the European coal in
the study period. In addition, the effects on the carbon market are not sufficiently
notable.
9.1.4 Original of Multifractality
It is generally argued that fat-tailed probability distribution and long-range corre-

lations of large and small fluctuations are two possible sources of multifractal nature
in (mono) financial time series, and other recent works have shown that multi-
fractality can be attributed to dual central limit-like convergence effects (He et al.
2014). However, due to financial time series are not sequences of independent
identically distributed random variables, the convergence theorem does not apply.
Multifractal detrended fluctuation analysis (MFDFA) and other methods which can
be used to study the auto-correlation of a single time series are applied to inves-
tigate the multifractality of the (mono) financial time series. Moreover, we can
quantify the contribution of long-range correlations and fat-fail by comparing the
multifractal degree between the original series and randomly shuffled series and
surrogated series, respectively. Multifractal detrended cross-correlation analysis
(MFDCCA) is used to analyze the cross-correlation between two time series and the
two origins of multifractality in power-law cross-correlated financial time series can
also be distinguished by comparing the multifractal degree between the original
series and randomly shuffled series and surrogated series, respectively (Wang et al.
2011a, b; Zhuang et al. 2014).
Table 9.7 Scaling exponents and multifractality degrees of the original, shuffled, and surrogated
series
DHq Dhq
EUA 0.6346 0.4477 0.392 0.8372 0.6581 0.5752
Brent 0.4439 0.4035 0.3884 0.5907 0.5847 0.5681
Coal 0.6861 0.5809 0.5675 0.8375 0.7871 0.7475
Electricity 0.6041 0.5551 0.4834 0.7966 0.7645 0.652
Gas 0.5944 0.5515 0.5489 0.7883 0.7424 0.7321
The shuffling procedure is described as follows: First, the pairs (p, q) of random
integer numbers are generated with p; q N, where N is the length of the time series
to be shuffled. Second, entries p and q are interchanged. Finally, the above steps are
repeated over 20N times to ensure that the original series are completely shuffled.
To study the contribution of the fat-tailed variations on multifractality, phase
randomization techniques are used. A similar idea was implemented to investigate
multifractality sources in financial returns. The time series generated by the phase
randomization procedure is generally called surrogate data. Various techniques of
phase randomization are available, and we perform amplitude adjusted Fourier
transform surrogate test (Table 9.7).
Figure 9.11 provides the change range of the Hurst index Hq of the original,
shuffled, and surrogated series in carbon and energy markets. Apparently, the range
of change for the Hurst index Hq is significantly reduced after the series is pro-
cessed using the contrasting fractal characteristics of the shuffled and original series.
This finding means that both long-range correlations and fat-tailed distributions
play important roles in contributing to multifractality.
9.1.5 Conclusions
This section used the maximum overlap wavelet transform approach to investigate
the nonlinear structure and cross-correlation between carbon and energy future
markets across different time scales. Furthermore, we utilized the statistical test
proposed by Podobnik et al. and the Granger causality test to further examine
whether the cross-correlation is significant on the whole, as well as the transmission
direction of the cross-correlations. To quantify the cross-correlation between carbon
and energy markets, we applied MFDCCA-MODWT, and we determined the
original multifractality through shuffling and surrogating procedures. Our empirical
results can be summarized as follows.
First, cross-correlation is relatively significant based on the analysis of MODWT
multiresolution and the significance of the statistic Qcc ðmÞ. Furthermore, with the
Granger causality test, the carbon price between energy price, such as coal, natural,
(a) EUA EUA

1 1.2
original series
0.9 shuffled series 1
surrogated series
0.8 0.8
0.7 0.6
Hq
Dq
0.6 0.4
0.5 0.2
0.4 0
-0.2
-10 -5 0 5 10 0.2 0.4 0.6 0.8 1 1.2
q hq
Brent Brent
0.8 1.2
1
0.7
0.8
0.6
0.6
Hq
Dq
0.4
0.5
0.2
0.4
0
-0.2
-10 -5 0 5 10 0.2 0.4 0.6 0.8 1
q hq
(b) Coal Coal

1 1.2
original series
shuffled series 1
0.8 surrogated series
0.8
0.6
Hq
Dq
0.6
0.4
0.2
0.4
0
0.2 -0.2
-10 -5 0 5 10 0.2 0.4 0.6 0.8 1 1.2
q hq
Electricity Electricity
1 1.2
0.9 1
0.8 0.8
0.7 0.6
Hq
Dq
0.6 0.4
0.5 0.2
0.4 0
-0.2
-10 -5 0 5 10 0.2 0.4 0.6 0.8 1 1.2
q hq
(c) Gas Gas

1 1.2
original series
0.9 shuffled series 1
surrogated series
0.8
0.8
0.7
0.6
Dq
Hq
0.6
0.4
0.5
0.2
0.4
0
0.3
0.2 -0.2
-10 -5 0 5 10 0 0.2 0.4 0.6 0.8 1
q hq
Fig. 9.11 Scaling exponents and multifractal spectra of the original, shuffled, and surrogated
series
and electricity prices, are maintained in a certain degree of causal

relationship. A bidirectional Granger causality between carbon and electricity prices
in future markets is determined as well. However, the Granger causality relationship
between the carbon and oil prices is not evident because numerous factors influence
oil prices. Given that the price of oil is mainly driven by the major political events
and abnormal climate events, establishing the transmission mechanism between
them is ineffective.
Second, our evidence indicates that the cross-correlation between carbon and
energy market price returns is persistent and multifractal. Thus, an increase in
carbon price tends to be followed by an increase in energy price. The existence of
long-range cross-correlations implies that past changes in carbon prices can
improve the predictability of energy future prices.
Finally, the DHq and Dhq of the original, shuffled, and surrogated series in the
carbon and energy future markets decrease in long-range correlations and fat-tailed
distributions, which significantly contribute to multifractality. This finding is in
agreement with the West Texas Intermediate crude oil, metal, and gold futures
markets according to Wang et al. and Guo et al., respectively.
Through the empirical study, our findings on long-range cross-correlations
present important modeling implications. For example, economic models that
incorporate long-range cross-correlations can better capture the interactions
between carbon and energy future markets. In this sense, the nonlinear structure
model is better than the conventional econometric model. On the one hand, from
the multifractality degree, we can conclude that the carbon market is more ineffi-
cient than the energy market. In other words, the carbon market is much more
vulnerable to external shock and then the risks of carbon market are much bigger.
Thus, investors should consider different market risks in order to formulate a rea-
sonable risk diversification strategy. On the other hand, European carbon futures
basically formed a good relationship with various energy markets and gradually
entered maturity. However, at the Copenhagen Climate Conference in December
2009, countries failed to form a valid agreement, which seemingly increased the
uncertainty of the carbon market. In addition, market transactions have insufficient
freedom, too much government intervention, and too few market main body and
other defects, which resulted in the carbon markets being more inefficient than
energy markets. Thus, financial regulators should strengthen financial regulation,
and improve the legal system and information disclosure policy. Thereby, investors
can timely and accurately grasp the real information and to make better investment
decisions.
9.2 Multifractal Features of EUA and CER Futures

Markets: MFDFA-EMD
The massive CO2 emissions from fossil fuel consumption and their influence on
climate change have become a major ecological and political issue worldwide. The
“Kyoto Protocol,” which was established in February 2005, aims to limit the total
global CO2 and other greenhouse gas emissions. In the same year, the European
Climate Exchange (ECX) futures were listed on the carbon dioxide emission futures
to assist the buyers and the sellers who are engaged in the emissions trading for risk
management. European Union Emissions Trading Scheme (EU ETS) is the largest
carbon market worldwide in terms of market value and trading volume according to
Zhang and Wei (2010). The two major instruments traded in the EU ETS are
Certified Emission Reduction (CER) credits since 2007 and the European Union
Allowances (EUA) since 2005. As participants in futures markets function as
hedgers and speculators, the market may be considered a place for transferring price
risks. Therefore, the prices of EU ETS, CER, and EUA futures markets are ana-
lyzed in this study.
The relationship between EU ETS spot and futures prices has been empirically
analyzed. The earliest time-series analysis of spot market prices conducted by Bruce
and Gao (1996); they reveal that the market fundamentals model is unsuitable for
EUA and recommend the use of generalized autoregressive conditional
heteroscedasticity (GARCH) model to simulate the price. Current studies are difficult
to compare because they cover different time periods and use different approaches.
Empirical methodologies are grouped into structured and econometric models.
Structured models used to outline carbon markets, and analyze carbon price move-
ments from the perspective of supply–demand equilibrium (Seifert et al. 2008).
Econometric models include linear models, such as Auto-Regressive and
Moving Average Model (ARMA) and vector autoregression (VAR). For nonlinear
models, rescaled range analysis (R/S) and modified R/S model are used by Feng
et al. (2011) and reveal that carbon market is a biased random walk characterized by
fractals. Empirical mode decomposition (EMD) is proposed by Huang (1971), an
analytical approach to assess nonlinear and non-stationary time series and
decompose time or price series into several independent intrinsic mode functions
(IMFs) and one residue. Since then, EMD is adopted for analysis of spot and futures
prices, as well as for forecasting and cross-correlation analysis. In the present study,
we use EMD to analyze the multiscale and multifractality of EUA and CER futures
market prices.
Although structured models are used to understand the characteristics of carbon
markets, actual application is difficult because constructing an appropriate demand
and supply model is complicated under the dynamic and changeable market con-
text. Econometric methods perform well for short term or price-series analysis and
forecasting but cannot satisfactorily explain the intrinsic driving force of carbon
price changes. The accuracy of R/S analysis is significantly reduced if the sample
data sequence is short term or non-stationary. Detrended fluctuation analysis
9.2 Multifractal Features of EUA and CER Futures Markets: MFDFA-EMD 191
(DFA) is proposed by Peng et al. (1994) to address the limitation of R/S analysis.
And the detrended fluctuation method “grafting” to a multifractal field and establish
multifractal DFA (MFDFA).
In the recent years, an increasing number of researchers, including Cao et al.
(2013), Dai (2009), have used DFA or MFDFA to analyze the multifractality of
stock markets; although this approach can effectively describe the multifractal and
long-term memory characteristics of non-stationary time-series method, it also
presents some limitations. For time-series analysis, MFDFA requires detrended
processing, in which polynomial fitting is commonly used. Nonetheless, polyno-
mial selection exhibits a variable mode ranging from 1 to k order. Moreover,
segmentation of the entire sequence after partition is intermittent because of
MFDFA formation. This phenomenon leads to intermittent polynomial fitting
adjacent segmentation on the interval, thus introducing new errors of the pseudo
wave and causing deviation of the wave function, resulting in distortion of the
scaling exponents.
Basing on daily price data of carbon emission rights in futures markets of
Certified Emission Reduction (CER) and European Union Allowances (EUA), we
analyze the multiscale characteristics of the markets by using empirical mode
decomposition (EMD) and multifractal detrended fluctuation analysis (MFDFA)
based on EMD. The complexity of the daily returns of CER and EUA futures
markets changes with multiple time scales and multilayered features. The two
markets also exhibit clear multifractal characteristics and long-range correlation.
We employ shuffle and surrogate approaches to analyze the origins of multifrac-
tality. The long-range correlations and fat-tail distributions significantly contribute
to multifractality. Furthermore, we analyze the influence of high returns on mul-
tifractality by using threshold method. The multifractality of the two futures mar-
kets is related to the presence of high values of returns in the price series.
9.2.1 MFDFA-EMD Methodology
EMD method is used to extract and eliminate the trend of the items from the
original sequence (Zachmann and Von Hirschhausen 2008; Gu and Zhou 2010).
A technique using sliding window technology is used to improve MFDFA interval
segmentation method and traditional MFDFA method of non-overlapping seg-
mentation changes for continuous overlapping. Although this process enables a
large increase in subinterval (2Ns to N s þ 1), it also eliminates pseudo fluctua-
tion errors generated by polynomial fitting because of the discontinuity of seg-
mentation data at the junction.
MFDFA-based EMD assumes that the first two steps of MFDFA are similar. In
the third step, EMD is used to decompose the series instead of polynomial
detrending. Considering a given time series xðtÞ, where t ¼ 1; 2; 3; . . .N, the EMD
algorithm can be described as follows:
Step 1: Identify all the maxima and minima of carbon price series xðtÞ;
Step 2: Generate their upper and lower envelopes, fmax ðtÞ and fmin ðtÞ, with cubic
spline interpolation;
In the EMD procedure, we have used the find and diff functions among the
Matlab functions library to find out the values of the carbon price with all the local
maximal and minimal positions and the spline function to form the cubic spline
interpolation.
Step 3: Calculate the point-by-point mean xðtÞ from the upper and lower
envelopes:
fmax ðtÞ þ fmin ðtÞ

xðtÞ ¼ ð9:9Þ
2
Step 4: Extract the mean from carbon price series and define the difference
between xðtÞ and xðtÞ as lðtÞ:
lðtÞ ¼ xðtÞ xðtÞ ð9:10Þ
Step 5: Verify whether the new series lðtÞ satisfies two conditions: one is the
numbers of local extreme and zero crossings in the whole data set must
be equal or differ by 1 at most and the other is the mean value of the
“upper envelope” (defined by the local maxima) and the “lower envel-
ope” (defined by the local minima) must be zero at any time point (Wang
et al. 2011a). If lðtÞ is an Intrinsic Mode Function (IMF), denote lðtÞ as
the ith IMF and replace xðtÞ with the residue rn ðtÞ ¼ xðtÞ lðtÞ. The ith
IMF is often denoted as yi ðtÞ and the i is called its index; otherwise,
replace xðtÞ with lðtÞ;
Step 6: Repeat above steps until the residue satisfies some stopping criteria. One
typical stopping criterion proposed by Cao et al. (2012)
fmax ðtÞ fmin ðtÞ

/ðtÞ ¼ ð9:11Þ
2

lðtÞ

cð t Þ ¼ ð9:12Þ
/ðtÞ
where /ðtÞ and cðtÞ represent the mode amplitude and the evaluation
function, respectively.
Thus the sifting is iterated until cðtÞ\h1 for some prescribed fractions ð1 /Þ
of the total duration, while cðtÞ\h2 for the remaining fractions, where h1 and h2 are
two thresholds aimed to guarantee globally small fluctuations in the mean while
taking into account locally large excursions. One can typically set / ¼ 0:05, h1 ¼
0:05 and h2 ¼ 10 h1 .
The algorithm stops if the conditions are satisfied, otherwise the algorithm
continues. Finally, the trend is given by
X
n
rn ðtÞ ¼ xðtÞ yi ðtÞ: ð9:13Þ
i¼1
where rn is a residue representing the trend of the time series.

We start MFDFA based on EMD from a given time series xðiÞ, where
i ¼ 1; 2; 3; . . .N, and
Step 1′: construct the cumulative sum
X
t
uð t Þ ¼ xi ; i ¼ 1; 2; 3; . . .N ð9:14Þ
i¼1
Step 2′: divide the new series uðtÞ into Ns disjoint segments of the equal size s,
where Ns ¼ int½N=s. Each segment is denoted by uv, such that
uv ðiÞ ¼ uðl þ iÞ, 1 i s, where l ¼ ðv 1Þs.
Step 3′: determine the local trend uev in each segment uv with polynomial fitting
method. MFDFA method is called MFDFA-l (MFDFA-1, MFDFA-2,
and so on) when a polynomial of the order l is adopted in this step. We
obtain the residual sequence using
ev ðiÞ ¼ uv ðiÞ ug
v ðiÞ; 1is ð9:15Þ
The third step of the algorithm in the baseline MFDFA implies the detrending of
the segments through polynomial fitting. One segment in the version based on
EMD is used to compute an EMD local trend for each segment uv as uev ¼ rn ðiÞ,
where uev is the local trend and rn ðiÞ is the local trend based on the EMD approach.
Step 4′: constructed the series of residuals using the trend function as follows:
ev ðiÞ ¼ uv ðiÞ rn ðiÞ; 1is ð9:16Þ
The detrended fluctuation function F(v, s) for the segment uv by using the
residuals determined in Eq. (9.13) and given by
1X s
½F ðv; sÞ2 ¼ ½ev ðiÞ2 ð9:17Þ
s i¼1
Step 5′: derive the overall detrended fluctuation function at the qth order as
( )1q
1 X Ns
F q ðsÞ ¼ ½F ðv; sÞq ð9:18Þ
Ns v¼1
where q represents any real value except q = 0. In case q = 0, the

formula is
( )
1 X Ns
F0 ðsÞ ¼ exp ln½F ðv; sÞ ð9:19Þ
Ns v¼1
Finally, a power-law relationship based on different timescales s is established

between Fq(s) and the time scale s
Fq ðsÞ sHq ð9:20Þ
Generally, the exponent Hq is dependent on q. The segment v with large vari-

ance (i.e., large deviation from the corresponding fit) for positive q dominates the
average Fq(s). Therefore, Hq describes the scaling behavior of the segments with
large fluctuations if q is positive. Generally, large fluctuations are characterized by a
low scaling exponent Hq for multifractal time series. The segment v with small
variance for negative q dominates the average Fq(s). Thus, the scaling exponent Hq
for negative q values describes the scaling behavior of segments with small fluc-
tuations and is characterized by large scaling exponents. Based on Cheng et al.
(2013), Liu et al. (2016), we can see that the kinds of fluctuation related to q have
persistence when Hq > 0.5, and anti-persistence when Hq < 0.5. Especially, when
q = 2, H2 is the Hurst exponent. In other words, for q = 2, the standard DFA-EMD
procedure is retrieved.
If 0.5 < H2 < 1, the correlations of time series are persistent (positive). An
increase is likely to be followed by another increase, and the larger H2, the stronger
of persistence. If 0 < H2 < 0.5, the correlations of time series are anti-persistent
(negative). An increase is likely to be followed by a decrease, and the smaller H2,
the stronger of anti-persistence. If H2 = 0.5, the time series display a random walk
behavior, and the marker is weakly efficient.
The conclusion of multifractality is obtained from the dependence of Hq on the
values of the fluctuation order q.
tðqÞ ¼ qHq 1 ð9:21Þ
The time series exhibits a multifractal nature if the multifractal exponent t(q) is a
nonlinear function of q. The important variable set hq * Dq is obtained using
Legendre transform and defined by
hq ¼ Hq þ qH 0 ðqÞ ð9:22Þ
Dq ¼ qhq tðqÞ ð9:23Þ
The singularity strength function hq characterizes the singularities of the time

series. The multifractal spectrum Dq describes the singularity content of the time
series.
To measure the degree of multifractality, define DH as
DHq ¼ Hqmax Hqmin ð9:24Þ
where DH is the range of the generalized Hurst exponent Hq. A high DH value
results to a strong degree of multifractality and high risks, and vice versa.
According to Chen et al. (2016), the strength of multifractality is also estimated
by the width of multifractal spectrum and given by
Dhq ¼ hqmax hqmin ð9:25Þ
9.2.2.1 Data Description
For futures market, we use daily price series provided by the ECX, which accounts
for about 90% of the total daily futures market transaction volume. The EU ETS for
carbon emission trading products are artificially divided into three separate stages.
The first and pilot period lasted from 2005 to 2007, and the second or relatively
“mature period” ranged from 2008 to 2012. The third period is from 2013 to 2020.
We select the settlement prices of EUA and CER futures provided by the ECX
from March 14, 2008 to December 31, 2012 as sample data with a total of 1231
samples. We select the annual contract expires in December because the largest
trading volume is obtained when the futures contract expires in December. EUA
futures market prices (EUAF) and CER futures market prices (CERF) sample data
units are in euros per ton of carbon dioxide equivalent. Data are derived from the
Wind database.
Future prices and returns of CER and EUA are illustrated in Fig. 9.12a, b,
respectively. The trend of EUAF and CERF index is identical, but the overall
tendency of EUA futures price is slightly higher than that of CER as shown in
Fig. 9.12a, b. Moreover, in October 2012 EUA futures prices shocked and CER
futures prices plummeted and were close to 0. Futures price trend of the two
markets was closely linked with the international climate negotiations. In December
2009, before the Copenhagen international climate conference negotiations, whe-
ther EUA or CER market, the futures prices are both rising, indicating that the
carbon finance markets is full of confidence in the forthcoming meeting agreement
(a)
30
CERF
EUAF
Futures prices
20
10
0
2008 2009 2010 2011 2012
year
(b)
0.5
0
Returns
-0.5
-1
2008 2009 2010 2011 2012
year
Fig. 9.12 Futures prices and returns of CERF and EUAF. Note We place the EUAF up translation
as 0.2 units in Fig. 9.1b to distinguish it from CERF
which is conducive to the development of the market. Under this anticipation, EUA
futures prices soared to around €17 ton. Similar to the EUA market, CER futures
prices also rose to €15 ton. However, the Copenhagen international climate con-
ference and the agreement did not play any binding effect of carbon emission
reduction. The agreement of the conference does not specify the emission reduction
of developed countries, and also doesn’t clearly point out the sources of funding and
technical support of energy saving and emission reduction in developing countries.
CER market reacted strongly to the agreement, its futures prices fell after the
conference.
And we can see that since 2010, the carbon finance market price fell all the way,
the reason can be summarized as follows: Firstly, the fall in euro exchange rate
severely affected the carbon financial markets whose transaction money is euro at
that time, the carbon finance market futures prices has made a new low record, and
investors were definitely not confident of the carbon finance market development;
Secondly, a downturn global economy and the CER market’s hit have put a price
pressure on the carbon credits by CER.
Although James and Edmister (1983) found there is no such cross-sectional
correlation exist between the price changes and volume, however, a considerable
number of studies have found that positive correlations exist between the price
changes and volume. Additionally, since the trading volumes data for the markets,
which we can get, are incomplete. Then, we can not draw conclusions whether there
is such a relationship between carbon futures prices and trading volumes.

