Expert Systems with Applications 42 (2015) 3868–3874

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Time series forecasting based on wavelet filtering

Tae Woo Joo 1, Seoung Bum Kim ⇑
School of Industrial Management Engineering, Korea University, Anam-Dong, Seongbuk-Gu, Seoul 136-713, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: Forecasting time series data is one of the most important issues involved in numerous applications in real
Available online 20 January 2015 life. Time series data have been analyzed in either the time or frequency domains. The objective of this
study is to propose a forecasting method based on wavelet filtering. The proposed method decomposes
Keywords: the original time series into the trend and variation parts and constructs a separate model for each part.
ARIMA Simulation and real case studies were conducted to examine the properties of the proposed method
Forecasting under various scenarios and compare its performance with time series forecasting models without wave-
Time series
let filtering. The results from both simulated and real data showed that the proposed method based on
Wavelet transforms
wavelet filtering yielded more accurate results than the models without wavelet filtering in terms of
mean absolute percentage error criterion.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction To address the limitations of Fourier transforms, wavelet trans-

A time series is a set of observations collected over time (Lai
et al., 2010). The analysis of time series data is one of the most
important areas in statistics in both theory and application. The
main objective of time series analysis is to obtain inherent struc-
tural characteristics, such as autocorrelations, trends, and seasonal
variations, and to use this information to formulate an appropriate
mathematical model for analysis and prediction (Anderson, 1971).
In general, time series data can be analyzed in either the time or
frequency domains. The most widely used methods in analyzing
the time domain include time series regression, decomposition
methods, exponential smoothing, and the Box–Jenkins autoregres-
sive integrated moving average (Li, Gai, Kang, Wu, & Wang, 2014).
Analyses in the frequency domain are usually conducted for
periodic and cyclical observations. The methodologies of the fre-
quency domain are based on Fourier transforms that allow us to
determine the number of frequency components and detect the
dominant cyclic frequency, all of which are embedded in the time
domain. However, Fourier transforms-based methods have limita-
tions in that they require an assumption of stationarity and pro-
duce no information associated with time (Croarkin & Tobias,
2006; Popoola, 2007; Tabak and Feitosa, 2010). Further, Fourier
transforms do not work well in large data, and thus, they can be
implemented only on an interval between 0 and 2p (Tran, 2006).
⇑ Corresponding author. Tel.: +82 2 3290 3397; fax: +82 2 929 5888.
E-mail addresses: mrblond002@gmail.com (T.W. Joo), sbkim1@korea.ac.kr
(S.B. Kim).
1
Tel.: +82 2 3290 3769.
forms that are localized in both the time and frequency domains
(Mallat, 1989) have been proposed (Morettin, 1996, 1997;
Mousa, Munib, & Moussa, 2005; Percival & Walden, 1999;
Priestley, 1996). More specifically, because wavelet basis functions
exist over a finite time limit and are typically irregular and asym-
metric, they are better suited for those time series analyses that
exhibit sharp discontinuities and local behavior (Graps, 1995).
Many prediction approaches based on wavelet transforms have
been developed recently (Chen, 2014; Nguyen, Khosravi,
Creighton, & Nahavandi, 2015). Nguyen et al. (2015) presents a
combination of wavelet features with fuzzy standard additive
model for medical diagnosis. In general, if a time series is non-
stationary, it is difficult to determine a relevant global model. For
example, Hydrological time series forecasting is a difficult task
because of its complicated nonlinear, non-stationary and multi-
scale characteristics (Di, Yang, & Wang, 2014). To overcome this
problem, local models based on wavelet transforms have been pro-
posed (Weigend & Mangeas, 1995; Zhang, Coggins, Jabri, Dersch, &
Flower, 2001). In addition, it is known that wavelet transforms
have the potential to increase the accuracy of time series predic-
tions (Ramsey, 1999; Schlüter & Deuschle, 2010). Once forecasting
models using wavelet transforms succeeded in eliminating noise
before preceding to construct a model (Alrumaih & Al-Fawzan,
2002), improvement in overall forecasting performance followed
(Chou, 2014; Ferbar, Čreslovnik, Mojškerc, & Rajgelj, 2009).
Other approaches that use wavelet transforms estimate compo-
nents in a structural time series model (Arino, 1995; Sang, 2013;
Wong, Wai-Cheung, Zhongjie, & Lui, 2003; Zhang et al., 2001).
Arino (1995) proposed a methodology that used wavelet

http://dx.doi.org/10.1016/j.eswa.2015.01.026
0957-4174/Ó 2015 Elsevier Ltd. All rights reserved.
 T.W. Joo, S.B. Kim / Expert Systems with Applications 42 (2015) 3868–3874
transforms to decompose time series into long-term trends and
seasonal components. Specifically, let y = {yt : t = 1, . . ., T} be a time
series set. y is decomposed into two sets z1 = {z1t : t = 1, . . ., T} and
z2 = {z2t : t = 1, . . ., T}, that is, y ¼ z1 þ z2 , where z1 and z2 , respec-
tively, represent long-term trends and seasonal behavior. Each
component is used to construct an appropriate model, and total
prediction accuracy can be obtained by aggregating the prediction
results from each component model. However, the generalizability
of this method is questionable because Arino’s method considered
only limited cases such as long-term trends and seasonal
variations.
Moreover, some methods applied linear or nonlinear predictive
models to wavelet coefficients (approximation and detail coeffi-
cients) at each level, and the final prediction was obtained by
inverting the predicted wavelet coefficients. Renaud, Starck, and
Murtagh (2002) introduced the multiscale autoregressive models
that use Haar wavelets and scale coefficients during decomposition.
A method that combines nonlinear models with wavelet coeffi-
cients has also been proposed (Chen, Qian, & Meng, 2013; Hadaś-
Dyduch, 2014; Rocha, Paredes, Carvalho, Henriques, & Harris,
2010; Soltani, 2002; Wang, 2014). Chen, Nicolis, and Vidakovic
(2010) classified the wavelet coefficients into trend, seasonal, and
high frequency components and used them to construct forecasting
models based on exponential smoothing, harmonic regression, and
autoregressive moving average with exogenous input (ARMAX)
model. However, these methods involve a high computational load
because they considered each level of a series. Moreover, consider-
ing each level of series could cause an overfitting problem.
In this paper, we propose a forecasting method that uses wave-
let filtering. Through wavelet transforms, the series is partitioned
into two parts (trends and variations). We then construct a sepa-
rate model for each part. By this partitioning, the proposed method
is especially useful for the time series with a large amount of noise.
The potential overfitting problem of this approach caused by using
two separate models is addressed by adjusting the decomposition
levels.
The rest of this paper is organized as follows. In Section 2, we
briefly introduce the basic concept of wavelet transforms. In
Section 3, we present our proposed forecasting approach based
on wavelet filtering. Section 4 presents a simulation study to
examine the properties of the proposed method and evaluate its
performance under various scenarios. Section 4.2 presents the
experimental results from real-life problems. Finally, Section 5
contains our concluding remarks.
2. Wavelet transforms
Wavelets have the advantages of the locality of the analysis and
their ability to handle multiscale information efficiently. Numer-
ous studies of wavelets have been conducted in the fields of sig-
nal/image processing (Meerwald & Uhl, 2001; Prasad & Iyengar,
1997; Rao & Bopardikar, 1998; Subasi, 2007; Avci & Derya, 2008),
statistics (Abramovich, Bailey, & Sapatinas, 2000; Antoniadis,
1999; Vidakovic, 1999), and manufacturing processes (Guo,
Linyan, Gang, & Song, 2008; Jeong, Lu, & Wang, 2006; Jin & Shi,
2001; Lada, Lu, & Wilson, 2002; Saravanan & Ramachandran,
2009). Although Fourier transforms give only the frequency infor-
mation of given data, wavelet transforms can simultaneously deli-
ver both the time and frequency localizations (Mallat, 1989).