Series Mean Std. Dev. Skewness Kurtosis Jarque–Bera Probability
CERF −0.0039 0.0423 −3.4236 37.3393 62,836.0400 0.0000
EUAF −0.0010 0.0273 0.0821 6.9214 789.4599 0.0000
Figure 9.12b illustrates the logarithmic returns of CER and EUA futures market,
respectively, as well as the significant volatility clustering and long-term memory
effect of the time series. The agglomeration effect on CER futures market is evident,
and high degree of volatility leads to high risks.
Table 9.8 shows the basic statistical properties of the log returns rate of CERF
and EUAF series. The variance of the EUAF is lower than that of the CERF,
resulting in lower volatility of market risk. The skewness of the CERF is −3.424,
revealing a clear left-side feature. The peak value of Kurtosis is 37.3393, which is
higher than the peak normal distribution 3.
The peak value of the EUAF is also higher than the normal distribution, indi-
cating that the characteristics of the log returns of the two market rate series exhibit
an evident peak tail distribution, which is consistent with the results of Jarque–Bera
test.
Jarque–Bera test results reject the null hypothesis at 1% significance level,
indicating that the logarithmic returns series of the CERF and EUAF does not obey
the efficient market hypothesis (Cao and Han 2014; Cao et al. 2014b, 2016). On the
whole, the returns time series of the two market are in line with asymmetric (such as
exist left-side feature), higher kurtosis and fat-tail distribution. That is to say, the
EMH does not conform to the carbon market, it is necessary to employ fractal
theory to study the long-term memory.
9.2.2.2 Decomposition Based on EMD Method
The IMFs and residues derived by applying EMD to the CERF and EUAF daily
returns series are shown in Fig. 9.13a, b, respectively. Nine IMF components and a
residue trend component are obtained. All IMFs are listed in the order by which
they are extracted, that is, from the highest frequency to the lowest frequency. The
last item in Fig. 9.13 is the residue. The following conclusions are formulated based
on Fig. 9.13 and Table 9.9: (1) all IMFs present changing frequencies and ampli-
tudes, which differ from any harmonic. This finding reflects the complex changes of
CERF and EUAF daily returns with multiple time scales and multilayering. (2) The
amplitudes of the IMFs decrease with changing frequency from high to low. For
example, all the amplitudes of IMF1 in Fig. 9.13 are higher than the other intrinsic
mode functions. The variance contribution of the IMF1 account is higher than 50%.
Hence, IMF1 is the dominant mode of the CERF and EUAF. (3) The amplitude of
fluctuation (IMF1 to IMF5) exhibits an evident “sustainability”. With the average
point of view that a past considerably fluctuates, the trend of the IMF1 indicates
IMFs and residual

(a)
0.5 0.2
0 0
-0.5 -0.2
0 500 1000 1500 0 500 1000 1500
0.2 0.05
0 0
-0.2 -0.05
0 500 1000 1500 0 500 1000 1500
0.05 0.02
0 0
-0.05 -0.02
0 500 1000 1500 0 500 1000 1500
0.02 0.01
0 0
-0.02 -0.01
0 500 1000 1500 0 500 1000 1500
0.01 0.05
residual
0 0
-0.01 -0.05
0 500 1000 1500 0 500 1000 1500
(b)
IMFs and resdiual
0.1 0.1
0 0
-0.1 -0.1
0 500 1000 1500 0 500 1000 1500
0.05 0.05
0 0
-0.05 -0.05
0 500 1000 1500 0 500 1000 1500
0.02 0.01
0 0
-0.02 -0.01
0 500 1000 1500 0 500 1000 1500
-3
x 10
0.01 5
0 0
-0.01 -5
0 500 1000 1500 0 500 1000 1500
-3 -3
x 10 x 10
5 2
residual
0 0
-5 -2
0 500 1000 1500 0 500 1000 1500
Fig. 9.13 IMFs and residues based on EMD: a CERF and b EUAF
Table 9.9 Measures of IMFs and residues for CERF and EUAF derived from EMD
Observed CERF EUAF
Mean Variance Variance Mean Variance Variance
period contribution period contribution
(%) (%)
IMF1 2.29 0.0011 50.9722 2.8 0.0005 55.0416
IMF2 6.06 0.0004 21.5675 8.2 0.0002 21.1243
IMF3 18.36 0.0003 14.9538 13.09 0.0001 12.8807
IMF4 24.12 0.0001 4.4706 28.6 0 5.7483
IMF5 51.25 0.0001 2.7404 42.41 0 2.7332
IMF6 123 0 0.8227 123 0 1.114
IMF7 410 0 1.141 307.5 0 0.8997
IMF8 615 0 0.2937 307.5 0 0.1698
IMF9 615 0 0.6092 1230 0 0.2607
Residue 0.0001 2.4288 0 0.0276
substantial fluctuation trends in the futures and vice versa. (4) The CERF and
EUAF components present different fluctuation cycles. For example, CERF
decomposition components with mean fluctuation cycles of 2–3, 6, 18–19, 24–25,
51–52, 123, and 615 days. However, the daily returns of the long-term memory are
mainly short-term mean fluctuation cycles (2–3, 6, 18–19, 24–25 days, etc.). This
finding reveals that the CERF and EUAF contains complex multiple time scales.
9.2.2.3 Multifractal Analysis of CER and EUA Futures
We also empirically analyze the multifractality with MFDFA-based EMD method

in this section. Figure 9.14 presents the generalized Hurst exponents Hq of CERF
and EUAF series when q varies from −10 to 10.
Figure 9.14 shows that the Hq values of CERF and EUAF are not constant and
present nonlinear decreasing trend with increasing order, indicating that the CERF
and EUAF markets are multifractal. The change range of CERF is also higher than
EUAF, suggesting that CERF series possesses stronger multifractal characteristics
in the sample interval with relatively high risks. The risk is relatively lower than
CERF, whereas EUAF possesses weaker multifractal characteristics (Table 9.10).
In particular, the H2 values of CERF and EUAF series based on the
MFDFA-EMD algorithm are 0.5213 and 0.5104, respectively. The values are
higher than 0.5, indicating that long -range correlation exists in the price of the
product in carbon emissions of the CER and EUA futures market.
In order to determine whether the calculated H2 values fall within its 95%
confidence interval, we have employed the fractionally autoregressive integrated
moving average (ARFIMA) process to carry out statistical tests.
(a) 1
CERF
0.8 EUAF
Hq
0.6
0.4
0.2
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
(b) 1
Dq
0.5
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
hq
Fig. 9.14 Generalized Hurst exponents and multifractal spectra of CERF and EUAF
Table 9.10 Scaling exponents and multifractality degrees of CERF and EUAF
DHq Dhq
CERF 0.5572 0.9679
EUAF 0.4056 0.7855
Table 9.11 The results of d = 0.0104 d = 0.0213

Hurst exponents estimated by
MFDFA-EMD for time series Theoretical value 0.5104 0.5213
generated by ARFIMA MFDFA-EMD 0.5194(0.0085) 0.5303(0.0071)
processes 95% confidence [0.5003–0.5385] [0.5143–0.5464]
interval
Note For each length of time series, we generate 10 pairs. The
results show the average of 10 auto-correlation exponents of
those generated pairs; in parentheses there are the means of
standard errors
As shown in Table 9.11, based on ARFIMA process, we simulate some artificial

time series with the length of N, which is the same as the actual sample. Meanwhile,
HMFDFAEMD ¼ 0:5194 (d = 0.0104) and HMFDFAEMD ¼ 0:5303 (d = 0.0213),
suggesting that the Hurst exponents seem to be closer to its theoretical value. From
the 95% confidence interval of them, the numerical values of the simulation are fall
within the confidence interval. In other words, this can be considered that their
values are more than 0.5.
The generalized Hurst index Hq of carbon emissions of the CER futures market
is higher than that of the EUA when q < 0. As all the generalized Hurst index Hq
are higher than 0.5, the effects of slight fluctuations on carbon emission of the
CERF are amplified. The state-persistence of carbon emissions of the CERF is
higher than that of the EUA futures. However, the internal factors of the market
significantly affect the price trend and the CERF possesses weaker anti persistence.
The generalized Hurst index Hq is lower than 0.5 when q > 2, and the effect of the
market prices intensifies if the futures market price of the carbon emission product
exhibits tremendous volatility. This phenomenon causes the futures market prices to
exhibit anti-persistence characteristics. External factors in the market affect the
entire futures market price.
The sample data of the CERF series are stronger than those of the EUAF in
terms of volatility degree (see Table 9.3). This part describes that the ability of
carbon emissions CER futures strongly captures external causes in the market, and
the reaction sensitivity of the information is also strong. Thus, the risk of the CER
futures market is higher than that of the EUA futures market. The multifractal
spectrum graph hq * Dq of carbon emission CER and EUA futures markets are
shown in Fig. 9.3b. The multifractal spectrum graphs of the two futures markets are
asymmetric, and the distribution range of the whole function curve is wide; this
finding suggests that the market price fluctuation is higher, and the price distribu-
tion is not uniform, regardless of carbon emissions CER futures market or the EUA
futures market. Moreover, Dhq [ 0 in Fig. 9.3b indicates that the incidences of the
high extreme points of carbon emissions futures prices are higher than that of low
extreme points.
The left end of CERF is significantly higher than the right, that is, the high event
of carbon emission CERF plays a leading role on the overall price. The opportunity
located in the low price is higher than that in the high price, particularly carbon
emission CERF increases. The left end of CERF is significantly lower than the
right, that is, the lower event of carbon emission EUAF mostly affects the overall
price. Thus, carbon emission EUAF declines as a whole.
9.2.3 Origins of Multifractality
9.2.3.1 Shuffling Procedure
Two main sources of multifractality are presented in Wang et al. (2011a, b), Zhou
(2009). One source is the long-range of large and small fluctuations, and the other
source is the fat-tail distribution of the return series. By comparing the multifractal
degree between the original series and randomly shuffled series, we quantify the
contribution of long-range correlations. The shuffling procedure is described as
follows: first, the pairs (p, q) of random integer numbers are generated with
Table 9.12 Scaling exponents and multifractality degrees of the original, shuffled, and surrogated
series
DHq Dhq
CERF 0.5572 0.5154 0.4141 0.7356 0.6826 0.5595
EUAF 0.4056 0.2037 0.1907 0.5558 0.3137 0.2995
p; q N, where N is the length of the time series to be shuffled. Second, the entries
p and q are interchanged. Finally, the above steps are repeated more than 20N times
to ensure that the original series are completely shuffled.
9.2.3.2 Surrogating Procedure
The classic method is proposed by Theiler et al. (1992) of quantifying fat-tail

contribution by comparing the multifractal degrees between the original and sur-
rogated series, which are created using Fourier phase randomization. The present
study uses a simple method proposed by Song and Shang (2011) and Kantelhardt
et al. (2002). For a given distribution, a series of random numbers is generated with
Gaussian distribution and rearranged to obtain the series frt ; t ¼ 1; 2; . . .N g, which
exhibits similar rank ordering to the original series fxt ; t ¼ 1; 2; . . .N g. Therefore, rt
should be ranked as n in the series frt ; t ¼ 1; 2; . . .N g if and only if xt is ranked n in
the original series fxt ; t ¼ 1; 2; . . .N g. The shuffling and surrogating procedure we
present here is consistent with the literatures.
For CERF, the change range of the Hurst index Hq is significantly reduced after
the series is processed using the contrast fractal characteristics of the shuffled and
original series. The original series Hq value is decreased from 0.8446 to 0.2874,
shuffled series is decreased from 0.8433 to 0.3279, and surrogated series is
decreased from 0.7167 to 0.3026 (Table 9.12). The differences vary from 0.5572 to
0.5154 and 0.4141, indicating that multi-fractal characteristics are significantly
reduced after the series is shuffled and surrogated.
The change range of the multifractal spectrum is shown in Fig. 9.15 The mul-
tifractal spectrum width of the CERF returns series significantly narrows after the
series is shuffled and surrogated. This finding suggests that the multifractal char-
acteristic of CERF returns series significantly reduces after processing. Therefore,
the multifractal characteristics of the CERF returns series return are explained by
the existence of significant long-term memory and fat-tail distribution. This con-
clusion is similar to that obtained with EUAF.
CERF
(a) 1
original
0.8 shuffled
surrogated
Hq
0.6
0.4
0.2
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
(b) 1
0.5
Dq
-0.5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
hq
EUAF
(c) 1
original
0.8 shuffled
surrogated
Hq
0.6
0.4
0.2
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
(d) 1
0.8
Dq
0.6
0.4
0.2
0.4 0.5 0.6 0.7 0.8 0.9 1
hq
Fig. 9.15 Scaling exponents and multifractal spectra of the original, shuffled, and surrogated
series
CERF
1
0.8
Δhq
0.6
0.4
0.2
1 2 3 4 5 6 7 8 9
λ
EUAF
0.8
original
shuffle
0.6
surrogate
Δhq
0.4
0.2
1 1.5 2 2.5 3 3.5 4
λ
Fig. 9.16 Multifractality of time series, in which returns higher than the threshold kr are deleted
and interpolated
9.2.4 Risk Analysis
Research on the risks of the financial market focuses on larger fluctuation. This
section investigates the influence of high returns on the multifractality by using the
method proposed by Kristoufek (2012) within the certain threshold kr, where
k 2 f1; 2; . . .10g and r is the standard deviation of price returns series. The values
higher than the specific threshold kr are eliminated, and the eliminated data points
are replaced by linear interpolation. The time series retains higher values as the
threshold kr increases. Thus, the complexity of financial markets under different
high returns is studied.
Figure 9.16 presents the dependence of the degree of multifractality Dhq, which
is defined by the range of singularity strength hq, on the threshold kr for original,
shuffled, and surrogated data. Complexity is related to the presence of very high
returns values in the two futures markets. However, the shuffled and surrogated
data, which remove nonlinearity from the original data, are independent from the
threshold kr. Futures markets values are created by non-trivial interactions between
heterogeneity agents and the influence of internal and external events. The results of
Fig. 9.16 indicate that more complex markets could possibly produce high returns.
9.2.5 Conclusion and Implications
In this section, we investigate the multifractality in the return series of the carbon
emission of CER and EUA future markets by using MFDFA-EMD. We also dis-
cuss the multiscales of the two futures markets, origins of multifractality, and effects
of high returns on multifractality by using EMD, shuffled and surrogated approa-
ches, and threshold method. The empirical findings are summarized as follows.
First, the CER and EUA future markets possess multifractal characteristics with
long-range correlation. The price of the CER futures market is higher than that of
the EUA, resulting in higher risks. Moreover, the higher event of carbon emission
CERF significantly affects the overall price. Generally, the prices of the carbon
emission CERF increase and those of the EUA futures market decrease.
Second, the DHq and Dhq of the original, shuffled, and surrogated series in the
CER and EUA future markets decrease in long-range correlations and fat-tail dis-
tributions, which significantly contribute to multifractality. This finding is in
agreement with the West Texas Intermediate crude oil, metal and gold futures
markets according to Maskawa et al. (2013), Morales and Callaghan (2012), Junior
and Franca (2012).
Finally, the complexity of carbon emission markets is related to the presence of
very high returns. However, the shuffle and surrogate data, which remove the
nonlinearity from the original data, are independent from the threshold Tr. These
results also indicate that more complex markets could possibly produce high
returns.
The summary of results is presented as follows.
(1) The presence of multiscales and multifractality suggests that nonlinear models,
such as multi-fractal model, may not be utilized when modeling the behavior of
carbon emission futures market prices.
(2) Carbon emission futures markets exhibit similar multifractal source to the crude
oil, metal, and gold futures market. This finding indicates that the carbon
emission futures market on multifractal sources presents general characteristics
of the market. MFDFA-EMD method is recommended for further research as
an extension of the multifractal detrended fluctuation analysis (Zachmann and
Von Hirschhausen 2008; Gu and Zhou 2010; Tsai 2013) to determine the
relationship of multifractality and long-range correlations. This method can
eliminate pseudo fluctuation errors generated by polynomial fitting caused by
the discontinuity of segmentation data at the junction.
(3) The effects of high returns on multifractality can be utilized to provide a risk
assessment method in complex financial markets. Investment risks are reduced
through transfer of financial tools and dispersion of the function. Financial
regulators should also strengthen financial regulation and improve the legal
system and information disclosure policies. Therefore, investors can timely and
accurately obtain actual information to formulate investment decisions.
9.3 Cross-Correlation Among Mainland China, US,

and Hong Kong Stock Markets VC-MF-DCCA
As a complex system, the financial market has become increasingly interconnected