Wavelet transforms are conducted from wavelet basis functions
that consist of a scale wavelet part /ðtÞ and a detail part wðtÞ:
/j;k ðtÞ ¼ 2j=2 /ð2j t  kÞ;
wj;k ðtÞ ¼ 2j=2 wð2j t  kÞ;
3869
Z Z
/ðtÞdx ¼ 1 wðtÞdx ¼ 0 /ðtÞ; wðtÞ 2 L2 ðRÞ;
where j and k denote, respectively, the scaling parameter and trans-
lation index. L2 ðRÞ is the space of square integrable real function
defined on the real line R. The data can be decomposed into the fol-
lowing wavelet series form:
X
n X
J X
n
yðtÞ ¼ cj;k /j;k ðtÞ þ dj;k wj;k ðtÞ; ð2Þ
k¼1 j¼1 k¼1
where cj;k represents the approximation coefficient at scale j and
location k; J is the decomposition level; dj;k represents the detail
coefficient at scale j and location k; and n is the size of the time ser-
ies data. Reconstruction can also be done through Mallet’s pyramid
algorithm (Burt & Adelson, 1983). Fig. 1 shows an example of the
overall process of decomposition and reconstruction by wavelet
transforms.
In Fig. 1, cj;k represents the coarse approximation (trend and
seasonality) and dj;k represents detailed information (noise or ran-
dom fluctuation). The difference between the first level approxi-
mation coefficients c1;k and original series y yields the detail
coefficients of the first level d1;k : To obtain c2;k ; c1;k is approximated
by a set of basis functions. As can be seen from Fig. 1, when the
level becomes higher, cj;k represents the overall pattern of the ori-
ginal series.
3. Proposed forecasting method
3.1. Data decomposition through wavelet filtering
Our proposed method consists of three steps: (1) wavelet trans-
forms to decompose the data in the trend part (TP) and variation
part (VP), (2) determination of the optimal decomposition level,
and (3) construction of forecasting models. First, wavelet trans-
forms are conducted to the original time series data to obtain the
same number of coefficients as the size of the data. Further, deno-
ising that shrinks the empirical detail wavelet coefficients toward
zero is performed to remove noise. We used soft thresholding in
this study (Donoho & Johnstone, 1994). A wavelet filtering method
is used to partition the data into the TP and VP (Chang and Yadama,
2010). More precisely, the TP can be obtained by setting all the
detail coefficients at all levels dj;k ðj ¼ 1; . . . ; JÞ to zero while main-
taining all coarse approximation coefficients, cj;k ðj ¼ 1; . . . ; JÞ. Simi-
larly, the VP can be obtained by setting all the coarse
approximation coefficients at the last level, cJ;k to zero while main-
taining all detail coefficients. The TP and VP can be considered the
overall trend and variation information in a time series. Fig. 2
shows how a series can be decomposed into the TP and VP by
wavelet filtering. Note that in this paper, we used Daubechies
wavelets as a basis function whose scale and detail wavelets are
shown in Fig. 3.
3.2. Choosing the optimal decomposition level
For modeling, the dataset is divided into training (80%), valida-
tion (10%), and testing (10%) sets. The training set is used to create
a model, and the testing set is used to evaluate the model. The val-
idation set is used to select the optimal parameters for the decom-
position level of wavelet transforms. To determine an adequate
decomposition level, we varied the number of levels for both the
TP and VP from one to six until the model delivered the best accu-
racy in terms of the minimum mean absolute percentage error
(MAPE). We believe a potential overfitting problem by using two
separate models (TP and VP) can be addressed by adjusting the
ð1Þ
decomposition levels in the wavelet transforms.
3870 T.W. Joo, S.B. Kim / Expert Systems with Applications 42 (2015) 3868–3874