worldwide. The recent bankruptcy of Lehman Brothers in 2008 and the Euro crisis
made the overall market crisis a global epidemic. Disruption in a single market may
exert a significant effect on other financial markets because of interdependence.
Several scholars have studied financial markets by using various approaches,
ranging from statistical physics and nonlinear dynamics to computer science.
In recent years, several scholars have developed different MF-DCCA methods to
describe the cross-correlation between two non-stationary time series; some of these
methods include MF-X-DFA (Zhou 2008); MF-X-DMA (Jiang and Zhou 2011),
which is based on MF-DMA (Gu and Zhou) and DMA (Alessio et al. 2002);
MF-HXA (Kristoufek 2012); MF-X-PF (Xie et al. 2015); and MF-DPXA (Qian
et al. 2015). Cao et al. (2014a, b) proposed the multifractal asymmetric detrended
cross-correlation analysis (MF-ADCCA) method to investigate asymmetric
cross-correlations in non-stationary time series. Zhang et al. (2015) also used
MF-ADCCA to study the asymmetric characteristics of cross-correlations between
PM2.5 concentration and meteorological factors. In addition, Cao et al. (2016)
developed an asymmetric MF-DCCA method that is conducted based on the dif-
ferent directions of risk conduction (DMF-ADCCA). Shi et al. (2014) introduced a
method called “multiscale multifractal detrended cross-correlation analysis”
(MM-DCCA); MM-DCCA may help present significantly richer information than
MF-DCCA by sweeping all the ranges of scale at which the multifractal structures
of complex systems are discussed. Yin and Shang (2013) used multiscale detrended
fluctuation analysis and multiscale detrended cross-correlation analysis to investi-
gate the auto-correlation and cross-correlation between American and Chinese
stock markets from 1997 to 2012. Liu et al. (2016) used empirical mode decom-
position (EMD), MF-DCCA, and principal component analysis (PCA) to propose
the EMD-MFDCCA-PCA method. When Zhao et al. (2011) studied the traffic
signals using the MF-DCCA method, they found that crossovers arising from
extrinsic periodic trends made the scaling behavior difficult to analyze. Therefore,
they introduced a Fourier filtering method to eliminate the trend effects and sys-
tematically investigate the multifractal cross-correlation of simulated and real traffic
signals. Pal et al. (2016) characterized the multifractal nature and power–law
cross-correlation between any pair of genome sequence through an integrative
approach that combines 2D MF-DCCA and chaos game representation. From their
analysis, they observed the existence of multifractal nature and power–law
cross-correlation behavior between any pair of genome sequences.
Several studies have investigated the volatility of the financial market from
different viewpoints. For example, scholars have analyzed the volatility clustering
of the financial market. Junior and Franca (2012) found that high-volatility markets
are directly related to strong correlations between markets. Maskawa et al. (2013)
discovered that market-wide price co-movement becomes prominent before and
9.3 Cross-Correlation Among Mainland China, US, and Hong Kong … 207
after a large price decline, such as an endogenous market crash. Wang et al. (2011a)
established long-term power–law cross-correlations in the absolute values of returns
that quantify risk and found that these cross-correlations decay significantly more
slowly than cross-correlations between returns. Podobnik et al. (2010) established
long-term magnitude cross-correlations in price fluctuation and physiological time
series; both of which are healthy and pathological. Moreover, a new methodology
was proposed to assess and quantify inter-market relations. The approach was based
on the correlations among market index, index volatility, market index cohesive
force, and meta-correlations. Although research on cross-correlation and volatility
is abundant, only a few studies have been conducted on volatility-constrained
cross-correlation. Therefore, a volatility-constrained multifractal detrended
cross-correlation analysis (VC-MF-DCCA) method was developed in the present
study to investigate the cross-correlation among the markets of Mainland China,
United States, and Hong Kong, as based on Ochiai and Nacher (2014).
Several studies have also investigated the risk contagion of stock markets.
Chinese scholars have studied the contagion effect of Mainland China and Hong
Kong stock markets. Zhang et al. (2006) found that the response to the events of the
mainland market is earlier than that of the Hong Kong stock market. Liao (2010)
discovered the single-direction spillover effect on liquidity from Shanghai to Hong
Kong markets in the beginning of the crisis. In the late stage of the crisis, a one-way
spillover effect on liquidity and volatility was observed from Hong Kong to
Shanghai markets. Several foreign scholars have also studied the contagion effect of
Chinese and American stock markets. Morales and Callaghan (2012) found that the
US crisis did not exert contagion effects on Asian economies. If US equity markets
influence Asian economies, then scholars should consider that the deceleration of
the US economy caused problems in Asian financial markets. Chen et al. (2016)
determined that US economic variables, such as dividend price ratio, dividend
yield, and industrial production, strongly forecast the future monthly volatilities of
the Chinese stock market. Considering that the conclusions of these studies are not
uniform, the present study also examined the directionality of influence between
stock markets.
This study focuses on multifractal detrended cross-correlation analysis of the
different volatility intervals of Mainland China, US, and Hong Kong stock markets.
A volatility-constrained multifractal detrended cross-correlation analysis
(VC-MF-DCCA) method is proposed to study the volatility conductivity of
Mainland China, US, and Hong Kong stock markets. Empirical results indicate that
fluctuation may be related to important activities in real markets. The Hang Seng
Index (HSI) stock market is more influential than the Shanghai Composite Index
(SCI) stock market. Furthermore, the SCI stock market is more influential than the
Dow Jones Industrial Average stock market. The conductivity between the HSI and
SCI stock markets is the strongest. HSI was the most influential market in the large
fluctuation interval of 1991–2014. The autoregressive fractionally integrated
moving average method is used to verify the validity of VC-MF-DCCA. Results
show that VC-MF-DCCA is effective.
9.3.1 VC-MF-DCCA Method
First, we define the average and standard deviation of Ri;t for the given period [ti, tf]
as follows:
1 X
EðRi;t Þ ¼ Ri;t ; ð9:26Þ
ðtf ti Þ ti t\tf
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 X
rðRi;t Þ ¼ Ri;t EðRi;t Þ2 : ð9:27Þ
tf ti ti t\tf

Second, we let k be a subset of all the time points tti t\tf . Thereafter, we
define the expectation value and standard deviation as follows:
1 X
EðRt ; kÞ ¼ Rt ; ð9:28Þ
#k t2k
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 X
rðRt ; kÞ ¼ Rt EðRt ; kÞ2 ; ð9:29Þ
#k t2k
where #k denotes the number of elements of k.

Third, we set

k½t1 ;t2 :a;b ¼ t 2 ½ti ; tf jt1 t\t2 ; a r Ri;t Ri;t \b r Ri;t ; ð9:30Þ
where a; b 2 ½0; þ 1Þ. In this study, we set Ds ¼ t2 t1 as three years. If Ds is less

than three years, then the volatility-constrained data are too minimal to use the
MF-DCCA method. If Ds is more than three years, then the results cannot accu-
rately reflect the dynamic condition of the stock market. We consider time series {x
(t)} and {y(t)}; {x(t)} 2 k, {y(t)} 2 k, and t1 t t2 .
New series are built as follows:
X
i
xðiÞ ¼ ½xðtÞ x; ð9:31Þ
t¼1
X
i
yðiÞ ¼ ½yðtÞ y; ð9:32Þ
t¼1
P#k P#k
where x ¼ #k
1
t¼1 xðtÞ and y ¼ #k
1
t¼1 yðtÞ.
{x(i)} and {y(i)} are then divided into #ks ¼ intð#k=sÞ non-overlapping seg-
ments of equal size s. We can calculate from the end of the time series with the
same method to avoid losing effective messages because the total length of the time
series is not exactly the segments’ integral multiple of s. Thus, a total of 2#ks
segments exist. The local variance for each of the segments is computed through a
least-square polynomial fit of the series shown as follows:
1X s
Fv ðsÞ ¼ jxv ðiÞ xev ðiÞjjyv ðiÞ yev ðiÞj: ð9:33Þ
s i¼1
For v ¼ 1; 2; . . .; 2#ks , we define the qth-order fluctuation function as

" #1=q
1 Xs
2#k
Fxy ðq; sÞ ¼ Fv ðsÞq=2 ; q 6¼ 0; ð9:34Þ
2#ks v¼1
" #
1 Xs
2#k
Fxy ð0; sÞ ¼ exp ln Fv ðsÞ ; q ¼ 0: ð9:35Þ
4#ks v¼1
If two time series have a long-range correlation, then Fxy ðq; sÞ will increase in
the power rate. The relationship of fluctuation is
Fxy ðq; sÞ Shxy ðqÞ : ð9:36Þ
If 0\hxy ð2Þ\0:5, then the time series have a long-range negative correlation. If
hxy ð2Þ ¼ 0:5, then the time series are not correlated. When 0:5\hxy ð2Þ\1, as the
value increases, the long-range correlation of the series becomes stronger.
In this study, we used the asymmetric property of the three indexes’ returns and
found that this property is important in quantifying the strength of influence from
one index to another. For example, if hxy ðqÞð0:5\hxy ðqÞ\1Þ of R1;t (R1;t is the base
asset) and R2;t is larger than R1;t and R2;t (R2;t is the base asset), then the main
directionality of the influence is from R1;t to R2;t .
9.3.2 Validation of the VC-MF-DCCA Method
The time series exhibits complex behavior; thus, we must determine if they can be
measured at successive time intervals. However, the coefficients are not robust and
can be misleading if outliers are present because real-world data are characterized
by a high degree of non-stationarity. We used artificial data generated by a
two-component fractionally autoregressive integrated moving average (ARFIMA)
model to test the universality of the VC-MF-DCCA method.
X
1 X
1
yi ¼ W an ðq1 Þyin þ ð1 WÞ an ðq2 Þy0in þ ei ; and ð9:37Þ
n¼1 n¼1
X
1 X
1
y0i ¼ ð1 WÞ an ðq1 Þyin þ W an ðq2 Þy0in þ e0i ; ð9:38Þ
n¼1 n¼1
where ei and e0i denote two independent and identically distributed Gaussian vari-
ables with zero mean and unit variance. an ðqÞ denotes the statistical weights defined
by an ðqÞ ¼ Cðn qÞ=ðCðqÞCð1 þ nÞÞ, where C denotes the Gamma function; q
are parameters ranging from −0.5 to 0.5 (related to the VC-MF-DCCA exponent,
ða ¼ 0:5 þ qÞ); and W is a free parameter that ranges from 0.5 to 1.0 and controls
the strength of the power-law cross-correlation between yi and y0i . Given the
two-component ARFIMA process of Eqs. (9.24) and (9.25), we generated new time
series yi and y0i , which are characterized by different values of an ðq1;2 Þ and W. We
set q ¼ 0:4, W ¼ 0:5 (maximum strength of power–law cross-correlation), and
ei ¼ e0i .
We used new time series yi and y0i to replace the original data to examine the
validity of VC-MF-DCCA. To avoid the contingency, we created three groups in
which the number of data is 500 (less than the original data; Group A), 700 (similar
to the original data; Group B), and 1000 (more than the original data; Group C). We
generated 10 time series in every group and calculated the average. The results are
shown in Table 9.13.
In Table 9.13, all the results of hxy are close to 0.9. The numerical values of the
simulation fall within the 95% confidence interval. This result means that the
VC-MF-DCCA method is valid. As the numbers increase, the standard errors
decrease, and the validity of the method improves.
To strengthen the validation, we set q ¼ 0:3, W ¼ 0:5 (maximum strength of
power–law cross-correlation), and ei ¼ e0i again. We created three more groups in
which the number of data is 500 (less than the original data; Group D), 700 (similar
to the original data; Group E), and 1000 (more than the original data; Group F). We
then generated 10 time series in every group and calculated the average. The results
are shown in Table 9.14.
In Table 9.14, all the results of hxy are close to 0.8. The numerical values of the
simulation also fall within the confidence interval. This result means the
VC-MF-DCCA method is valid. As the number of data increases, the standard
errors decrease, and the validity of the method improves.
Therefore, VC-MF-DCCA is valid. Data validity improves as the number of data
increases.
Table 9.13 Results of ARFIMA and VC-MF-DCCA methods
Group hxy ½0; 1=2Þ Standard errors hxy ½1=2; 1Þ Standard errors hxy ½1; þ 1Þ Standard errors
A 0.8292(0.7413, 0.9171) 0.0389 0.8713(0.7935, 0.9491) 0.0344 0.7010(0.6095, 0.7925) 0.0404
B 0.8784(0.8247, 0.9321) 0.0237 0.8952(0.8410, 0.9493) 0.0239 0.7037(0.6733, 0.7341) 0.0134
C 0.9583(0.9389, 0.9778) 0.0086 0.9876(0.9763, 0.9989) 0.0050 0.9803(0.9700, 0.9910) 0.0046
Note Values in parentheses are the 95% confidence interval
9.3 Cross-Correlation Among Mainland China, US, and Hong Kong …
211
212
Table 9.14 Results of ARFIMA and VC-MF-DCCA methods

Group hxy ½0; 1=2Þ Standard errors hxy ½1=2; 1Þ Standard errors hxy ½1; þ 1Þ Standard errors
D 0.7479(0.6582, 0.8375) 0.0396 0.7233(0.6743, 0.7722) 0.0216 0.7064(0.6485, 0.7643) 0.0256
E 0.7398(0.6732, 0.8065) 0.0295 0.8525(0.8093, 0.8957) 0.0191 0.7933(0.7465, 0.8401) 0.0207
F 0.9247(0.8913, 0.9581) 0.0147 0.9746(0.9610, 0.9883) 0.0060 0.9881(0.9821, 0.9940) 0.0026
Note Values in parentheses are the 95% confidence interval
9 Multifractal Detrend Method with Different Filtering
9.3.3.1 Data Description
The daily closing prices of the Shanghai Composite Index (SCI), Dow Jones
Industrial Average (DJIA), and Hang Seng Index (HSI) from January 12, 1991 to
June 30, 2015 were used as sample data. These three markets are representative and
have a close relationship (Tseng and Li 2012). Many factors, such as statutory
holidays, affect the trading time of Mainland China, US, and Hong Kong stock
markets. Therefore, the non-overlapping data of Mainland China, the United States,
and Hong Kong were removed. A total of 5625 data groups were obtained for
empirical analysis.
We let Pi;t , i ¼ ð1; 2; 3Þ be the closing prices of SCI, DJIA, and HSI at time
tðti t\tf Þ, where ti and tf denote the initial and final time points for the data sets.
The logarithmic rate of return is defined as
Ri;t ¼ ln Pi;t ln Pi;ðt1Þ : ð9:39Þ
The logarithmic returns of SCI, DJIA, and HSI at time t are provided respec-
tively as follows:
R1;t ¼ ln SCIt ln SCIt1 ; R2;t ¼ ln DJIAt ln DJIAt1 ; and

ð9:40Þ
R3;t ¼ ln HSIt ln HSIt1 :
9.3.3.2 Effect of Different Volatility-Constrained Intervals
To explore the influence of the volatility constraint on long-range correlation, we

used the VC-MF-DCCA method on three-year data. Owing to the asymmetric
feature of the stock markets, the effect of the volatility constraint on long-range
correlation may differ from one stock market to another. We defined the
volatility-constrained stock market as the base stock market. Considering the
rationality of the volatility-constrained interval, we selected the three intervals of
a ¼ 0; b ¼ 1=2; a ¼ 1=2; b ¼ 1 and a ¼ 1; b ¼ 1.
Figure 9.17 shows the results of VC-MF-DCCA between SCI and DJIA in
different volatility-constrained intervals. The numbers on the X axis represent three
years (1991–2004). The base stock markets in Fig. 9.1a, b are SCI and DJIA,
respectively. Figure 9.1 suggests that hxy ½1; 1Þ from 1991 to 1999 was much lower
than that in the other intervals. However, from 2009 to 2011, hxy ½1; 1Þ achieved the
maximum value, and hxy ½1=2; 1Þ achieved the minimum value. From 2006 to 2008,
hxy ½1=2; 1Þ achieved its highest value. hxy ½0; 1=2Þ showed a similar trend as hxy
without the volatility constraint.
Figure 9.18 shows the results of VC-MF-DCCA between SCI and HSI. We
selected three intervals, i.e., a ¼ 0; b ¼ 1=2; a ¼ 1=2; b ¼ 1; and a ¼ 1; b ¼ 1.
(a) (b)
1 0.8
hxy hxy
0.9 hxy[0,1/2) 0.7 hxy[0,1/2)
hxy[1/2,1) hxy[1/2,1)
0.8 hxy[1,inf) hxy[1,inf)
0.6
0.7
0.5
hxy
hxy
0.6
0.4
0.5
0.3
0.4
0.3 0.2
0.2 0.1
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
year year
Fig. 9.17 Results of VC-MF-DCCA method between SCI and DJIA
(a) (b)
0.9 0.9
hxy hxy
hxy[0,1/2) 0.8 hxy[0,1/2)
0.8
hxy[1/2,1) hxy[1/2,1)
hxy[1,inf) 0.7 hxy[1,inf)
0.7
0.6
0.6
hxy
hxy
0.5
0.5
0.4
0.4
0.3
0.3 0.2
0.2 0.1
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
year year
Fig. 9.18 Results of VC-MF-DCCA between SCI and HSI
The base stock markets in Fig. 9.2a, b are SCI and HSI, respectively. Figure 9.2
also shows that the fluctuation trends of hxy ½0; 1=2Þ and hxy without the volatility
constraint are similar. From 1991 to 1996, hxy ½1; 1Þ was lower than that in the
other intervals but achieved the highest value from 2009 to 2011. hxy ½1=2; 1Þ
achieved its maximum value from 2006 to 2008.
Figure 9.19 shows the results of VC-MF-DCCA between DJIA and HSI in three
different intervals of a ¼ 0; b ¼ 1=2; a ¼ 1=2; b ¼ 1; and a ¼ 1; b ¼ 1. The base
stock markets in Fig. 9.3a, b are DJIA and HSI, respectively. hxy ½1; 1Þ was not the
lowest from 1991 to 1999 but was the highest from 2009 to 2011. hxy ½0; 1=2Þ
presented a similar trend as hxy without the volatility constraint, and hxy ½1=2; 1Þ
attained its maximum value from 2006 to 2008.
The analysis of the three figures shows the following. First, hxy ½0; 1=2Þ presented
a similar trend as hxy without the volatility constraint. The data are similar to the
original data because the interval has a slight fluctuation. Second, hxy ½1=2; 1Þ
attained its maximum value from 2006 to 2008 and immediately decreased from
2009 to 2011; thus, it is related to the financial crisis in 2008. Lastly, hxy ½1; 1Þ from
1991 to 1996 is lower than that in the other intervals in Figs. 9.1 and 9.2. However,
the situation is different in Fig. 9.3. The stock market of Mainland China developed
(a) (b)
0.8 1
hxy
hxy
hxy[0,1/2) 0.9
0.7 hxy[0,1/2)
hxy[1/2,1)
hxy[1/2,1)
hxy[1,inf) 0.8 hxy[1,inf)
0.6
0.7
0.5
hxy
hxy
0.6
0.4
0.5
0.3
0.4
0.2 0.3
0.1 0.2
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
year year
Fig. 9.19 Results of VC-MF-DCCA between DJIA and HSI
from 1991 to 1996, and few significant fluctuations occurred. hxy ½1; 1Þ had the
maximum value in all the figures from 2009 to 2011. It is also related to the global
financial crisis. The financial crisis caused a large fluctuation in global stock
markets.
9.3.3.3 Directionality of Stock Markets’ Correlation
The degree of influence from one stock market to another differs because of the
asymmetric feature of stock markets. We introduced the volatility-constrained
correlation to identify the direction of correlation (Cao et al. 2014b). In relation to
Eqs. (9.37) and (9.38), we defined the correlation coefficient as follows:
1 X ðR1;t EðR1;t ; kÞÞ ðR2;t EðR2;t ; kÞÞ

CðR1;t ; R2;t ; kÞ ¼ : ð9:41Þ
#k t2k rðR1;t ; kÞ rðR2;t ; kÞ
when ti t1 \t2 tf , the two different types of correlations are as follows.
F½a; bðsÞ ¼ CðR1;t ; R2;t ; k½s;s þ Ds;a;b Þ; ð9:42Þ
and
CðsÞ ¼ CðR1;t ; R2;t ; k½s;s þ Ds;0;1 Þ: ð9:43Þ
Fisher’s Z transformation was then implemented for Correlation C.
ZðCÞ ¼ ð1=2Þ logðð1 þ CÞ=ð1 CÞÞ: ð9:44Þ
We subtracted the constrained correlation F ½1; 1Þ to determine the net effect of

volatility and thus strengthen the correlation. Then, Z1 Zc can be interpreted as
the net volatility effect on correlation.
(a) (b)
0.05 0.04
(Z1-Zc) of SCI base asset - (Z1-Zc) of DJIA base asset

SCI base asset
0.04 DJIA base asset
0.03
0.03
0.02 0.02
0.01
Z1-Zc
0.01
0
-0.01 0
-0.02
-0.01
-0.03
-0.04 -0.02
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9
year
Fig. 9.20 Comparison of Z1 Zc between SCI and DJIA
(a) (b)
0.16 0.06
(Z1-Zc) of SCI base asset - (Z1-Zc) of HSI base asset
SCI base asset

0.14 HSI base asset
0.04
0.12
0.02
0.1
0.08 0
Z1-Zc
0.06 -0.02
0.04
-0.04
0.02
-0.06
0
-0.02 -0.08
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9
year
Fig. 9.21 Comparison of Z1 Zc between SCI and HSI
Figure 9.20 shows the results of Z1 Zc between SCI and DJIA. Figure 9.4a
shows the net volatility effect on correlation Z1 Zc with a different base market.
Figure 9.4b shows the difference between the ðZ1 Zc Þ of SCI base market and the
ðZ1 Zc Þ of DJIA base market. The last bar in Fig. 9.4b denotes the difference in
the average of ðZ1 Zc Þ with SCI and the average of ðZ1 Zc Þ with DJIA. The
average of Z1 Zc with SCI is higher than that with DJIA. This result implies that
SCI is more influential to DJIA than the opposite case.
The comparison of Z1 Zc between SCI and HSI is shown in Fig. 9.21.
Figure 9.21a shows the net volatility effect on correlation Z1 Zc constrained by
SCI and HIS markets. Figure 9.21b shows the difference between the ðZ1 Zc Þ of
SCI base market and the ðZ1 Zc Þ of HSI base market. On the average, Z1 Zc
constrained by the HSI market is larger than that constrained by the SCI market.
(a) (b)
0.07 0.015
(Z1-Zc) of DJIA base asset - (Z1-Zc) of HSI base asset

0.06 0.01
0.005
0.05
0
0.04
Z1-Zc
-0.005
0.03
-0.01
0.02
-0.015
DJIA base asset

0.01 -0.02
HSI base asset
0 -0.025
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9
year
Fig. 9.22 Comparison of Z1 Zc between DJIA and HSI
This result suggests that the HSI market has a greater effect on the SCI market than
the opposite case.
Figure 9.22a shows a comparison of Z1 Zc constrained by DJIA and HSI
markets. Determining which is larger between DJIA and HSI base markets is
difficult when Fig. 9.22a is used. However, the average of Z1 Zc constrained by
HSI is larger than that constrained by DJIA. This finding also indicates that HSI is
more influential to DJIA than the opposite case. This conclusion is similar to the
conclusions drawn from Figs. 9.20 and 9.21.
We conclude from the above analysis that HSI is the most influential stock
market among the three. The SCI stock market is more influential than the DJIA
market. We used the US closing price at 4 a.m. and Chinese closing price at 3 p.m.
as the closing prices of the same day because of the time difference. Thus, we can
discriminate the directionality of influence between different stock markets. This
knowledge can help identify the origin of market instabilities. The constrained
intervals fluctuate significantly given the parameter a ¼ 1; b ¼ 1. Therefore, the
HSI stock market is the most influential market in the significantly fluctuating
interval of 1991–2014 and is the origin of market instability.
9.3.4 Conclusions
This section investigated the volatility-constrained multifractal detrended

cross-correlation among Mainland China, US, and Hong Kong stock markets. The
main conclusions are as follows.
(1) We found that hxy ½0; 1=2Þ exhibits a similar trend as hxy without a volatility
constraint by using the VC-MF-DCCA method. hxy ½1=2; 1Þ attained its maxi-
mum value from 2006 to 2008 and immediately decreased from 2009 to 2011.
Thus, it is related to the financial crisis in 2008. hxy ½1; 1Þ has the maximum
value in all figures from 2009 to 2011. It is also related to the global financial
crisis. This result indicates that the financial crisis may have caused a large
fluctuation in global stock markets.
(2) The HSI stock market is more influential than the SCI stock market. The SCI
stock market is more influential than the DJIA stock market. The conductivity
between HSI and SCI stock markets is the strongest. The HSI stock market is
the most influential market in the large fluctuating intervals from 1991 to 2014.
(3) We used the ARFIMA method and found that the VC-MF-DCCA method is
valid. The validity of the method improves as the number of data increases.
The VC-MF-DCCA method studies multifractal detrended cross-correlation with
different fluctuation intervals. The length of the fluctuation intervals is defined in
this method. The method can reflect large and small fluctuations in real stock
markets. Our findings enrich the study on cross-correlation. Moreover, the corre-
lation analysis indicates the directionality of correlation of a stock market. These
two methods can help investors avoid risks and make rational decisions. Moreover,
the results of these two methods can be used as a reference by governments. These
two methods can inspire governments to emulate and absorb the virtues of other
stock markets. Thus, China’s stock market can improve.
References
E. Alberola, J. Chevallier, B. Chèze, Price drivers and structural breaks in European carbon prices
2005–2007. Energy Policy 36, 787–797 (2008)
A. Bruce, H.Y. Gao, Applied Wavelet Analysis with S-Plus (Springer, New York Inc., 1996)
D.W. Bunn, C. Fezzi, Interaction of European Carbon Trading and Energy Prices (2007)
G.X. Cao, Y. Han, Does the weather affect the Chinese stock markets? Evidence from the analysis
of DCCA cross-correlation coefficient. Int. J. Mod. Phys. B 28, 1450236 (2014)
G.X. Cao, L.B. Xu, J. Cao, Multifractal detrended cross-correlations between the Chinese
exchange market and stock market. Phys. A 391, 4855–4866 (2012)
G.X. Cao, J. Cao, L.B. Xu, Asymmetric multifractal scaling behavior in the Chinese stock market:
based on asymmetric MF-DFA. Phys. A. 392(4), 797–807 (2013)
G.X. Cao, J. Cao, L.B. Xu, L.Y. He, Detrended cross-correlation analysis approach for assessing
asymmetric multifractal detrended cross-correlations and their application to the Chinese
financial market. Phys. A 393, 460–469 (2014a)
G.X. Cao, Y. Han, W.J. Cui, Y. Guo, Multifractal detrended cross-correlations between the CSI
300 index futures and the spot markets based on high-frequency data. Phys. A 414, 308–320
(2014b)
G.X. Cao, Y. Han, Q.C. Li, W. Xu, Asymmetric MF-DCCA method based on risk conduction and
its application in the Chinese and foreign stock markets. Phys. A (2016). https://doi.org/10.
1016/j.physa.2016.10.002
S.M. Chen, L.Y. He, Welfare loss of China’s PM2.5 pollution: how to make personal vehicle
transportation policy? China Econ. Rev. 31(4), 106–118 (2014)
References 219
H. Cheng, J. Huang, Y. Guo et al., Long memory of price–volume correlation in metal futures
market based on fractal features. Trans. Nonferrous Met. Soc. China 23(10), 3145–3152 (2013)
J. Chen, F. Jiang, H.Y. Li, W.D. Xu, Chinese stock market volatility and the role of U.S. economic
variables. Pacific-Basin Finance J. 39, 70–83 (2016)
J. Chevallier, A model of carbon price interactions with macroeconomic and energy dynamics.
Energy Econ. 33, 1295–1312 (2011)
J. Chevallier, Time-varying correlations in oil, gas and CO2 prices: an application using BEKK,
CCC and DCC-MGARCH models. Appl. Econ. 44, 4257–4274 (2012)
F.J. Convery, L. Redmond, Market and price developments in the European Union emissions
trading scheme. Rev. Environ. Econ. Policy 1, 88–111 (2007)
M. Dai, Multifractal analysis of a measure of multifractal exact dimension. Nonlinear Anal. 70(2),
1069–1079 (2009)
G. Daskalakis, R.N. Markellos, Are electricity risk premia affected by emission allowance prices?
Evidence from the EEX, Nord Pool and Powernext. Energy Policy 37, 2594–2604 (2009)
Z. H. Feng, L.L. Zou, Y.M. Wei, Carbon price volatility: evidence from EU ETS. Appl. Energy 88
(3), 590–598 (2011)
C. Fezzi, D.W. Bunn, Structural interactions of European carbon trading and energy prices.
J. Energy Markets 2, 53–69 (2009)
G.F. Gu, W.X. Zhou, Detrending moving average algorithm for multifractals. Phys. Rev. E 82,
011136 (2010)
L.Y. He, W.S. Xie, Predictability and market efficiency in agricultural futures markets-a
perspective from price-volume correlation based on wavelet coherency analysis. Fractals. 23(2),
1550003: 1–14 (2015)
L.Y. He, S. Yang, W.S. Xie et al., Contemporaneous and asymmetric properties in the
price-volume relationships in China’s agricultural futures markets. Emerg. Markets Finance
Trade 50(sup1), 148–166 (2014)
N.E. Huang, Z. Shen, S.R. Long et al., The empirical mode decomposition and the Hilbert
spectrum for nonlinear and non-stationary time series analysis. Proc R Soc Lond A: Math Phys
Eng Sci. 454, 903–995 (1971)
C. James, R.O. Edmister, The relation between common stock returns trading activity and market
value. J. Finance 38(4), 1075–1086 (1983)
Z.Q. Jiang, W.X. Zhou, Multifractal detrending moving-average cross-correlation analysis. Phys.
Rev. E 84, 664–675 (2011)
L.S. Junior, I.D.P. Franca, Correlation of financial markets in times of crisis. Phys. A 391,
187–208 (2012)
J.L. Kanen, Carbon Trading and Pricing (Environmental Finance Publications, 2006)
J.W. Kantelhardt, S.A. Zschiegner, E. Koscielny-Bunde et al., Multifractal detrended fluctuation
analysis of nonstationary time series. Phys. A 316(1), 87–114 (2002)
R. Khalfaoui, M. Boutahar, H. Boubaker, Analyzing volatility spillovers and hedging between oil
and stock markets: evidence from wavelet analysis. Energy Econ. (2015)
N. Koch, Dynamic linkages among carbon, energy and financial markets: a smooth transition
approach. Appl. Econ. 46, 715–729 (2014)
L. Kristoufek, Multifractal height cross-correlation analysis: a new method for analyzing
long-range cross-correlations. Europhys. Lett. 95, 68001 (2012)
S.G. Liao, Study on the spillover effect between Shanghai and Hongkong stock market under the
background of subprime crisis. J. Shanghai Li Xin Account. Coll. 3, 43–57 (2010). (in
Chinese)
M. Lineesh, C.J. John, Analysis of non-stationary time series using wavelet decomposition. Nat.
Sci. 8, 53–59 (2010)
H.M. Liu, J.C. Zhang, Y.J. Cheng, C. Lu, Fault diagnosis of gearbox using empirical mode
decomposition and multi-fractal detrended cross-correlation analysis. J. Sound Vib. 385,
350–371 (2016)
M. Mansanet-Bataller, A. Pardo, E. Valor, CO2 prices, energy and weather. Energy J. 73–92
(2007)
J. Maskawa, J. Murai, K. Kuroda, Market-wide price co-movement around crashes in the Tokyo
stock exchange. Evol. Inst. Econ. Rev. 10, 81–92 (2013)
K. Minu, M. Lineesh, C.J. John, Wavelet neural networks for nonlinear time series analysis. Appl.
Math. Sci. 4, 2485–2495 (2010)
L. Morales, B.A. Callaghan, The current global financial crisis: do Asian stock markets show
contagion or interdependence effects? J. Asian Econ. 23, 616–626 (2012)
G.P. Nason, R. Von Sachs, Wavelets in time-series analysis. Philos. Trans. R. Soc. Lond. Ser.
A Math. Phys. Eng. Sci. 357, 2511–2526 (1999)
T. Ochiai, J.C. Nacher, Volatility-constrained correlation identifies the directionality of the
influence between Japan’s Nikkei 225 and other financial markets. Phys. A 393, 364–375
(2014)
M. Pal, V.S. Kiran, P.M. Rao, P. Manimaran, Multifractal detrended cross-correlation analysis of
genome sequences using chaos-game representation. Phys. A 456, 288–293 (2016)
C.-K. Peng, S.V. Buldyrev, S. Havlin et al., Mosaic organization of DNA nucleotides. Phys. Rev.
E 49(2), 1685–1689 (1994)
D.B. Percival, H.O. Mofjeld, Analysis of subtidal coastal sea level fluctuations using wavelets.
J. Am. Stat. Assoc. 92, 868–880 (1997)
B. Podobnik et al., Cross-correlations between volume change and price change. Proc. Natl. Acad.
Sci. U S A 106, 22079–22084 (2009a)
B. Podobnik, I. Grosse, D. Horvatic, S. Ilic, P.Ch. Ivanov, H.E. Stanley, Quantifying
(2009b)
B. Podobnik, D. Horvatic, A.M. Petersen, H.E. Stanley, Cross-correlations between volume
change and price change. Proc. Natl. Acad. Sci. U S A 106, 22079–22084 (2009c)
B. Podobnik, D. Wang, D. Horvatić, I. Grosse, H.E. Stanley, Time-lag cross-correlations in
collective phenomena. Europhys. Lett. 90, 68001 (2010)
B. Podobnik, Z.-Q. Jiang, W.-X. Zhou, H.E. Stanley, Statistical tests for power-law
cross-correlated processes. Phys. Rev. E 84, 066118 (2011)
X.Y. Qian, Y.M. Liu, Z.Q. Jiang, B. Podobnik, W.X. Zhou, H.E. Stanley, Detrended partial
cross-correlation analysis of two nonstationary time series influenced by common external
forces. Phys. Rev. E 91, 062816 (2015)
O. Renaud, J.-L. Starck, F. Murtagh, Prediction based on a multiscale decomposition. Int.
J. Wavelets Multiresolut. Inf. Process. 1, 217–232 (2003)
J. Seifert, M. Uhrig-Homburg, M. Wagner, Dynamic behavior of CO2 spot prices. Environ. Econ.
Manag. 56(2), 180–194 (2008)
W.B. Shi, P.J. Shang, J. Wang, A.J. Lin, Multiscale multifractal detrended cross-correlation
analysis of financial time series. Phys. A 403, 35–44 (2014)
R. Smale, M. Hartley, C. Hepburn, J. Ward, M. Grubb, The impact of CO2 emissions trading on
firm profits and market prices. Clim. Policy 6, 31–48 (2006)
S. Soltani, D. Boichu, P. Simard, S. Canu, The long-term memory prediction by multiscale
decomposition. Sig. Process. 80, 2195–2205 (2000)
J. Song, P.J. Shang, Effect of linear and nonlinear filters on multifractal detrended cross-correlation
analysis. Fractals 19, 443–453 (2011)
S. Suhartono, S. Subanar, Development of model building procedures in wavelet neural networks
for forecasting non-stationary time series. Eur. J. Sci. Res. 34, 416–427 (2009)
J. Theiler, S. Eubank, A. Longtin, B. Galdrikian, J.D. Farmer, Testing for nonlinearity in time
series: the method of surrogate data. Phys. D 77–94 (1992)
I.C. Tsai, Volatility clustering, leverage, size, or contagion effects: the fluctuations of Asian real
estate investment trust returns. J. Asian Econ. 27, 18–32 (2013)
J.J. Tseng, S.P. Li, Quantifying volatility clustering in financial time series. Int. Rev. Financial
Anal. 23, 11–19 (2012)
D. Wang, B. Podobnik, D. Horvatić, H.E. Stanley, Quantifying and modeling long-range cross
correlations in multiple time series with applications to world stock indices. Phys. Rev. E.
Statistical Nonlinear & Soft Matter Physics, 83(2), 046121 (2011a)
References 221
Y. Wang, Y. Wei, C. Wu, Detrended fluctuation analysis on spot and futures markets of West
Texas Intermediate crude oil. Phys. A 390, 864–875 (2011b)
W.J. Xie, Z.Q. Jiang, G.F. Gu, X. Xiong, W.X. Zhou, Joint multifractal analysis based on the
partition function approach: analytical analysis, numerical simulation and empirical applica-
tion. New J. Phys. 17, 27–30 (2015)
Y. Yin, P.J. Shang, Modified DFA and DCCA approach for quantifying the multiscale correlation
structure of financial markets. Phys. A 392, 6442–6457 (2013)
G. Zachmann, C. Von Hirschhausen, First evidence of asymmetric cost pass-through of EU
emissions allowances: examining wholesale electricity prices in Germany. Econ. Lett. 99,
465–469 (2008)
Y.F. Zhang, Z.T. Shi, Y.G. Chen, The comparative study of effectiveness between Hongkong
stock market and the mainland stock market. J. Financ. Res. 312, 33–40 (2006). (in Chinese)
C. Zhang, Z.W. Ni, L.P. Ni, Multifractal detrended cross-correlation analysis between PM2.5 and
meteorological factors. Phys. A 438, 114–123 (2015)
Y.J. Zhang, Y.M. Wei, An overview of current research on EU ETS: evidence from its operating
mechanism and economic effect. Appl. Energy 87(6), 1804–1814 (2010)
X.J. Zhao, P.J. Shang, A.-J. Lin, G. Chen, Multifractal Fourier detrended cross-correlation analysis
of traffic signals. Phys. A 390, 3670–3678 (2011)
W.X. Zhou, Multifractal detrended cross-correlation analysis for two non-stationary signals. Phys.
Rev. E 77, 06621 (2008)
W.X. Zhou, The components of empirical multifractality in financial returns. EPL (Europhys.
Lett.) 88(2), 28004 (2009)
J. Zhou, Multiscale analysis of international linkages of REIT returns and volatilities. J. Real
Estate Finance Econ. 45, 1062–1087 (2012a)
X. Zhuang, Y. Wei, B. Zhang, Multifractal detrended cross-correlation analysis of carbon and
crude oil markets. Phys. A 399, 113–125 (2014)
Chapter 10
Risk Analysis Based on Multifractal
Detrended Method
10.1 Asymmetric MF-DCCA Method Based on Risk