Fig. 1. A process of decomposition and reconstruction by wavelet transforms.

Fig. 2. Trend and variation parts by wavelet filtering.

Fig. 3. Daubechies 2 wavelets.

3.3. Forecasting modeling study is an autoregressive integrated moving average (ARIMA)

model that is defined as follows:
Once the time series is decomposed into the TP and VP at its
optimal decomposition levels, we constructed the forecasting
rd yt ¼ /1 y1 þ /2 y2 þ    þ /p ytp þ at  h1 at1  h2 at2    
model for each part. The forecasting model that we used in this  hq atq ð3Þ
 T.W. Joo, S.B. Kim / Expert Systems with Applications 42 (2015) 3868–3874 3871

The order of an ARIMA model can be represented by the notation

ARIMA (p, d, q), where p, d, and q are, respectively, the order of
the autoregressive, integrated, and moving average parts of the
model. Because ARIMA models require an assumption of stationa-
rity, we used an augmented Dickey–Fuller test to check whether
or not the series is stationary (Dickey & Fuller, 1979). If the test indi-
cates that the series is nonstationary, we first difference the series.
The augmented Dickey–Fuller test is implemented for the differ-
enced series; if the series is still nonstationary, we again first differ-
ence the differenced series. The above process is repeated until the
series is stationary. For data with seasonality, we used seasonal
ARIMA (SARIMA) models that can take seasonal variations into
account. ARIMA (or SARIMA) models for the TP and VP are sepa-
rately determined based on penalty function statistics such as
Akaike’s Information Criterion (Wang, Kang, Shen, Chang, &
Chung, 2010). Finally, total prediction accuracy can be obtained
by aggregating the prediction results from each of the TP and VP
models.

4. Simulation

4.1. Simulation data

We conducted a simulation to evaluate the performance of the

proposed method based on wavelet filtering and compared it
with the ARIMA and SARIMA models without wavelet filtering.
In our simulation study, we considered the time series with and
without seasonal variations. Fig. 4 shows eight examples of time
series data that exhibit increasingly linear trends without sea-
sonal variations ((1)–(4)) and with seasonal variations ((5)–(8)).
We generated different noise levels from the normal distributions
with mean zeros and four different standard deviations (r = 5, 10,
20 and 30).
In the examples of seasonality, we generated the dataset by
using a function that combines sine and cosine. The total data size
was 132, with a period of 12. To measure the performance of our
methods, we ran 200 replications and calculated MAPE, which
can be computed by the following equation:
n  
1X ^
yi  yi   100; Fig. 4. Time series data with linearly increasing trend with different amounts of
MAPE ¼ ð4Þ
n i¼1  yi  noise: (1), (5): r = 5; (2), (6): r = 10; (3), (7): r = 20; (4), (8): r = 30.

where yi and y ^i are, respectively, the actual and predicted values.

Table 1 shows the average values of MAPE from 200 experi-
Table 1
ments between ARIMA (or SARIMA) models without wavelet filter- Comparative results between ARIMA (or SARIMA) and the proposed method for
ing and from the proposed procedure that constructs ARIMA (or simulated datasets. The values inside parentheses indicate the standard deviations of
SARIMA) models with wavelet filtering. The values inside the MAPE.
parenthesis indicate standard deviations. The results showed that Scenario ARIMA (or SARIMA) models ARIMA (or SARIMA) models
for all scenarios but one (scenario 1) the proposed forecasting pro- without wavelet filtering with wavelet filtering
cedure yielded smaller MAPEs for the testing set than the ARIMA (proposed method)
(or SARIMA) models did. We can see that for the datasets with a (1) 2.11 (1.18) 2.12 (0.66)
small amount of noise (scenarios 2, 5, 6), the proposed procedure (2) 4.75 (2.08) 4.29 (1.35)
was more accurate, but not to a significant degree. However, for (3) 8.73 (2.76) 7.51 (1.79)
(4) 12.33 (3.40) 11.28 (2.5)
scenarios with a large amount of noise (scenarios 3, 4, 7, 8), the
(5) 1.95 (0.44) 1.59 (0.38)
proposed procedure yielded significantly better prediction accu- (6) 3.90 (0.94) 3.07 (0.68)
racy than pure ARIMA (or SARIMA) models. These simulation (7) 7.84 (1.92) 6.09 (1.34)
results clearly demonstrated the usefulness of the proposed proce- (8) 12.28 (3.83) 9.30 (1.91)
dure for forecasting.