Conduction and Its Application
In recent years, the economic exchanges between countries have become frequent
and close because of the acceleration of finance unity accompanied by the economic
globalization. As a barometer of the national economy, the stock index has grad-
ually presented the linkage. Although a certain gap between developed countries
and financial system needs to be addressed, the influence of China, India, and Brazil
as representatives of the emerging countries has been increasing and their stock
markets have been presenting certain linkage characteristics with developed
countries. However, economic linkage will also lead to financial risk. Boyer et al.
(2006) concluded that local financial risks can be expanded or diffused because of
the existence of the risk conduction phenomenon. Moreover, the risk of transmis-
sion exists not only in each financial market in one country, but also in the markets
of different countries. The Asian financial crisis of 1997 and the 2008 subprime
mortgage crisis are two typical examples. Another fact that cannot be ignored is the
absence of a unified paradigm for the mutual effects of the financial markets that
reflect economic development. For example, affected by many factors, such as
economic and political positions, the financial impact of the US on the Chinese
financial markets is obviously bigger than its impact on the US financial markets.
Furthermore, market movements are often stronger during the rise and fall asso-
ciated with global financial turmoil, and the stock market of developing countries
appear relatively independent of the developed countries during generally stable
periods. This phenomenon has something to do with the rapid economic devel-
opment of developing countries. Therefore, the risk transmission of domestic and
foreign financial markets may have asymmetric characteristics.
The study of risk conduction between the stock markets at home and abroad
started early. The results and methods of research have been relatively rich.
Scholars in the early times mainly used the calculation of the linear correlation

224 10 Risk Analysis Based on Multifractal Detrended Method
coefficient of the stock markets to analyze the risk of conduction effect. When using
linear correlation to study the risk of transmission relationship between the markets,
the changes in different periods are analyzed. For example, Baig and Goldfajn
(1999) studied the risk conduction effect on the stock markets of Thailand,
Malaysia, Indonesia, South Korea, and the Philippines. The empirical results show
the significant correlation among national stock markets during an economic crisis.
However, the correlation among the stock markets is not static and the method used
to depict the correlation gradually developed from static correlation coefficient into
time-varying correlation coefficient. Examples of the correlations are the dynamic
conditional correlation (multivariate) GARCH (DCC-MGARCH); the long memory
VAR-DCC-GARCH model, and multivariate fractionally integrated asymmetric
power ARCH (FIAPARCH) dynamic conditional correlation (DCC) approach. The
Granger causality test method proposed by Granger (1969) is another approach that
has been widely used to analyze the risk conduction effect (Lee and Yang 2014;
Chang et al. 2013). The traditional methods, such as the GARCH model, VAR
model, and Granger causality test, can detect the asymmetric correlation. However,
most of them depend on a specific model form or the threshold value. In recent
years, some scholars have proposed the nonlinear Granger causality test method
Baek and Brock (1992) based on the linear Granger causality test method, which
compensates for the limitation of the linear Granger causality test method.
Alzahrani et al. (2014) attempted to investigate the non-linear Granger causality
between the wavelet transformed spot and futures oil prices. Their findings con-
sistently indicate the bidirectional causality between the spot and futures oil markets
at different time scales and under non-linear causality assumptions. Fernandez
(2014) applied the nonlinear Granger causality test to examine the spillovers of the
US subprime crisis to Asian and European economies and to determine the extent of
the effect of the crisis on the currency and stock markets. The empirical results
indicate that volatility effects partly induce nonlinear causality.
In addition, a large number of studies on the financial market exist, particularly
empirical analyses of the long-range autocorrelation between stock and exchange
markets. The results demonstrate that the stock markets and other financial markets
have the nonlinear characteristics of the multifractal. Hurst index is one of the
commonly used methods to measure the long-range correlation of the financial time
series. Numerous methods to estimate the Hurst value of the time series are
available as long as the fractal exists.
Alvarez-Ramirez et al. (2009) proposed the asymmetric detrended fluctuation
analysis (A-DFA) as an extension of the DFA method to detect the asymmetry
scaling behavior of time series. Cao et al. (2013) employed the A-DFA method to
explore the asymmetric multiple scaling behavior of the Chinese stock market, and
the empirical results show that the degree of multifractality of the rising trend in the
Chinese stock market is stronger than the downward trend. Moreover, the corre-
lation of the asymmetry is obvious in the large fluctuation and the occurrence of
major events will increase the asymmetry of the stock market. Subsequently, the
MF-ADCCA method was proposed by Cao et al. to study the asymmetric cross
correlation in the Chinese RMB exchange rate market, Chinese stock markets, and
10.1 Asymmetric MF-DCCA Method Based on Risk Conduction … 225
foreign exchange rate market. However, the asymmetric transmission direction is

not involved. By contrast, no correlation study on the risk conduction effect
between China and the domestic and foreign stock markets exists in the current
literature. This is surprising considering that the MF-DCCA and its derivative
methods are only able to describe the long-term correlation and multifractal char-
acteristics of two non-stationary time series, which is insufficient to analyze the
correlation of conduction direction.
Therefore, we construct an asymmetric MF-DCCA method based on the dif-
ferent directions of risk conduction (DMF-ADCCA) and by using the traditional
MF-DCCA. We employ a nonparametric model, which is not dependent on the
model, to study the risk conduction effect on the stock markets of the Chinese as
well as those of developed and developing countries to ensure that the empirical
results are more objective and robust.
The acceleration of economic globalization gradually shows the linkage of the
stock markets in various counties and produces a risk conduction effect. An
asymmetric MF-DCCA method is conducted based on the different directions of
risk conduction (DMF-ADCCA) and by using the traditional MF-DCCA. To ensure
that the empirical results are more objective and robust, this study selects the stock
index data of China, the US, Germany, India, and Brazil from January 2011 to
September 2014 using the asymmetric MF-DCCA method based on different risk
conduction effects and nonlinear Granger causality tests to study the asymmetric
cross-correlation between domestic and foreign stock markets. Empirical results
indicate the existence of a bidirectional conduction effect between domestic and
foreign stock markets, and the greater influence degree from foreign countries to
domestic market compared with that from the domestic market to foreign countries.
10.1.1 Asymmetric MF-DCCA Method Based on Different

Risk Conduction
Based on the traditional MF-DCCA method, we construct an asymmetric

MF-DCCA method based on the different risk conduction directions in this paper.
The specific steps are as follows:
Considering two time series fxð1Þ ðtÞg and fxð2Þ ðtÞg of the same length N, where
m ¼ 1; 2; . . .N Dt, and assuming that time series fxð2Þ ðtÞg lag Dt and fxð1Þ ðtÞg do
not lag, the profile is changed into:
X
m
yð1Þ ðmÞ ¼ ðxð1Þ ðtÞ xð1Þ Þ ð10:1Þ
t¼1
X
m
yð2Þ ðmÞ ¼ ðxð2Þ ðt þ DtÞ xð2Þ ðt þ DtÞÞ ð10:2Þ
t¼1
where m ¼ 1; 2; . . .; N Dt; the remaining steps are the as steps 2 to 5 of the

method, which have been introduced in the second part of the literature. Then, we
discuss the influence of fxð1Þ ðtÞg on fxð2Þ ðtÞg. However, this approach ignores the
autocorrelation of time series fxð2Þ ðtÞg. Although the profile can effectively alleviate
the autocorrelation effect after eliminating the trend, we make further improvements
to attain high rigorousness.
We believe that adding a noise during the construction of the profile yð2Þ ðmÞ can
weaken the influence of the autocorrelation of the time series fxð2Þ ðtÞg to a certain
degree.
Therefore, we imitate the Granger causality test ideas based on earlier studies
and decide to combine the time series fxð1Þ ðt þ Dt1 Þg (fxð1Þ ðtÞg lag Dt1 ), fxð2Þ ðtÞg
lagging Dt2 is fxð2Þ ðt þ Dt2 Þg for convenience. However, Dt1 is not equal to Dt2
because it can change Dt1 to control the autocorrelation attenuation and gain of
fxð2Þ ðtÞg caused by the noise. Hence, the new profile is as follows:
X
m
yð1Þ ðmÞ ¼ ðxð1Þ ðtÞ xð1Þ Þ ð10:3Þ
t¼1
X
m
yð2Þ ðmÞ ¼ ðxð2Þ ðt þ Dt2 Þ xð2Þ ðt þ Dt2 Þ þ axð1Þ ðt þ Dt1 Þ axð1Þ ðt þ Dt1 ÞÞ;
t¼1
ð10:4Þ
where Dt1 Dt2 ,m ¼ N Dt1 ; Dt1 \Dt2 , m ¼ N Dt2 .

We introduce parameter a before fxð1Þ ðt þ Dt1 Þg and the value of a must be
small enough so as not to affect the general characteristic of fxð2Þ ðtÞg. Below is the
reference to the discrete signal cross-correlation function (assuming the two time
series fXt g, fYt g).
8
< 1 P xn þ m y m 0
> Nm1
b xy;biased ðmÞ ¼ N n¼0

R ð10:5Þ
>
: 1 b y
N R x ð m Þ m\0
such that,
8 Pm ð2Þ
ðx ðt þ DtÞÞ2
> Pm ð1Þ
< a ¼ ðNDt Þt¼1 Dt1 Dt2 ; m ¼ N Dt1
ðx ðt þ DtÞÞ2
P1 m t¼1 ð10:6Þ
>
:a ¼
ð2Þ
ðx ðt þ DtÞÞ2
Pm ð1Þ
t¼1
2 Dt1 Dt2 ; m ¼ N Dt2
ðNDt2 Þ t¼1
ðx ðt þ DtÞÞ
The profile yð2Þ ðmÞ emphatically diminishes the effects of the auto-correlation
brought by fxð2Þ ðtÞg because the value of a is small. The other steps are the same as
steps 2 to 5 of the method, which have been introduced in the second part of
literature (Chen et al. 2008). In this paper, the Hurst exponent H is defined as in
Chen et al. (2008). When q ¼ 2, the Hurst exponent is marked with Hð2Þ.
Therefore, we can check the influence of fxð1Þ ðtÞg on fxð2Þ ðtÞg when fxð2Þ ðtÞg
lags. On the contrary, we can check out the influence of fxð2Þ ðtÞg on fxð1Þ ðtÞg when
fxð1Þ ðtÞg lags.
10.1.2 Nonlinear Causality Testing

fXt g and fYt g are two time series that are assumed to be strictly stationary and
weakly dependent, where t ¼ 1; 2: The Baek and Brock (1992) nonlinearity
Granger causality testing approach can be described as follows:
The m-length lead vector of fXt g is denoted as Xtm and the a-length and b-length
a b
lag vectors of fXt g and fYt g are denoted as Xta and Xtb , respectively.
Xtm ðXt ; Xt þ 1 ; Xt þ 2 ; . . .; Xt þ m1 Þ; ð10:7Þ
Ytm ðYt ; Yt þ 1 ; Yt þ 2 ; . . .; Yt þ m1 Þ; ð10:8Þ

a
Xta ðXt ; Xt þ 1 ; Xt þ 2 ; . . .; Xta Þ; ð10:9Þ
b
Xtb ðXt ; Xt þ 1 ; Xt þ 2 ; . . .; Xtb Þ ð10:10Þ
For the given values of m, a and b 1 and e [ 0, the stock returns fYt g fail the
nonlinearity Granger cause fXt g if:
a a
PrðXtm Xsm \eXta a
Xsa \e; Yta a
Ysa \eÞ
m a ð10:11Þ
m
¼ Prð X X \e X X a
\eÞ;
t s ta sa
where PrðÞ denotes the probability and kk denotes the maximum norm. The
probability on the left-hand side of Eq. (10.10) is the conditional probability that
two arbitrary m-length lead vectors of fXt g are within a distance e of each other,
given that their corresponding a-length lag vectors of fXt g are at a distance shorter
than the e of each other.
To test whether the null hypothesis H0 : fYt g does not satisfy the nonlinearity
Granger cause fXt g, expressing the conditional probabilities in terms of the cor-
ðm þ a;b;eÞ ðm þ a;eÞ
responding ratios of joint probabilities is useful. Let C1C2ða;b;eÞ and C3C4ða;eÞ denote
the ratios of the joint probabilities corresponding to the left-hand and right-hand
sides of Eq. (10.10). These joint probabilities are defined as:
mþa
b b
C1ðm þ a; b; eÞ PrðXta m þ a
Xsa \e; Ytb Xsb \eÞ;
a
b b
C2ða; b; eÞ PrðXta Xsaa
\e; Ytb Ysb \eÞ; ð10:12Þ
mþa
m þ a
C3ðm þ a; eÞ Prð Xta Xsa \e;

C4ða; eÞ PrðX a X a \eÞ:
ta sa
For the given values of m, a and b 1 and e > 0, the null hypothesis H0 can
then be expressed as:
C1ðm þ a; b; eÞ C3ðm þ a; eÞ
H0 : ¼ : ð10:13Þ
C2ða; b; eÞ C4ða; eÞ
Let I ¼ ðZ1 ; Z2 ; eÞ denote a kernel that is equal to 1 when two conformable

vectors Z1 and Z2 are within the maximum-norm distance e of each other and 0
otherwise.

1; kZ1 Z2 k e
I ¼ ðZ1 ; Z2 ; eÞ ¼ ð10:14Þ
0; kZ1 Z2 k [ e
The correlation–integral estimators of the joint probabilities in Eq. (10.11) can

then be written as:
2 XX b b
þa mþa
C1ðm þ a; b; e; kÞ Iðxmta ;xsa ; eÞ Iðytb ; ysb ; eÞ;
kðk 1Þ t\s
2 XX
C2ða; b; e; kÞ Iðxata ;xasa ; eÞ Iðybtb ; ybsb ; eÞ;
kðk 1Þ t\s
XX ð10:15Þ
2 þa mþb
C3ðm þ a; e; kÞ Iðxmta ;xsb ; eÞ;
kðk 1Þ t\s
2 XX
C4ða; e; kÞ Iðxata ;xbsb ; eÞ:
kðk 1Þ t\s
where t, s ¼ maxða; bÞ þ 1; . . .; T m þ 1, n ¼ T maxða; bÞ m þ 1.

For the given values of m 1, Lx 1, Ly 1, and e [ 0. Using the joint
probability estimators in Eq. (10.16), the test statistics is defined as:
a
C1ðm þ a; b; eÞ C3ðm þ a; eÞ
n1=2 Nð0; r2 ðm; a; b; eÞÞ: ð10:16Þ
C2ða; b; eÞ C4ða; eÞ
Here, the variance is consistently estimated (Verdelhos et al. 2014). Given that
the values of m, a and b 1 and e [ 0, if fXt g does not satisfy the nonlinearity
Granger cause fYt g, then the test statistic in Eq. (10.15) distributes asymptotically
as a normal distribution with a zero mean and a constant variance. The conventional
critical values apply when they are adopted to test the null hypothesis that stock
price fXt g does not nonlinearly Granger-cause fYt g because the test statistics are
asymptotically normal. A similar procedure is followed to test the hypothesis that
fYt g does not nonlinearly Granger-cause fXt g. For the values of r2 , Hiemstra and
Jones (1994) provided a very complex formula. For simplicity, e ¼ 1:5r and m ¼ 1
are cited in the paper based on a related study (Verdelhos et al. 2014).
10.1.3 Sample Selection and Descriptive Statistics Analysis
We choose the US and Germany as the representatives of developed countries, and

India and Brazil as the representatives of emerging markets to conduct a compre-
hensive and multi-angle research of the cross correlation between the domestic and
foreign stock markets. Therefore, the closing price is used as the sample selection
that would be able to represent the national stock markets, namely, Shanghai
Composite Index (SSCI), the Standard & Poor’s 500 index (S&P 500), Frankfurt,
Germany (DAX), Bombay sensitive 30 index (BSESN), and Brazil Purvis Palmer
index (BVSP). The Chinese accession to the world trade organization and pro-
motion of the link with the economic development of countries in the world started
in November 10, 2001. The sample interval selected is January 1, 2002 to
September 26, 2014. We obtain 2895 pairs of price data by eliminating the stock
markets session overlaps. The sample data is from the Yahoo Finance website
(www.finance.yahoo.com/).
We focus on the logarithmic price returns rt , that is, rt ¼ logðPt Þ logðPt1 Þ. Pt
represents the closing price of day t. Therefore, for the return series, the empirical
data number is 2858 (denoted by rSSCI, rS&P500, rDAX, rBSESN, rBVSP) and
the descriptive statistical results are shown in Table 10.1.
Table 10.1 Descriptive statistics for the return series