4.2. Case study
In this section we demonstrated the applicability of the pro-
posed forecasting method by using eight real life datasets. Table 2
shows a set of time series plots of the dataset. The values under the
name of the dataset indicate the size of the data. The first five data-
sets (Viscosity, DVD, Towel, KOSPI, and House) lack seasonal
variations. The last three datasets (Net Profit, Passengers, Max
Power) exhibit seasonal variations. To identify the relevant periods
of the time series with seasonal variations, we used Fourier trans-
forms. The Viscosity dataset contains the daily viscosity of chemi-
cal product XR-22. DVD contains a weekly sales market, and Towel
contains the weekly sales of absorbent paper towels. These three
datasets are available in the book ‘‘Forecasting, Time Series, and
3872 T.W. Joo, S.B. Kim / Expert Systems with Applications 42 (2015) 3868–3874

Table 2
A set of time series plots of eight real datasets. The values under the name of the dataset indicate the size of the data.

No. Datasets Plot No. Datasets Plot

1 Viscosity (150) 42
5 House (143) 200

40 190

180
38
170
36
160
34
150
32
140

30
130

28 120

26 110

24 100
0 50 100 150 0 50 100 150

2 DVD (161) 90
6 Net Profit (120) 1100

1000
80
900
70
800

60 700

600
50
500
40
400

30 300

200
20 0 20 40 60 80 100 120
0 20 40 60 80 100 120 140 160 180

3 Towel (120) 20
7 Passengers (132) 2.2
x 10
6

2
15

1.8

10
1.6

1.4
5
1.2

0 1

0.8

-5
0 20 40 60 80 100 120 0.6
0 20 40 60 80 100 120 140

4 KOSPI (273) 1.6

x 10
6
8 Max Power (365) 8000

7500
1.55

7000
1.5

6500
1.45

6000
1.4

5500
1.35

5000
1.3

4500
1.25

4000
1.2 0 50 100 150 200 250 300 350
0 50 100 150 200 250 300

Regression: An Applied Approach (Fourth Edition)’’ (Bowerman, (Lee, 2009). The Passengers data exhibits seasonal variations with
O’Connell, & Koehler, 2005). KOPSI contains a series of stock prices a period of 12 and contains the monthly number of passengers at
from the Korea Exchange (KRX, http://www.krx.co.kr). The House Jeju International Airport from January 2003 to December 2013;
dataset contains the monthly house prices for first time buyers this dataset can be obtained from the Korea Airports Corporation
from February 2002 to December 2013 in the UK and were (http://www.airport.co.kr). Max Power contains the daily maxi-
obtained from the Office for National Statistics website (http:// mum electricity demand in 2013 in South Korea. This dataset con-
www.ons.gov.uk). The Net Profit dataset exhibits a linearly increas- tains a seasonal pattern with the period of 7 and can be obtained
ing trend with seasonal variations (period 12) that is available in from the Electric Power Statistics Information System (http://
the book ‘‘Statistical Methods for Prediction (Second Edition)’’ www.kpx.or.kr).

Table 3
The comparative results of MAPE between the ARIMA (or SARIMA) models with and without wavelet filtering.