Mean Max Min S.D. Ske Kur J–B
rSSCI 5.40e−05 0.039235 −0.055431 0.007326 −0.193353 7.640787 2582.494***
rS&P500 8.22 e−05 0.047586 −0.041126 0.005783 −0.132180 12.70689 11,228.79***
rDAX 9.24e−05 0.046893 −0.042683 0.007146 −0.021459 8.005565 2983.933***
rBSESN 0.000319 0.069444 −0.055572 0.007045 −0.152248 11.97696 9607.466***
rBVSP 0.000215 0.059397 −0.052532 0.008209 −0.131032 7.672055 2607.541***
Note ***denotes 1% significance levels; symbols “Max”, “Min’’, “S.D.”, “Ske”, “Kur” denote maximum, minimum,
standard deviation, skewness and kurtosis respectively. The “J–B” denotes Jarque–Bera statistics testing for the
normality assumption
Table 10.1 shows that for the five return series, the skewness is different from 0
and the kurtosis is different from 3, which indicates departures from normality.
Moreover, the Jarque–Bera statistics are significant at the 1% significance level.
Thus, the normality assumption of returns can be rejected; thus, the series have the
typical characteristics of long-term memory.
10.1.4 Simulation Analysis
To compare the performance of DMF-ADCCA in the detection of cross-correlation

between two non-stationary time series, we further determine whether the calcu-
lated values of Hurst exponent H fall within its 95% confidence interval. To
investigate the influence of the periodic trend, we employ a periodic
two-component fractionally autoregressive integrated moving average (ARFIMA)
process.
Table 10.2 shows some artificial time series with the length of N (2000 or
10,000 data points) that we simulated based on the ARFIMA process.
As shown in Table 10.2, with the increase of the lag order, the values of Hurst
exponent HDMFADCCA obtained by DMF-ADCCA method based on the simulated
series shows a fluctuating trend, and the volatility trend of the value ranges from lag
10 to lag 50. On the one hand, if the length of the time series is short,
DMF-ADCCA for lag = 10, lag = 30, and lag = 50 outperform the DMF-ADCCA
for lag = 1 and lag = 5. Currently, the cross-correlation exponent estimated by the
DMF-ADCCA method is more accurate. For example, when T = 1000 and
L = 2000, HDMFADCCA = 0.8904 (lag = 10), 0.8906 (lag = 30), and 0.8905
(lag = 50). From the 95% confidence interval, the numerical values of the
Table 10.2 The simulation results of DMF-ADCCA method (q1 ¼ q2 ¼ 0:4 and W ¼ 0:5)
L lag No trend T = 1000
HDMF-ADCCA 95% confidence HDMF-ADCCA 95% confidence
interval interval
2000 1 0.8395(0.0116) (0.8133,0.8658) 0.8806(0.0194) (0.8367,0.9245)
5 0.8400(0.0118) (0.8137,0.86620) 0.8894(0.0159) (0.8533,0.9254)
10 0.8429(0.0153) (0.8082,0.8776) 0.8904(0.0155) (0.8554,0.9253)
30 0.8442(0.0173) (0.8050,0.8833) 0.8906(0.0154) (0.8557,0.9255)
50 0.8430(0.0144) (0.8103,0.8757) 0.8905(0.0158) (0.8547,0.9262)
10,000 1 0.8364(0.0107) (0.8121,0.8608) 0.9472(0.0646) (0.8009,1.0934)
5 0.8412(0.0123) (0.8133,0.8691) 0.9476(0.0648) (0.8015,1.0938)
10 0.8417(0.0123) (0.8138,0.8695) 0.9536(0.0591) (0.8198,1.0874)
30 0.8457(0.0118) (0.8191,0.8723) 0.9570(0.0620) (0.8167,1.0972)
50 0.8424(0.0119) (0.8154,0.8694) 0.9567(0.0622) (0.8163,1.0970)
Note Values in parentheses are the mean of standard errors
simulation fall within the confidence interval. We can deduce that the
cross-correlations are showing upward fluctuations in a certain lag period (ap-
proximately 10 days).
On the other hand, if the length of the time series is long, the DMF-ADCCA
values for lag = 10, lag = 30, and lag = 50 are slightly overestimated compared
with those of the DMF-ADCCA for lag = 1 and lag = 5. For example, when
T = 1000 and L = 10,000, HDMFADCCA = 0.9536 (for lag = 10),
HDMFADCCA = 0.9570 (for lag = 30), and HDMFADCCA = 0.9567 (for lag = 50),
whereas the theoretical value is only 0.9 (see Table 10.2). The existence of
crossovers, which leads to the slight overestimation of the cross-correlation expo-
nents, could be the reason. In the 95% confidence interval, the numerical values of
the simulation fall within the confidence interval. Moreover, the cross-correlations
are showing upward fluctuations at a certain lag period (approximately 5 days).
In summary, if the time series is short, the DMF-ADCCA values for lag = 10,
lag = 30, and lag = 50 may be the better choice. However, if the time series is long,
the DMF-ADCCA values for lag = 1 and lag = 5 may slightly perform better than
those of lag = 10, lag = 30, and lag = 50. Moreover, since the empirical time series
are usually short, the DMF-ADCCA values for lag = 10, lag = 30, and lag = 50
may be more accurate because the estimations are closer to the theoretical values for
short time series.
10.1.5 Empirical Analysis of Asymmetric MF-DCCA

Method
Using the combined with the asymmetric MF-DCCA method based on different risk
conduction directions proposed in the second part, we discuss the influence of
synchronization to the cross-correlation between domestic and foreign stock mar-
kets and attempt to investigate the risk conduction direction characteristics between
them.
If the second part of the model is constructed as:
m
X
yð2Þ ðmÞ ¼ axð1Þ ðt þ Dt1 Þ þ xð2Þ ðt þ Dt2 Þ axð1Þ ðt þ Dt1 Þ xð2Þ ðt þ Dt2 Þ ;
t¼1
ð10:17Þ
where fxð1Þ ðtÞg and fxð2Þ ðtÞg have a completely linear relationship, then, making
Dt1 equal to Dt2 is unnecessary because abating the correlation to fxð2Þ ðtÞg causes
instability. However, previous literature suggests that Chinese and foreign stock
markets also show a nonlinear relationship. For convenience, the lag time Dt1
equals Dt2 , with values from 1 to 50 (50 days of trading time is approximately two
months).
When q = 2, rSSCI * rS&P500, rSSCI * rDAX, rSSCI * rBSESN,

rSSCI * rBVSP, the former lags and the latter does not lag, and the Hurst
exponents are marked with HDrSSCI rS&P500 ð2Þ,HDrSSCI rDAX ð2Þ, HDrSSCI rBSESN ð2Þ, and
HDrSSCI rBVSP ð2Þ, respectively. Figures 10.1, 10.2, 10.3, and 10.4 describe the eight
types of subscript indexes for different lag time variations.
Figures 10.1, 10.2, 10.3, and 10.4 show that either the “domestic stock market
lags and the foreign stock market does not lag” or “foreign stock market lags and
domestic stock market does not lag.” The degree of sustainability of
cross-correlations (or long-range correlations/persistence) are showing upward
fluctuations at a certain lag period (see Figs. 10.1, 10.2, 10.3, and 10.4 marked
lines, approximately 15 days). If the lag period of one market is more than 15 days,
the degree of sustainability of cross-correlations among the markets will be
reduced. This observation suggests that a bidirectional risk conduction effect exists
between domestic and foreign stock markets and the risk conduction effect will not
last for a long time but for only 15 days in general. Moreover, the influence is
relatively slight for the lag period beyond 15 days (Fig. 10.5).
Fig. 10.1 HDrSSCI rS&P500 ð2Þ and HrSSCI rS&P500 ð2Þ for different lag time variations
Fig. 10.2 HDrSSCI rDAX ð2Þ and HrSSCI rDAX ð2Þ for different lag time variations
Fig. 10.3 HDrSSCI rBSESN ð2Þ and HrSSCI rBSESN ð2Þ for different lag time variations (color figure
online)
Fig. 10.4 HDrSSCI rBVSP ð2Þ and HrSSCI rBVSP ð2Þ for different lag time variations (color figure online)
Fig. 10.5 H DT ð2Þ for different lag time variations

Based on the existence of the bidirectional risk conduction effect between the
domestic and foreign stock markets, we further investigate which market conduc-
tion effect is stronger. Therefore, we calculate the difference between the domestic
and foreign stock market lagging scale indexes as follows (mark as H DT ð2Þ):
HrDT
SSCI rS&P500
ð2Þ ¼ HDrSSCI rS&P500 ð2Þ HrSSCI DrS&P500 ð2Þ ð10:18Þ
HrDT
SSCI rDAX
ð2Þ ¼ HDrSSCI rDAX ð2Þ HrSSCI DrDAX ð2Þ ð10:19Þ
HrDT
SSCI rBSESN
ð2Þ ¼ HDrSSCI rBSESN ð2Þ HrSSCI DrBSESN ð2Þ ð10:20Þ
HrDT
SSCI rBVSP
ð2Þ ¼ HDrSSCI rBVSP ð2Þ HrSSCI DrBVSP ð2Þ; ð10:21Þ
where DT ¼ Dt1 ¼ Dt2 ¼ 1; 2; 3; . . .; 50.

The values of H DT ð2Þ are greater than zero, suggesting that when the domestic
stock market lags, the influence of both developed and developing countries on the
stock market to the domestic stock market is greater than the that of the domestic
stock market on foreign stock markets. On the contrary, when the foreign stock
market lags, we can conclude the opposite. Developed countries, such as the US
and Germany, and their capital markets are more perfect and have stronger influ-
ence on the world economy. Hence, their risk conduction effects are stronger.
However, the stronger risk conduction effect of India, Brazil, and other developing
countries over the Chinese stock market is difficult to understand. Although the
stock markets of India and Brazil are less developed, an observation that is easily
overlooked by people, they are more mature than the Chinese stock market. India’s
Mumbai Stock Exchange, the oldest stock market in Asia, which was opened in
1875 by the British, was put on the right track in 1957. All stock trading processes
became online in 1995; in 2001 derivates were put out. Moreover, India’s private
listed companies are more mature and the PE ratio is relatively low. The devel-
opment patterns of India pay more attention to consumption instead of investment,
domestic demand rather than exports, service industry rather than
non-manufacturing industry, and high and new technology industries rather than
labor intensive industry. Therefore, its economic characteristics also have a positive
effect on the healthy stock market. Sao Paulo Stock Market is Latin America’s
biggest stock market, undertaking 90% of new and additional issuance of shares of
Latin America’s capital markets. The Bovis Palmer index selected in this paper is
composed of the 63 largest Sao Paulo stocks. The largest weight in the stock market
of Brazil, as a resource-intensive country, is the resource companies, which are
robust. In addition, the openness of Brazil’s stock market is high and the QFII in
2009 accounted for 30% of its stock market capitalization. However, the economic
fundamentals of the Chinese stock market do not have the advantages of the third
industry of India and have no rich resources of emerging markets such as Brazil. As
a result, the Chinese stock market is more influenced by these countries.
10.1.6 Nonlinear Granger Causality Test Based

on the Remover of Long Memory
We employ the nonlinear causality test method proposed by Baek–Brock to define

the causal relationship between domestic and foreign stock markets. We compare
the results using the multifractal method. This method has been widely adopted by
scholars after the modification of Hiemstra and Jones. To avoid the autoregressive
phenomenon, the time series for the Grainger causality test must be stationary.
Thus, we first use the augmented Dickey–Fuller test (ADF test) for the unit root test
for all stock market returns series, where the optimal lag period is generated based
on the Akaike and Schwarz information criteria. The results are shown in
Table 10.3.
These results indicate that both the domestic and foreign stock market return
series have the intercept term, but not the trend. The optimal lag period is 27. The
values of t-Statistics are less than the critical value of the 1% significance level and
significant at the 1% significance level. Thus, the null hypothesis can be rejected,
which indicates that the series is stationary. The values of Hurst index and fractal
difference order for each return series are presented in Table 10.4.
Table 10.4 shows that the domestic and foreign stock markets return series have
long-term memory characteristics, and that the existence of long-term memory
characteristics may affect the results of the nonlinear Granger causality test. Hurst
(1951) put forward the rescaled range analysis (R/S analysis) to study the statistical
characteristics of long-term memory time series. We can determine the long-term
memory features by computing the Hurst index. If the scaling exponent H(2) > 0.5,
the correlations of the time series are long-range persistent (positive). If the scaling
exponent H(2) < 0.5, the correlations of the time series are anti-persistent
Table 10.3 ADF test Series Lag period t-statistic 1% critical value
rSSCI 27 (c, 0) −53.81202*** −3.432448
rS&P500 27 (c, 0) −59.97664*** −3.432448
rDAX 27 (c, 0) −54.87766*** −3.432448
rBSESN 27 (c, 0) −51.33331*** −3.432448
rBVSP 27 (c, 0) −40.47894*** −3.432448
Note (c, 0) representation of time series have no intercept trend.
The null assumption is that the time series exist unit root.
***Denotes the null assumption is rejected at 1% significance
level
Table 10.4 Hurst exponent rSSCI rS&P500 rDAX rBSESN rBVSP

H(2) and fractional difference
order d H(2) 0.5867 0.5298 0.5570 0.5246 0.5132
d 0.0867 0.0298 0.0570 0.0246 0.0132
(negative). If the scaling exponent H(2) = 0.5, the time series display a random
walk behavior and the market is weakly efficient.
The following is based on the R/S analysis method and fractional order differ-
ence algorithm employed to study the nonlinear Granger causality between the
stock returns in the domestic and foreign stock markets after the elimination of
long-term memory. According to Ellis et al. (2004), the fractional order difference
d can be solved by the Hurst index, that is, d ¼ Hð2Þ 0:5. We determine the
fractional order difference of each market’s return series, then conduct the nonlinear
Granger test to the series after the fractional difference. The test results are listed in
Table 10.5.
Table 10.5 shows that bidirectional nonlinear Granger causality relationships
exist between the Chinese and US stock markets and Indian and Brazilian stock
markets when a ¼ b ¼ 2. The nonlinear relationship gradually becomes insignifi-
cant with the increase of the lag order, but the empirical results indicate that the
Chinese and US stock markets and the Indian and Brazilian stock markets do have
an interaction effect relationship. This is consistent with the results of an earlier
study, which showed that along with the deepening of the financial globalization
process, the dynamic correlation coefficient of the Chinese stock market with the
rest of the countries has been enhanced significantly. At the same time, the study
explained that the Chinese stock market has changed from one with a closed state to
one with a certain degree of openness. The Chinese stock market is becoming more
and more connected to the world markets and its linkage trend with the foreign
stock market is becoming obvious. Hence, the volatility of the Chinese stock market
is affected by the external stock market and vice versa. This characteristic is also
exhibited by the stock market risk conduction.
However, our finding suggests that unidirectional Granger causality relation-
ships exist between the Chinese and German stock markets during the study period
(a ¼ b ¼ 1; 2; . . .; 5). On the one hand, the Chinese stock market is still an
emerging market to a certain extent; thus, the international financial risks easily
enter the Chinese stock market through the trade and financial market channels.
This is consistent with the literature (Xue et al. 2015) that indicates that the German
stock market does Granger-cause the Chinese stock market, but the Chinese stock
market does not Granger-cause the German stock market. On the other hand,
Chinese stock market returns are strongly influenced by all kinds of uncertain
information; thus, predicting the mean of the market return through the returns of
the other markets is difficult. However, this does not mean that the two markets are
independent.
The above conclusions show that the influence of the stock markets of developed
countries on those of developing countries remains very strong. Fluctuations in the
US and German stock markets will basically cause significant changes in the
Chinese stock markets. Moreover, an imbalance effect on the stock markets exists
between developing and developed countries. For example, the Chinese stock
market has a significant volatility effect on the German stock market, but the
influence of the Chinese stock market on the German stock market is not
significant.
10.1
Table 10.5 Nonlinear Granger causality test based on the remover of long memory
a¼b rSSCI ! rS&P500 rS&P500 ! rSSCI rSSCI ! rDAX rDAX ! rSSCI rSSCI ! rBSESN rBSESN ! rSSCI rSSCI ! rBVSP rBVSP ! rSSCI
1 2.962*** 3.147*** 1.323 1.643** 4.740*** 4.776*** 2.444*** 2.166**
*** *** *** *** *** ***
2 3.662 3.939 0.680 2.734 4.724 4.888 2.651 3.097***
* *** *** *** ***
3 1.628 4.106 0.100 2.671 4.386 3.750 1.022 2.884***
*** *** *** ***
4 0.582 3.001 −0.380 2.807 3.150 2.806 −0.092 1.268
5 0.369 2.474*** −0.520 3.038*** 1.916** 1.332 −0.073 0.667
Note a ¼ b denotes the residual series of lag order number; rSSCI ! rS&P500 means the original hypothesis: rSSCI isn’t a Granger causality of rS&P500, the others
have the same meaning. The values in the table are the t values and the corresponding level of significance,***, **, *denote 1, 5, 10% significance levels
respectively; e ¼ 1:5r, m = 1; to return series not eliminating long memory, we also do the lag of 1–5 order nonlinear granger causality test, its each lag order
significant levels is similar to those in Table 10.5
Asymmetric MF-DCCA Method Based on Risk Conduction …
237
10.1.7 Conclusions
We construct the asymmetric MF-DCCA method based on different risk conduction

methods to investigate the asymmetric cross-correlation between domestic and
foreign stock markets. Moreover, we employ the nonlinear Granger causality test
based on the removal of long-term memory to validate our study. The empirical
results are as follows:
(1) Bidirectional conduction effect exists between domestic and foreign stock
markets, and the degree of influence of foreign countries on domestic market is
greater than that of the domestic market on foreign countries.
(2) Bidirectional nonlinear Granger causality relationships exist between the
Chinese and foreign stock markets. In addition, only a unidirectional Granger
causality relationship exists between the Chinese and German stock markets
during the study period.
Although the asymmetric MF-DCCA method based on the risk conduction
direction and the nonlinear Granger causality test are two different methods, the
results obtained by these methods are generally consistent. We can conclude that
the Chinese stock market is more vulnerable to the effect of foreign stock markets,
especially during a period of decline. Furthermore, the Chinese stock market is
defective and unable to resist risk completely. We believe that the following aspects
of the Chinese stock market should be improved.
First, financial market supervision and the ability to combat market manipulation
should be enhanced. Although current laws and regulations on the supervision of
the stock market in China have been relatively perfect, law enforcement is relatively
lacking. Therefore, the relevant government departments must be strengthened in
terms of supervision to promote further standardization, transparency, fairness, and
soundness in financial market.
Second, investment knowledge and risk awareness of investors should be
improved. At the same time, investors should be led gradually from speculation to
investment and from short-term operation to long-term operation. Financial
knowledge via multiple channels should be promoted to enable investors to pay
attention to the linkage with foreign stock markets and to adjust the investment
strategy immediately when the foreign stock market is volatile.
The traditional asymmetric MF-DCCA method is unable to analyze the
cross-market risk conduction direction quantitatively. However, we always refer to
the judgment of the direction of risk conduction when analyzing the stocks, futures,
and other financial market risks. Thus, the asymmetric MF-DCCA based on risk
conduction direction method proposed in this paper has a wide application value in
terms of judging the risk conduction direction. In addition, the research on the
direction of the risk conduction is based on the sample interval, and the use of
“volatility constraint” on the overall study sample area between short intervals of
risk conduction direction judgment is the future direction of our research.
However, in this paper, we are just tried to construct an asymmetric MF-DCCA

method based on the different directions of risk conduction (DMF-ADCCA) and by
using the traditional MF-DCCA which has been widely used. It is possible to use
other MF-DCCA methods (such as MF-X-DMA etc.) to construct the similar
approach to empirical analysis of the asymmetric risk conduction effect on the
financial markets. The empirical results may more accuracy since differences exist
in the trend elimination method and the processing of the original data profile,
which may both have an impact on the cross-correlation exponent estimated by
these methods. It is worth studying further.
10.2 Extreme Values Evaluation Based on Detrended