Data ARIMA models without wavelet filtering ARIMA models with wavelet filtering (proposed method)
Parameters MAPE (%) Parameters MAPE (%)
Optimal decomposition level TP VP
Viscosity (0, 0, 2) 4.81 3 (2, 0, 0) (2, 0, 1) 5.09
DVD (2, 1, 1) 5.38 3 (2, 1, 1) (2, 0, 2) 4.02
Towel (0 ,1, 1) 6.64 4 (2, 2, 2) (2, 0, 1) 6.03
KOSPI (2, 1, 2) 2.88 5 (2, 2, 2) (2, 0, 1) 2.47
House (1, 1, 1) 3.19 2 (2, 2, 2) (2, 0, 1) 2.44
Net Profit (0, 2, 2) (0, 2, 2)s = 12 6.55 4 (2, 2, 2) (0, 2, 2) (0, 2, 2)s = 12 4.83
Passengers (2, 2, 2) (0, 2, 2)s = 12 13.21 4 (2, 2, 2) (2, 2, 2) (0, 2, 2)s = 12 11.13
Max Power (2, 2, 2) (0, 2, 2)s = 7 11.02 5 (2, 1, 2) (1, 0, 1) (0, 2, 2)s = 7 7.85
 T.W. Joo, S.B. Kim / Expert Systems with Applications 42 (2015) 3868–3874
Table 3 shows the comparative results with the real data that
were obtained with ARIMA (or SARIMA) models without and with
wavelet filtering. Our proposed method uses the models with
wavelet filtering. The results demonstrated that in seven of eight
cases the proposed approach with the forecasting model with
wavelet filtering yielded smaller MAPEs than the ARIMA (or SAR-
IMA) models. In particular, in the Passenger and Max Power cases,
the proposed method significantly outperformed the SARIMA
model. Note that various optimal decomposition levels were found.
In summary, these results from real life examples demonstrated
the usefulness and applicability of the proposed method.
5. Conclusions
In this paper we have proposed a forecasting method based on
wavelet filtering. Time series data usually comprises many local
bumps and peaks with different scales. Wavelets have the advan-
tage of the locality of the analysis and their ability to handle mul-
tiscale information efficiently. The main idea of the proposed
method is decomposition of the original data into the TP and VP
that, respectively, represent the overall trend and variations
through wavelet filtering and subsequent construction of separate
forecasting models for each of them. A potential overfitting
problem by using two models was addressed by adjusting the
decomposition levels in the wavelet transforms. Technically, to
determine an adequate decomposition level, we changed the num-
ber of levels for both the TP and VP from one to six until we found
the model that minimizes MAPE. In the current study, we selected
the decomposition level somewhat arbitrarily. Nevertheless, both
the simulation and real life examples demonstrated the usefulness
and applicability of the proposed forecasting method. In particular,
the proposed method is useful for situations where seasonality
with a large amount of noise exist.
There are several interesting directions for future research. A
more comprehensive simulation should be conducted to evaluate
the proposed forecasting method in terms of efficiency and appli-
cability. Another interesting and important future research issue
is the derivation of the variance of the forecasting outcome so as
to construct the reliable predication interval. Further, we may
use different wavelet basis functions other than Daubechies wave-
lets used in this study to examine their impact to forecasting accu-
racy. We believe that the fundamental value of the present study
lies in its integration of wavelet transforms and traditional fore-
casting methods to increase forecasting accuracy.
Acknowledgements
The authors sincerely thank the referees for their helpful sug-
gestions and comments, which greatly improved the quality of
the paper. This research was supported by This research was sup-
ported by Brain Korea PLUS and Basic Science Research Program
through the National Research Foundation of Korea funded by
the Ministry of Science, ICT and Future Planning (2013007724).
References
Abramovich, F., Bailey, T. C., & Sapatinas, T. (2000). Wavelet analysis and its
statistical applications. The Statistician, 49, 1–29.
Alrumaih, R. M., & Al-Fawzan, M. A. (2002). Time series forecasting using wavelet
denoising: An application to Saudi stock index. Journal of King Saud University,
Engineering Sciences, 2(14), 221–234.
Anderson, T. W. (1971). The statistical analysis of time series. New York: John Wiley &