Fluctuation Analysis
The stock market plays an important role as the barometer of national economy.
Therefore, the frequency and strength of extreme events in stock markets have a
huge effect on the economy and social life, attracting the attention of more scholars.
The reasonable forecasting of the fluctuation of stock price, the appearance of
extreme event forecasting, and the effective risk of fluctuation of stock price
forecasting have become the research focus of government regulators, financing
institutions, scholars, and investors.
Extreme events do not occur in the scope of normal systematic state of its own
evolution, but in the course of systematic evolution or the abnormal state caused by
the interference outside the system. In other words, the state of events seriously
deviates from the average behavior. Events that rarely occur can be called extreme
events in the statistical sense. Numerous studies focus on extreme events, and most
of them focus on the extreme events of the climate. For example, Karagiannidis
et al. (2012) used data from 280 stations in Europe to analyze the climatic char-
acteristic of rainfall in this region. They discovered that the extreme rainfall events
have a downward trend in most of the regions in Europe. Burgueño et al. (2014)
conducted a research on extreme climatic temperature in Catalonia, Spain.
Verdelhos et al. (2014) separated and classified the extreme wind events according
to wind speed.
The global financial crisis, such as the subprime crisis in the US and the debt
crisis in Europe, resulted in a world economy in turmoil. Stock risk is usually
measured using the VaR model. However, this model is only used to measure risk
in a normal market. To evaluate the risk of extreme events, Gnedenko (1943)
proposed the extreme value theory. Longin (1996) was the first to apply this theory
in the financial field. Bao and Le (2006) used different methods to calculate the VaR
values of four countries, and argued that this extreme value model is suitable for the
financial crisis. Some scholars regard the stock market as a complex system. They
studied the extreme events of stock markets from the dynamics angle and applied
the Omori rule (the relation between the times of aftershock exceeding the assigned
earthquake magnitude and time) from seismology to the large changes in stock
market. For example, Selcuk (2004) suggested that the Omori rule existed after a
major fall in the stock returns on 27 October 1996 and 31 August 1998 by studying
the daily data of 10 emerging stock markets, such as Argentina, Brazil, Hong Kong,
Indonesia, Korea, Mexico, Philippines, Singapore, Taiwan, and Turkey. Weber
et al. (2007) discovered that the Omori rule is not only applicable to extreme events
of the stock market such as the market crash and stock plunge, but also to the
evolution of some variables after the medium shock of the stock market based on
the research of Selcuk. Malevergne and Sornette (2001) studied the Standard &
Poor’s 500 index after 19 October 1987 (Black Monday) and discovered that its
invisible variance had a damping trend in power law index with periodic fluctuation.
Zawadowski et al. (2004) used 15 min of data to study share price, volatility, and bid
ask spread evolution of daily extreme events in NYSE and NASDAQ. Their results
showed that stock price reversed and appeared to peak in extreme events, and then
damped in the power rate. Siokis (2014) analyzed extreme economic events in
Europe from 2009, and regarded the index of stock market as the index of economic
activities. Then, they analyzed the function of government economic aid.
To define extreme events, several studies on climatic extreme events defined
extreme events by the threshold. If the value exceeds the threshold, the value can be
defined as extreme value and the event can be defined as extreme event. The
percentile method is employed in many studies to define the threshold. The
threshold exceeds some tercile in the percentile method. A value that exceeds the
threshold is considered extreme events. Bartholy and Pongrácz (2010) and Gemmer
et al. (2011) used 99, 97, 95 and 90% to define the threshold of rainfall in the
region. This method is influenced by human factors and does not consider the
characteristics of the evolution of the data or the system, such as scale invariance,
long-range correlations, and so on. The threshold which is obtained by the per-
centile method, appeared to be uncertain.
This paper focuses on the comparative analysis of extreme values in Chinese and
American stock markets based on the detrended fluctuation analysis
(DFA) algorithm using the daily data of Shanghai composite index and Dow Jones
Industrial Average. The empirical results indicate that the multifractal detended
fluctuation analysis (MF-DFA) method is more objective than the traditional per-
centile method. The range of extreme value of Dow Jones Industrial Average is
smaller than that of Shanghai composite index, and the extreme value of Dow Jones
Industrial Average is more time clustering. The extreme value of the Chinese or
American stock markets is concentrated in 2008, which is consistent with the
financial crisis in 2008. Moreover, we investigate whether extreme events affect the
cross-correlation between the Chinese and American stock markets using multi-
fractal detrended cross-correlation analysis algorithm. The results show that
extreme events have nothing to do with the cross-correlation between the Chinese
and American stock markets.
10.2 Extreme Values Evaluation Based on Detrended Fluctuation Analysis 241
10.2.1 Threshold Estimation Method
The Hurst exponent obtained by MF-DFA method assures the long-range corre-
lation of the systemic evolution, and the extreme events do not (or very little) affect
the long-range correlation of the entire system. Therefore, the DFA method con-
firms the thresholds of extreme events.
We consider time series xi , and n is the length of series. ði ¼ 1; 2; 3; . . .; N Þ. The
algorithm to measure threshold is as follows:
(1) Find the maximum xmax and minimum xmin in the series xi ;
(2) Calculate the midpoint (R) of xi ;
(3) Starting with xmax , remove the values in xi section fxi ; xi xmax d kg until
xi ¼ R to obtain a new series yj ðj ¼ xmax d kÞ, where d is the interval of the
section and k ¼ 1; 2; 3; . . .; ðxmax RÞ=d. Starting with xmin , remove the values
in xi section fxmin ; xi xmin þ d kg until xi ¼ R to obtain the new series
yj ðj ¼ xmin þ d kÞ, where d is the interval of the section and
k ¼ 1; 2; 3; . . .; ðR xmin Þ=d. The value of d represents the resolution ratio of
this method. The resolution ratio is higher as d is smaller. In this paper, we set
d ¼ 0:0001;
(4) Calculate the index of long-range correlation of every new time series yj ; that is,
the value of DFA is denoted as Dj ;
(5) When the changes of Dj tend to gently converge to the original values hq , j is
the threshold of xi at this time.
In this paper, the method can also calculate the threshold with the rise and
decline of the data. The threshold is acceptable in this manner.
In this part, we chose the daily closing prices of the Shanghai composite index
(SCI) and the Dow Jones Industrial Average (DJIA) as sample data. To guarantee
the timeliness of this paper, we chose a sample range from 4 January 2006 to 30
April 2014 because the Chinese reform in non-tradable shares became effective in
May 2005. Several factors affect the trading time of Chinese and American stock
markets such as statutory holidays and so on. We removed the non-overlapping data
between Chinese and American stock markets. We obtained 1957 groups data in the
final empirical analysis.
We used the logarithmic rate of return ½rt ¼ logðPt Þ logðPt1 Þ
as the data of
empirical analysis. Therefore, the returns of SCI and DJIA in t minute (r1;t and r2;t )
are given as follows:
r1;t ¼ logðSCIt SCIt1 Þ and r2;t ¼ logðDJIAt DJIAt1 Þ ð10:22Þ
Figure 10.6 shows the values of Dj in the original data, rising trend, and down
trend of SCI. The first inflection point in the positive section was j ¼ 0:024 on 16
January 2007 and its rate of return was 0.0242. The values of Dj were smooth and
steady with the values of DFA of the original data, suggesting that if j 0:024, then
the points removed contained few messages of the evolution of the entire system.
Removing those points would not affect the rule of evolution of the system itself.
The evolution rule of series yj has the same mechanism as the original series fxi g.
Hence, Dj is almost equal to the values of DFA of the original data. These points
can be regarded as extreme values of current series. Therefore, j ¼ 0:024 is con-
sidered as the positive threshold of the current series. Similarly, j ¼ 0:029 is
considered the negative threshold of the current series. Its rate of return on 30 May
2007 was −0.02919. As the second inflection point in Fig. 10.1, j ¼ 0:039; that is,
its rate of return on 19 September 2008 was 0.039235, Dj was almost equal to the
values of DFA of the original data and tends to be more smooth and steady. In other
words, j ¼ 0:039 was also a positive threshold of this series, but the frequency of
the extreme events was smaller than j ¼ 0:024. Similarly, j ¼ 0:04, that is, its rate
of return on 27 February 2007 was −0.0402, and the frequency of the extreme
events was also smaller than j ¼ 0:029.
Figure 10.7 shows the values of Dj of the original data, rising trend, and down
trend of DJIA change with j. Similarly, the first inflection point in the positive
section was j ¼ 0:0275, that is, its rate of return on 21 November 2008 was
1
Original data
0.95 Rising trend
Down trend
0.9
0.85
0.8
J=0.024
Dj
0.75 J=0.039
0.7 J=-0.029
J=-0.04
0.65
0.6
0.55
-0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
j
Fig. 10.6 The changes of values of DFA of series yj about the original data, rising trend and down
trend of SCI
1.1
Original data
Rising trend
1
Down trend
0.9
J=0.035
J=0.0275
0.8
Dj
0.7 J=-0.035
0.6 J=-0.0265
0.5
0.4
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
j
Fig. 10.7 The changes of values of DFA of series yj about the original data, rising trend and down
trend of DIJA
0.027524. The values of Dj tend to be smooth and steady with the values of DFA of
the original data, suggesting that when j 0:0275, the points removed contained
few messages of the evolution of the entire system. Removing those points do not
affect the rule of the evolution of the system itself. The evolution rule of series Yj
has the same mechanism as the original series fxi g. Hence, Dj is almost equal to the
values of DFA of the original data. These points are considered the extreme values
of the current series. Therefore, j ¼ 0:0275 is considered the positive threshold of
the current series. Similarly, j ¼ 0:0265 is considered the negative threshold of
the current series. Its rate of return on 27 October 2008 was −0.026553. In the
second inflection point in Fig. 10.2, j ¼ 0:035, that is, the rate of return on 14
October 2008 was 0.0355496. Dj was almost equal to the values of DFA of the
original data and tend to be smooth and steady. In other words, j ¼ 0:035 is also a
positive threshold of this series, but the frequency of the extreme events is smaller
than j ¼ 0:0275. Similarly, j ¼ 0:035, that is, the rate of return on 15 October
2008 was −0.03561, and the frequency of the extreme events was also smaller than
j ¼ 0:0265.
10.2.3 Simulation Analysis
To evaluate the accuracy of the MF-DFA method used to acquire the thresholds, we
compare MF-DFA method with the percentile method, which is adopted widely to
acquire thresholds. The percentile method adopts the method of probability

statistics and considers the rate of returns that exceeds some tercile as the extreme
values of the rate of returns. The method is as follows:
The daily rates of returns of Chinese and American stock markets are arranged in
ascending order from 4 January 2006 to 30 April 2014, and then x1 ; x2 ; . . .; xn is
obtained. The threshold of percentile method is expressed as
x ¼ ð1 aÞxj þ axj þ 1 ; ð10:23Þ
where j ¼ floorðpðn þ 1ÞÞ is the serial number in the ascending daily rate of returns
(floor is a rounding function in Matlab), p is the percentiles, and a ¼ pðn þ 1Þ j is
the weight coefficient. In this paper, p is selected as 0.95.
Tables 10.6 and 10.7 show the thresholds of SCI and DJIA estimated using
MF-DFA and percentile methods. Regardless of the thresholds of SCI or DJAI, the
positive thresholds obtained by the MF-DFA method is higher than that of per-
centile method, and the negative thresholds obtained by MF-DFA method is lower
than that of percentile method. Therefore, the range of extreme events obtained by
MF-DFA method is smaller than that of the percentile method. In addition, the
frequency of percentile method exceeds 10 and different from the frequency of
extreme events that occur in real life.
Moreover, we compared the MF-DFA method with the percentile method
combined with real life extreme events. Figures 10.3 and 10.4 represent the date of
the extreme events of SCI and DJAI, respectively. The numbers 1–100 in the
abscissa represent January 2006 to April 2014 and the numbers in the ordinate
represent the days. The black points in the graph represent the dates of extreme
events. The points with the blue box showed the dates of extreme events, which
exceeded the first inflection point. The points with the red box showed the dates of
extreme events, which exceeded the second inflection point.
Table 10.6 Threshold values estimated by MF-DFA and percentile method of SCI
MF-DFA method Percentile method (95%)
Threshold Frequency Threshold Frequency
Positive daily rate of returns 0.0245 10
Negative daily rate of returns −0.029 10
Positive daily rate of returns 0.039 2 0.0115736304 97
Negative daily rate of returns −0.04 2 −0.0128998139 97
Table 10.7 Threshold values estimated by MF-DFA method and percentile method of DJAI
MF-DFA method Percentile method(95%)
Threshold Frequency Threshold Frequency
Positive daily rate of returns 0.0275 5
Negative daily rate of returns −0.0265 5
Positive daily rate of returns 0.035 2 0.0081132553 97
Negative daily rate of returns −0.035 2 −0.0086499126 97
35
30
25
20
Day
15
10
0
0 10 20 30 40 50 60 70 80 90 100
Month
Fig. 10.8 The dates of extreme events of SCI (color figure online)
Figure 10.8 shows that the black points are dispersed. Scattered distribution is
sparse, but the points with the blue box are relatively concentrated and focused on
2008. The space between the two red dash lines indicates the year 2008. The
subprime crisis in America occurred in 2008. This crisis shocked the Chinese stock
market. The number of points with the red box is limited, indicating that they only
represent some special situations. Therefore, we adopt the first inflection point as
the threshold using MF-DCCA to study the correlation.
Similarly, Fig. 10.9 shows that the black points have scattered distribution, and
the points with the blue box are also focused on August and October 2008, indi-
cating a period when the share prices of two house loan magnates in the USA
slumped and resulted in the subprime crisis. The space between the two red dash
lines indicates the period from August to October 2008. The points with the red box
validated the fact further. However, the number of points with the red box is
35
30
25
20
Day
15
10
0
0 10 20 30 40 50 60 70 80 90 100
Month
Fig. 10.9 The dates of extreme events of DJAI (color figure online)
limited. We choose the first inflection point as the threshold in the subsequent
analysis.
10.2.4 Time-Clustering of Extreme Events
To study the time-clustering of extreme events of SCI and DJAI, we adopt the
coefficient of variation to measure. Assume Vi ¼ jPt þ 1 Pt j (Du et al. 2013), for
Pt ; Pt þ 1 , which represent the closing price at t and t þ 1 moments. Similarly,
si ¼ ti þ 1 ti for ti þ 1 ; ti implied that the time of events arranged by serial number i.
Inter-event time is denoted by s. We then define the coefficient of variation as
follows:
rs
Cv ¼ ; ð10:24Þ
h si
where rs is the standard deviation of the events Vi and hsi is the mean inter-event
time. Poissonian process has Cv ¼ 1, but a clusterized process is characterized by
Cv [ 1.
Table 10.8 shows that Cv of SCI obtained by MF-DFA and the percentile
method is greater than 1, and Cv acquired by percentile method is greater than that
of MF-DFA method, implying that Cv obtained by percentile method is more
Table 10.8 Cv under different thresholds

Threshold MF-DFA Percentile Threshold MF-DFA Percentile
method method(95%) method method(95%)
Cv (SCI) 3.76888 16.6539 Cv 21.0380 12.6605
(DJAI)
time-clustering than that of MF-DFA method. Although Cv acquired by percentile

and MF-DFA methods are greater than 1, Cv obtained by MF-DFA method is
greater than the percentile method. The extreme events acquired by the MF-DFA
method are more time-clustering than that of the percentile method. The result
contradicted with that in Du et al. (2013), which indicates that Cv is greater and the
threshold is larger. Therefore, we chose the tercile again and calculated Cv . We
chose 90% and Cv was 12.00967. The result is smaller than the threshold of 95%,
confirming the conclusion of Yuan et al. (2012).
The percentile method is significantly affected by sample size and the fitting
distribution function compared with MF-DFA method without considering dynamic
mechanism and the background of extreme events. The percentile method tends to
be experiential and subjective. Therefore, the MF-DFA method is more reasonable
than the percentile method based on the frequency of extreme events, the range of
threshold, and their consistency with real life.
10.2.5 Effect of Extreme Value on the Cross-Correlation

of Chinese and American Stock Markets
First, using the sorting method, we found 20 daily rate of returns data that exceeded
the threshold of SCI and 10 daily rate of returns data that exceeded the threshold of
DJAI. We deleted the data in 30 different dates and analyzed the cross-correlation to
compare the cross-correlation of SCI and DJAI after removing the extreme values.
Figure 10.10 plots one to three–order scaling exponent hðqÞ of MF-DCCA
method for SCI and DJAI after deleting the extreme values. Figure 10.5 shows the
fluctuation trend of hðqÞ after removing the extreme values is almost similar to the
fluctuation trend without removing the extreme values. As q increased, hðqÞ
declined from approximately 0.65 to approximately 0.40. This result suggested that
the scaling exponent hðqÞ was not a constant, which indicates that multifractal
feature existed in the cross-correlation of the rate of returns of Chinese and
American stock markets. In addition, when 7 q 10, Hxy ðqÞ was smaller than 0.5.
When 10 q 4, Hxy ðqÞ was greater than 0.5; that is, when 10 q 4,
long-range cross-correlation existed between Chinese and American stock markets.
When 7 q 10, a negative long-range cross-correlation existed between Chinese
and American stock markets. Therefore, the thresholds are independent on the
cross-correlation of Chinese and American stock markets with no interference. This
0.7
pre-MF-DCCA-1
pre-MF-DCCA-2
0.65 pre-MF-DCCA-3
MF-DCCA-1
MF-DCCA-2
0.6
MF-DCCA-3
0.55
q
h
0.5
0.45
0.4
0.35
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 10.10 hðqÞ of MF-DCCA method of SCI and DJAI after deleting the extreme values
conclusion is the same with that of Yang et al. (2008). The extreme events do not
(or slightly) affect the long-range correlation of the entire system.
10.2.6 Discussion
10.2.6.1 Choice of the Value of Parameter d
To verify the conclusions above, parameter d should be changed to determine if the

thresholds are reasonable. When d = 0.001, we obtain the two figures below.
Figure 10.11 shows that when d = 0.001, the values of Dj of the original data,
rising trend, and down trend of SCI change with j. The first inflection point in the
positive section was j ¼ 0:024. The values of Dj tend to be smooth and steady with
the values of DFA of the original data. Thus, j ¼ 0:024 is considered as the positive
threshold of the current series. Similarly, j ¼ 0:029 is considered as the negative
threshold of the current series. As the second inflection point is j ¼ 0:039 as shown
in Fig. 10.1, Dj is almost equal to the values of DFA of the original data and tends
to be smooth and steady. Therefore, j ¼ 0:039 is also a positive threshold of this
series, but the frequency of the extreme events is smaller than j ¼ 0:024. Similarly,
j ¼ 0:04 and the frequency of the extreme events is also smaller than j ¼ 0:029.
0.95
Origianl data
Rising trend
0.9
Down trend
0.85
0.8 J=0.024
Dj
0.75
J=0.039
J=-0.029
0.7 J=-0.04
0.65
0.6
0.55
-0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
j
Fig. 10.11 the changes of values of DFA of series yj about the original data, rising trend and
down trend of SCI
This result is the same with the threshold we acquired above. Therefore, we con-
sider the thresholds obtained by MF-DCCA as robust.
Figure 10.12 shows that when d = 0.001, the values of Dj of the original data,
rising trend, and down trend of DJIA change with j. Similarly, the first inflection
point in the positive section is j ¼ 0:0275. The values of Dj tend to be smooth and
steady with the values of DFA of the original data. Therefore, j ¼ 0:0275 is con-
sidered the positive threshold of the current series. Similarly, j ¼ 0:0265 is
considered the negative threshold of the current series. As the second inflection
point is j ¼ 0:035 as shown in Fig. 10.12, Dj is almost equal to the values of DFA
of the original data and tends to be smooth and steady. Therefore, j ¼ 0:035 is also
a positive threshold of this series, but the frequency of the extreme events is smaller
than j ¼ 0:0275. Similarly, j ¼ 0:035 and the frequency of the extreme events
were also smaller than j ¼ 0:0265. Similarly, the thresholds are the same as the
thresholds we acquired above, further verifying the robustness of the results.
1.1
Origianl data
Rising trend
1
Down trend
0.9
0.8
J=0.0275
J=0.035
Dj
0.7
J=-0.035
0.6
J=-0.0265
0.5
0.4
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
j
Fig. 10.12 the changes of values of DFA of series yj about the original data, rising trend and
down trend of DIJA
10.2.6.2 Replacing of the Extreme Values
In this section, we replace the extreme values with the moving averages to verify
whether the thresholds are independent from the cross-correlation of Chinese and
American stock markets. For the parameter of the method of moving average n, we
chose n ¼ 10. Then, we obtain the following results:
Figure 10.13 shows the fluctuation trend of hðqÞ with q in which the extreme
values are replaced by the moving averages. The result is almost similar to the
fluctuation trend without removing the extreme values. As q increased, hðqÞ also
declined from approximately 0.65 to approximately 0.4. When 7 q 10,Hxy ðqÞ is
smaller than 0.5, and when 10 q 4, Hxy ðqÞ is greater than 0.5. A long-range
cross-correlation existed between Chinese and American stock markets when
10 q 4, whereas a negative long-range cross-correlation existed between
Chinese and American stock markets when 7 q 10.
The conclusion can be validated further by using the method of moving average;
that is, the thresholds are independent from the cross-correlation of Chinese and
American stock markets, there is no interference.
0.7
MF-DCCA-1
MF-DCCA-2
0.65 MF-DCCA-3
post-MF-DCCA-1
post-MF-DCCA-2
0.6
post-MF-DCCA-3
0.55
q
h
0.5
0.45
0.4
0.35
-10 -8 -6 -4 -2 0 2 4 6 8 10
q
Fig. 10.13 hðqÞ of MF-DCCA method of SCI and DJAI replaced with the moving averages
10.2.7 Conclusions
This paper investigated the thresholds and the cross-correlation between the
Chinese and American stock markets using the MF-DFA method. Our conclusions
are as follows:
(1) Compared with the traditional percentile method, the results obtained from the
MF-DFA method were more robust. Similarly, the time-clustering of extreme
events of DJAI was stronger than SCI, which suggests that DJAI was more
sensitive to extreme events than SCI.
(2) Thresholds were closely linked to real life. From the beginning of the equity
division of Chinese stock markets in May 2005, the thresholds were focused in
the year 2008 in Chinese and American stock markets. The share prices of two
house loan magnates in the US slumped and caused the subprime crisis in
August 2008. This crisis shocked the American and Chinese stock markets.
Therefore, thresholds in the Chinese and American stock markets are found in
September and October of that same year.
(3) By using the MF-DCCA method, we found that the cross-correlation for the
thresholds has no major effect on the two markets. Therefore, removing the
extreme values is not necessary when the cross-correlation of two markets is
investigated.
The advantages of the MF-DFA method compared with the percentile method
increases the popularity of MF-DFA method in the financial filed. The method also
reflects the advantages of interdisciplinary and creation and inspires the government
to emulate the American stock market and absorb the virtues of American stock
market mechanism. They should formulate reasonable mechanisms consistent
Chinese policies. In this manner, our stock market will improve. Moreover, the
range of threshold obtained using MF-DFA method can provide a reference for the
management organizations and investors to avoid the risks and improve the man-
agement of investments. Extreme events do not influence the cross-correlation.
Hence, we can conduct research on cross-correlation without considering the
extreme events. Our findings enrich the study on cross-correlation.
10.3 Research Prospect
10.3.1 Risk Measurement
Wei et al. (2013) combined the multifractal theory with the extreme value theory to
prove the superiority of the ARFIMA-MFV-EVT model with EVT theory and can
reflect the value of the risk. Suo et al. (2015) used the time scale for risk detection to
analyze and compare China’s spot and futures market volatility. Mainik et al.
(2015) used the multivariate extremum theory to invest in the daily returns of the
S & P 500 stock to minimize risk and found that the extreme risk index was
significantly better than the traditional method. At present, although scholars have
been involved in the fractal method of financial risk measurement, but did not take
into account the financial time series of asymmetric features, and there is no fractal
method used to measure the extreme risk of financial markets risk. In addition, the
use of fractal methods to measure the risk of the study did not consider the fractal
method itself, the existence of the error, because China’s stock market sample
length is not large enough (especially after the Shenzhen-Hong Kong), the use of
fractal method failure probability should also be considered in the risk In the range.
Therefore, it is worthy to study the asymmetric risk measure based on different ups
and downs and the financial extreme risk measure based on DFA-E.
10.3.2 Cross Market Investment Research
In 1952, Professor Markowitz put forward the concept of a portfolio, established a

modern portfolio theory, the risk into the system and non-systematic risk, so as to
guide investors to optimize their investment behavior. Since then, Sharpe, Lintner,
etc. in order to strengthen the application of the theory, the micro-research is turned
to the whole market, which is simplified into a single factor relationship based on
10.3 Research Prospect 253
the market index and it is found that the return of the capital assets and the risk
follows the linear relationship between the capital assets under the equilibrium
market condition, that is, the capital asset pricing based on the mean-variance
model Model (CAPM).CAPM believes that non-systemic risk can be eliminated
through decentralized investment, and systemic risk requires risk compensation, the
higher the systemic risk of an asset, the greater the expected return. However, due
to the premise that the premise is too harsh, many economists try to study the
pricing theory under certain conditions of weakening. Since then, the new pricing
theory has been put forward, such as the existence of a large number of
non-market-oriented asset investment pricing theory, arbitrage pricing theory
(APT) and asset yield and average consumption growth rate linear relationship
model (CCAPM), etc. (Mayers 1972; Ross 1976; Breeden et al. 1989). However, in
essence, these models are still confined to CAPM, provided that the yield of
financial assets should be subject to a normal distribution, that is, the financial
market itself to meet the “effective market hypothesis” (EMH). However, the yield
of securities is not independent of each other, and its change does not obey the
random walk model, the probability distribution is not a normal distribution, and
presents a “peak fat tail” characteristics. The traditional theory of finance also
attempts to make some work on CAPM. The well-known work is the three-factor
pricing model proposed by Fama and French in the 1993 and 1996 papers.
However, regardless of the number of factors used in the model, the multi-factor
model itself can not escape the scope of the linear model. In fact, the linearity of the
other model factors is a subjective inference, except that the linearity of the risk
factor-standard deviation r is ensured by the capital asset pricing model. In fact, the
Hurst index is a non-linear factor. In this regard, Raei and Mohammadi (2008) give
a fractional capital asset pricing model (FCAPM), and define the corresponding
fractional beta coefficient, the systemic risk of the interpretation and therefore can
be improved. These studies also mark the relationship between risk and profit from
the traditional linear transformation is non-linear.
Moreover, the CAPM model is essentially a revenue-risk model, and its risk
measure is mainly based on variance, but in practice, when the market does not
meet the prerequisites for an effective market hypothesis, the variance and the
accuracy of the CAPM indicator Greatly reduced. Some scholars have jumped out
of this framework and reconstructed the risk measurement model. Weiyu et al.
(2005) established a market risk measure Rf based on two main parameters of
multi-scale fractal spectrum. However, this risk measure still does not account for
the overall risk of multiple assets as a whole system. Therefore, it is worthwhile to
construct a fractional-order CAPM model based on the portfolio.
References
J. Alvarez-Ramirez, E. Rodriguez, J. Carlos Echeverria, A DFA approach for assessing

asymmetric correlations. Phys. A 388(12), 2263–2270 (2009)
M. Alzahrani, M. Masih, O. Al-Titi, Linear and non-linear Granger causality between oil spot and
futures prices: a wavelet based test. J. Int. Money Financ. 48, 175–201 (2014)
E. Baek, W. Brock, A general test for nonlinear Granger causality: bivariate model, Iowa State
University and University of Wisconsin at Madison working paper (1992)
T. Baig, I. Goldfajn, Financial market contagion in the Asian crisis. IMF Staff Paper 46(2), 167–
195 (1999)
J. Bartholy, R. Pongrácz, Analysis of precipitation conditions for the Carpathian Basin based on
extreme indices in the 20th century and climate simulations for 2050 and 2100. Phys. Chem.
Earth 35, 43–51 (2010)
B.-H. Boyer, T. Kumagai, K. Yuan, How do crises spread? Evidence from accessible and
inaccessible stock indices. J. Finance 61(2), 957–1003 (2006)
T. Breeden, R. Douglas, G. Michael, H.L. Robert, Empirical tests of the consumption‐oriented
CAPM. J. Finance 44(2), 231–262 (1989)
A. Burgueño, X. Lana, C. Serra, M.D. Martínez, Daily extreme temperature multifractals in
Catalonia (NE Spain). Phys. Lett. A 378, 874–885 (2014)
G.X. Cao, W. Xu, Multifractal features of EUA and CER futures markets by using multifractal
detrended fluctuation analysis based on empirical model decomposition. Chaos Solitons
Fractals 83, 212–222 (2016a)
G.-X. Cao, W. Xu, Nonlinear structure analysis of carbon and energy markets with MFDCCA
based on maximum overlap wavelet transform. Phys. A 444, 505–523 (2016b)
G.X. Cao, J. Cao, L.-B. Xu, Asymmetric multifractal scaling behavior in the Chinese stock market:
based on asymmetric MF-DFA. Phys. A 392(4), 797–807 (2013)
G.-X. Cao, J. Cao, L.-B. Xu et al., Detrended cross-correlation analysis approach for assessing
asymmetric multifractal detrended cross-correlations and their application to the Chinese
financial market. Phys. A 393, 460–469 (2014a)
G.-X. Cao, Y. Han, W.-J. Cui et al., Multifractal detrended cross-correlations between the CSI 300
index futures and the spot markets based on high-frequency data. Phys. A 414, 308–320
(2014b)
T. Chang, S.C. Cheng, G. Pan et al., Does globalization affect the insurance markets? Bootstrap
panel Granger causality test. Econ. Model. 33(2), 254–260 (2013)
S. Chen, S.-F. Li, X.-X. Li, Empirical study on co-movement effect between “China Concept”
shares, shares, Chinese and Abroad Securities Market. J. Manag. Sci. 21(4), 105–110 (2008).
(in Chinese)
H.B. Du, Z.F, Wu, N. Zhang, S.W. Zong, X.J. Meng, The charateristics of changes of extreme
temperature and rainfall events the recent 60a in Dandong. Sci. Geogr. Sin. 4, 467–474 (2013)
(in Chinese)
C. Ellis, P. Wilson, Another look at the forecast performance of ARFIMA models. Int. Rev.
Financ. Anal. 13(1), 63–81 (2004)
V. Fernandez, Linear and non-linear causality between price indices and commodity prices.
Resour. Policy 41, 40–51 (2014)
M. Gemmer, T. Fischer, T. Jiang et al., Trends in precipitation extremes in the Zhujiang River
Basin, South China. J. Clim. 24, 750–761 (2011)
B. Gnedenko, Sur la distribution limite du terme maximum d’une se’rie ale’atoir. Ann. Math. 44,
423–453 (1943)
M.A.S. Granero, J.E.T. Segovia, J.G. Pérez, Some comments on Hurst exponent and the long
memory processes on capital markets. Phys. A 387(22), 5543–5551 (2008)
C.W.J. Granger, Investigating causal relations by econometric models and cross-spectral methods.
Econometrica 37(3), 424–438 (1969)
References 255
C. Hiemstra, J.D. Jones, Testing for linear and nonlinear Granger causality in the stock
price-volume relation. J. Financ. 49(5), 1639–1664 (1994)
H.E. Hurst, The long-term storage capacity of reserviors. Trans. Am. Soc. Civ. Eng. 116(1), 770–799
(1951)
A.F. Karagiannidis, T. Karacostas, P. Maheras, T. Makrogiannis, Climatological aspects of
extreme precipitation in Europe, related to mid-latitude cyclonic systems. Theor. Appl.
Climatol. 107, 165–174 (2012)
T.H. Lee, W. Yang, Granger-causality in quantiles between financial markets: using copula
approach. Int. Rev. Financ. Anal. 33(5), 70–78 (2014)
F. Longin, The asymptotic distribution of extreme stock market return. J. Bus. 69, 383–408 (1996)
G. Mainik, G. Mitov, L. Rüschendorf, Portfolio optimization for heavy-tailed assets: extreme risk
index vs Markowitz. J. Empirical Finance 32, 115–134 (2015)
Y. Malevergne, D. Sornette, Testing the Gaussian copula hypothesis for financial assets
dependences. Quant. Financ. 3, 231–250 (2001)
D. Mayers, Nonmarketable assets and capital market equilibrium under uncertainty. Stud. Theor.
Capital Markets 1, 223–248 (1972)
R. Raei, S. Mohammadi, Fractional return and fractional CAPM. Appl. Financ. Econ. Lett. 4(4),
269–275 (2008)
S.A. Ross, The arbitrage theory of capital asset pricing. J. Econ. Theor. 13(3), 341–360 (1976)
F. Selcuk, Financial earthquakes sftershocks and scaling in emerging stock markets. Phys. A 333,
306–316 (2004)
F.M. Siokis, European economics in crisis: a multifractal analysis of disruptive economic events
and the effects of financial assistance. Phys. A 395, 283–293 (2014)
Y.Y. Suo, D.H. Wang, S.P. Li, Risk estimation of CSI 300 index spot and futures in China from a
new perspective. Econ. Model. 49, 344–353 (2015)
T. Verdelhos, P.G. Cardoso, M. Dolbeth, M.A. Pardal, Recovery trends of a Scrobicularia plana
populations after restoration measures, affected by extreme climate events. Mar. Environ. Res.
98, 39–48 (2014)
P. Weber, F.Z. Wang, V.C. Irena, H. Shlomo, H.E. Stanley, Relation between volatility
correlations in financial markets and Omori processes occurring on all scales. Phys. Rev. E 76,
016109-1–016109-6 (2007)
Y. Wei, W. Chen, Y. Lin. Measuring daily value-at-risk of SSEC index: a new approach based on
multifractal analysis and extreme value theory. Phys. A 392(9), 2163–2174 (2013)
Y. Xue, L.S. Jia, W.Z. Teng et al., Long-range correlations in vehicular traffic flow studied in the
framework of Kerner’s three-phase theory based on rescaled range analysis. Commun.
Nonlinear Sci. Numer. Simul. 22(1), 285–296 (2015)
P. Yang, W. Hou, G.L. Fen, Confirming the thresholds of extreme events based on the multifractal
detrended fluctuation analysis. Chin. J. Phys. 57, 5333–5342 (2008). (in Chinese)
Y. Yuan, X.T. Zhuang, Z.Y. Liu, W.Q. Huang, Time-clustering behavior of sharp fluctuation
sequences in Chinese stock markets. Chaos Solitons Fractals 45, 838–845 (2012)
A.G. Zawadowski, J. Kertész, G. Andor, Large price changes on small scales. Phys. A 344, 221–226
(2004)

2018 Book MultifractalDetrendedAnalysisM

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

2018 Book MultifractalDetrendedAnalysisM

Caricato da

Copyright:

Formati disponibili

Guangxi Cao

ISBN 978-981-10-7915-3 ISBN 978-981-10-7916-0 (eBook)

© Springer Nature Singapore Pte Ltd. 2018

Printed on acid-free paper

3.2.3 Sources of Multifractality . . . . . . . . . . . . . . . . . . . . . . . . 28

5.3 Empirical Analysis on International Gold Markets . . . . . . . . . . . 95

8.2.2 Comparative Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

10.2 Extreme Values Evaluation Based on Detrended Fluctuation

© Springer Nature Singapore Pte Ltd. 2018 1

1.1 A Historical Evolution of Fractal Methods

1.2 Application Areas

E. Alessio, A. Carbone, G. Castelli, V. Frappietro, Second-order moving average and scaling of

2.1.1 R/S and Modiﬁed R/S

log ððR=SÞn Þ ¼ log c þ H log n; ð2:1Þ

Equivalently, the value H can be obtained by estimating the slope of

log Hn ¼ log c þ ðH 0:5Þ log n; ð2:3Þ

where log Hn ¼ log ½ðR=SÞn EðR=SÞn . The Hurst exponent H estimated by

2.1.2 DFA Method

The DFA method can be summarized as the following Peng (1994):

2.3 Estimation and the Descriptive Statistics

2.3.2 Descriptive Statistics

2.4 Relationship Between the Two Time-Varying Hurst

We discuss the relationship between the time-varying Hurst exponents of the

2.4.1 Unit Root Test

Table 2.4 ADF test for time-varying Hurst exponent series

2.4.2 Cointegration Test

2.4.3 Granger Causality Test

We investigated the evolution of time-varying Hurst exponents of the Chinese

stock market, the market-oriented level of them need be improved through

E. Abounoori, M. Shahrazi, S. Rasekhi, An investigation of forex market efﬁciency based on

© Springer Nature Singapore Pte Ltd. 2018 21

Assume that there is a time series fyðiÞg, i ¼ 1; 2; . . .; N, where N is the length of

for each segment v, v ¼ 1; 2; . . .; N s , and

for any q 6¼ 0, and

Fq ðsÞ shðqÞ ð3:6Þ

Through Eqs. (3.4) and (3.6), we can obtain

The partition function can be deﬁned as

Then we can obtain

sðqÞ ¼ qhðqÞ 1 ð3:11Þ

By Legendre Transformation, we can obtain the relationship between s(q) and

Through Eqs. (3.11) and (3.12), we can obtain

a ¼ hðqÞ þ qh0 ðqÞ

Thus the Da can be estimated by the parameters a, b.

3.1.2 Partition Function

To keep our description as self-contained as possible, let us review briefly the

rðiÞ ¼ logðTði þ tÞ=TðiÞÞ; ð1  t  250Þ ð3:16Þ

Na ðdÞ df ðaÞ ; ðd ! 0Þ ð3:19Þ

where q can range from −∞ to +∞ theoretically. When q is positive, the partition

Sq ðdÞ dsðqÞ ; ðd ! 0Þ ð3:21Þ

Therefore, another way to characterize a multifractal series is the singularity

f ðaÞ ¼ aq sðqÞ ð3:24Þ

Through Eqs. (3.20), (3.22) and (3.23), we can obtain

3.2 Empirical Analysis on Developed-Emerging

f ðiÞ ¼ jlnðTði þ DtÞ=TðiÞÞj; ði  L DtÞ ð3:26Þ

In our case, because our data are daily records, namely, Dt ¼ 1.

3.2.2 Multifractal Spectrum Analysis

By means of the above-mentioned model, we ﬁrst of all obtained lnFq(s) versus

lns lns q=0

3.2.3 Sources of Multifractality

In current literature, two major sources of multifractality are widely acknowledged

we shuffled the series, it implies that there is no influence by nonlinear temporal

1.2 China 1.0

unchanged; thereby, the fat-tailed probability distribution is by no means the

3.2.4 Comparative Analysis

where log Hn ¼ log ½ðR=SÞn EðR=SÞn . The Hurst exponent H estimated by

rðiÞ ¼ logðTði þ tÞ=TðiÞÞ; ð1 t 250Þ ð3:16Þ

f ðiÞ ¼ jlnðTði þ DtÞ=TðiÞÞj; ði L DtÞ ð3:26Þ

Fq ðsÞ sHxy ðqÞ : ð4:6Þ

Hxy ðqÞ ½Hxy ðqÞ þ Hxy ðqÞ=2: ð4:15Þ