Sons.
Antoniadis, A. (1999). Wavelets in statistics: A review. Journal of the Italian Statistical
Society.
Arino, M. (1995). ‘‘Time Series Forecasts via Wavelets: An Application to Car Sales in
the Spanish Market’’. Discussion Paper No. 95–30, Institute of Statistics and
Decision Sciences, Duke University.
3873
Avci, Engin, & Derya, A. (2008). A novel approach for digital radio signal
classification: Wavelet packet energy–multiclass support vector machine
(WPE–MSVM). Expert Systems with Applications, 34(3), 2140–2147.
Bowerman, B. L., O’Connell, R. T., & Koehler, A. B. (2005). ‘‘Forecasting, time series,
and regression: An applied approach’’. South-Western.
Burt, P. J., & Adelson, E. H. (1983). The Laplacian pyramid as a compact image code.
IEEE Transactions on Communications, 31, 532–540.
Chang, S. I., & Yadama, S. (2010). Statistical process control for monitoring non-
linear profiles using wavelet filtering and B-spline approximation. International
Journal of Production Research, 48(4), 1049–1068.
Chen, G. (2014). Automatic EEG seizure detection using dual-tree complex wavelet-
Fourier features. Expert Systems with Applications, 41, 2391–2394.
Chen, H., Nicolis, O., & Vidakovic, B. (2010). Multiscale forecasting method using
ARMAX models. Current Development in Theory and Applications of Wavelets, 4,
267–287.
Chen, N., Qian, Z., & Meng, X. (2013). Multistep wind speed forecasting based on
wavelet and gaussian processes. Mathematical Problems in Engineering, 2013,
1–8.
Chou, Chien-Ming (2014). Application of set pair analysis-based similarity forecast
model and wavelet denoising for runoff forecasting. Water, 6, 912–928.
Croarkin, C., & Tobias, P. (2006). NIST/Sematech e-Handbook of Statistical Methods.
U.S. Commerce Department. Available at <http://www.itl.nist.gov/div898/
handbook/index.htm>.
Di, C., Yang, X., & Wang, X. (2014). A four-stage hybrid model for hydrological time
series forecasting. PLOS ONE, 9(8), e104663.
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive
time series with a unit root. Journal of the American Statistical Association, 74,
427–431.
Donoho, D. L., & Johnstone, I. M. (1994). Ideal denoising in an orthonormal basis
chosen from a library of basis. Comptes Rendus de l’Académie des Sciences Paris,
319, 1317–1322.
Ferbar, L., Čreslovnik, D., Mojškerc, B., & Rajgelj, M. (2009). Demand forecasting
methods in a supply chain: Smoothing and denoising. International Journal of
Production Economics, 118, 49–54.
Graps, A. (1995). An introduction to wavelets. IEEE Computational Science and
Engineering, 2(2), 50–61.
Guo, X., Linyan, S., Gang, L., & Song, W. (2008). A hybrid wavelet analysis and
support vector machines in forecasting development of manufacturing. Expert
Systems with Applications, 35(1), 415–422.
Hadaś-Dyduch, M. (2014). Non-classical algorithm for time series prediction of the
range of economic phenomena with regard to the interaction of financial
market indicators. Chinese Business Review, 13(4), 221–231.
Jeong, M. K., Lu, J. C., & Wang, N. (2006). Wavelet based SPC procedure for
complicated functional data. International Journal of Production Research, 44(4),
729–744.
Jin, J., & Shi, J. (2001). Automatic feature extraction of waveform signals for in-
process diagnostic performance improvement. Journal of Intelligent
Manufacturing, 12(3), 267–268.
Lada, E. K., Lu, J. C., & Wilson, J. R. (2002). A wavelet-based procedure for
process fault detection. IEEE Transaction on Semiconductor Manufacturing, 15(1),
79–90.
Lai, C. P., Chung, P. C., & Tseng, V. S. (2010). A novel two-level clustering method for
time series data analysis. Expert Systems with Applications, 37(9), 6319–6326.
Li, G., Gai, Z., Kang, X., Wu, Z., & Wang, Y. (2014). ESPSA: A prediction-based
algorithm for streaming time series. Expert Systems with Applications, 41,
6098–6105.
Lee, W. W (2009). Statistical methods for prediction (2nd ed.). Free Academy.
Mallat, S. G. (1989). A theory for multi resolution signal decomposition: The wavelet
representation. IEEE Transaction on Pattern Analysis and Machine Intelligence, 11,
674–693.
Meerwald, P., & Uhl, A. (2001). ‘‘A survey of wavelet-domain watermarking
algorithms’’. In Proceedings of SPIE, electronic imaging, security and
watermarking of multimedia contents III, 4314 (pp. 505–516), USA.
Morettin, P. A. (1996). ‘‘From fourier to wavelet analysis of time series’’. In A. Pratt
(Ed.), Proceedings of computational statistics (pp. 111–122).
Morettin, P. A. (1997). Wavelets in statistics. Resenhas, 3, 211–272.
Mousa, R., Munib, Q., & Moussa, A. (2005). Breast cancer diagnosis system based on
wavelet analysis and fuzzy-neural. Expert Systems with Applications, 28(4),
713–723.
Nguyen, T., Khosravi, A., Creighton, D., & Nahavandi, S. (2015). Classification of
healthcare data using genetic fuzzy logic system and wavelets. Expert Systems
with Applications, 42, 2184–2197.
Percival, D. B., & Walden, A. T. (1999). Wavelet methods for time series analysis.
Cambridge: Cambridge University Press.
Popoola, A. O. (2007). Fuzzy-Wavelet Method for Time Series Analysis, Disertasi,
Department of Computing, School of Electronics and Physical Sciences,
University of Surrey, Surrey.
Prasad, L., & Iyengar, S. S. (1997). Wavelet analysis with applications to image
processing. CRC Press.
Priestley, M. B. (1996). Wavelets and time-dependent spectral analysis. Journal of
Time Series Analysis, 17, 85–104.
Ramsey, J. B. (1999). The contribution of wavelets to the analysis of economic and
financial data. Philosophical Transactions of the Royal Society of London A, 357,
2593–2606.
Rao, R. M., & Bopardikar, A. S. (1998). Wavelet transforms: Introduction to theory and
applications. Addison Wesley Longman.
3874 T.W. Joo, S.B. Kim / Expert Systems with Applications 42 (2015) 3868–3874
Renaud, O., Starck, J. L., & Murtagh, F. (2002). ‘‘Wavelet-Based Forecasting of Short
and Long Memory Time Series.’’ Université de Genève, Faculté des sciences
économiques et sociales, Département d’économétrie.
Rocha, T., Paredes, S., Carvalho, P., Henriques, J., & Harris, M. (2010). ‘‘Wavelet-based
time series forecast with application to acute hypotensive episodes prediction’’.
In Engineering in medicine and biology society (EMBC), 2010 annual international
conference of the IEEE, 2403–2406.
Sang, Y. F. (2013). Improved wavelet modeling framework for hydrologic time series
forecasting. Water Resources Management, 27(8), 2807–2821.
Saravanan, N., & Ramachandran, K. I. (2009). Fault diagnosis of spur bevel gear box
using discrete wavelet features and decision tree classification. Expert Systems
with Applications, 36(5), 9564–9573.
Schlüter S., & Deuschle, C. (2010). ‘‘Using wavelets for time series forecasting? Does
it pay off?’’ IWQW, University of Erlangen-Nuremberg.
Soltani, S. (2002). On the use of wavelet decomposition for time series prediction.
Neurocomputing, 48, 267–277.
Subasi, A. (2007). EEG signal classification using wavelet feature extraction and a
mixture of expert model. Expert Systems with Applications, 32(4), 1084–1093.
Tabak, B. M., & Feitosa, M. A. (2010). Forecasting industrial production in Brazil:
Evidence from a wavelet approach. Expert Systems with Applications, 37(9),
6345–6351.
Tran, L. (2006). ‘‘From fourier transforms to wavelet analysis: Mathematical
concepts and examples’’.
Vidakovic, B. (1999). Statistical modeling by wavelets. New York: John Wiley & Sons.
Wang, C. C., Kang, Y., Shen, P. C., Chang, Y. P., & Chung, Y. L. (2010). Applications of
fault diagnosis in rotating machinery by using time series analysis with neural
network. Expert Systems with Applications, 37(2), 1696–1702.
Wang, J. (2014). A hybrid wavelet transform based short-term wind speed
forecasting approach. The Scientific World Journal.
Weigend, A. S., & Mangeas, M. (1995). Nonlinear gated experts for time series:
Discovering regimes and avoiding overfitting. International Journal of Neural
Systems, 6, 373–399.
Wong, H., Wai-Cheung, I., Zhongjie, X., & Lui, X. (2003). Modelling and forecasting
by wavelets, and the application to exchange rates. Journal of Applied Statistics,
30, 537–555.
Zhang, B. L., Coggins, R., Jabri, M. A., Dersch, D., & Flower, B. (2001). Multiresolution
forecasting for futures trading using wavelet decompositions. IEEE Transactions
on Neural Networks, 12, 765–775.