Econometrics of Financial Markets

Markus Haas

Summer term 2016

April 9, 2016
 Econometrics of Financial Markets

• Markus Haas, markus.haas@qber.uni-kiel.de

• Sebastian Müller, s.mueller@qber.uni-kiel.de

• For organizational/exam issues: info@qber.uni-kiel.de (Ms Ethel

Fritz)

• Course materials (slides, old exams) in OLAT

• MATLAB course (“Übung”): Sebastian Müller

1
 Econometrics of Financial Markets

• written exam, more information to follow

• usual registration procedure

2
 Course Outline

• Introduction

• Properties of financial returns

• Time series concepts

• Statistical models for returns:

(i) Univariate GARCH models and extensions
(ii) Dynamic conditional score models: The Beta–t–EGARCH process
(iii) Stochastic volatility models
(iv) Realized volatility
(v) Multivariate GARCH

3
 Textbooks

• Alexander, C. (2008), Practical Financial Econometrics, Wiley, New York

• Christoffersen, P. (2012), Elements of Financial Risk Management, 2e,

Elsevier

• Franke/Härdle/Hafner (2015). Statistics of Financial Markets, Springer.

• Harvey, A. C. (2013). Dynamic Models for Volatility and Heavy Tails,

Cambridge University Press.

• McNeil/Frey/Embrechts (2015): Quantitative Risk Management,

Revised Edition, Princeton University Press.

• Mills, T. C., Markellos (2008), The Econometric Modelling of Financial

Time Series, 2e, Cambridge University Press.

• Taylor, S. (2005), Asset Price Dynamics, Volatility, and Prediction,

Princeton University Press, New Jersey.

4
 Textbooks

• Tsay, R.S. (2010), Analysis of Financial Time Series, 2e, Wiley, New
York.

5
 Other useful books

• Andersen/Davis/Kreiß/Mikosch, eds. (2009), Handbook of Financial

Time Series, Springer

• Bauwens/Hafner/Rombouts, eds. (2012), Handbook of Volatility Models

and Their Applications, Wiley

6
 Asset Return Definitions

• Let Pt be the price of an asset at time t (stock, stock index, exchange

rate,...).

• If there are no dividends or other payments, the (single–period) discrete

net return is
Pt − Pt−1 Pt
Rt = = − 1, (1)
Pt−1 Pt−1
whereas the discrete gross return is

Pt
1 + Rt = . (2)
Pt−1

• Often the returns defined in (1) and (2) are multiplied by 100 to be
interpretable in terms of percentage returns.

7
 Asset Return Definitions

• If an asset pays a dividend, Dt, between time t − 1 and t, then the return
is
Pt + Dt − Pt−1 Pt − Pt−1 Dt
Rt = = + , (3)
Pt−1 Pt−1 Pt−1
where the term (Pt − Pt−1)/Pt−1 is the capital gain, and Dt/Pt−1 is the
dividend yield.

8
 Asset Return Definitions

• The continuously compounded or log–return of an asset is

 
Pt
rt = log = log Pt − log Pt−1 = log(1 + Rt), (4)
Pt−1

and the percentage log–return is (4) multiplied by 100.

• Its name derives from the fact that (4) can be rewritten as

Pt = Pt−1ert , (5)

i.e., rt is the rate of return under continuous compounding between time

t − 1 and time t, i.e.,
h  r t n i
Pt−1 lim 1+ = Pt−1ert = Pt. (6)
n→∞ n

9
 Asset Return Definitions

• Since ex ≥ 1 + x, we always have

rt = log(1 + Rt) ≤ Rt. (7)

• For small returns,1

Rt2 Rt3 Rt4

rt = log(1 + Rt) = Rt − + − + · · · ≈ Rt , (8)
2 3 4

so that rt may serve (and is often used in practice) as a reasonable

approximation to the discrete return Rt.

1
Note that expansion (8) is valid only for Rt ∈ (−1, 1]; the approximation is very good at least for Rt
between −0.1 (–10%) and 0.1 (10%), cf. Table 1.

10
 Asset Return Definitions

• E.g., daily returns are very rarely outside the range from −10% to 10%,
and then the choice of the return definition will be of minor importance.

Table 1: Discrete and continuous returns

100 × Rt –1.00 –2.50 –5.00 –10.00 –15.00
100 × log(1 + Rt) –1.01 –2.53 –5.13 –10.54 –16.25
100 × Rt 1.00 2.50 5.00 10.00 15.00
100 × log(1 + Rt) 1.00 2.47 4.88 9.53 13.98
Continuous percentage return rt = 100 × log(1 + Rt).

11
 discrete return Rt
0.25
log-return log(1 + Rt )
0.2

0.15

0.1
Rt , log(1 + Rt )

0.05

−0.05

−0.1

−0.15

−0.2

−0.25

−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25
Rt

12
 Asset Return Definitions

• Both return definitions have their advantages and disadvantages.

• E.g., in time series analysis, continuously compounded returns are often

used because multi–period returns are sums of single–period returns.

• That is, if rt(τ ) denotes the (multi–period) return from time t − τ to

time t, we have
 
Pt
rt(τ ) = log
Pt−τ
 
Pt Pt−1 Pt−τ +1
= log ···
Pt−1 Pt−2 Pt−τ
     
Pt Pt−1 Pt−τ +1
= log + log + · · · + log
Pt−1 Pt−2 Pt−τ
τ
X −1
= rt−i.
i=0

13
 Asset Return Definitions

• This is not the case for the discrete return, where

τY
−1
1 + Rt(τ ) = (1 + Rt−i), (9)
i=0

and sums of random variables can be more easily handled than products.

• Also, due to limited liability, most asset prices have a lower bound of
zero, i.e., Rt ≥ −1, which is easier to preserve in models based on
continuous returns.2

2
However, distributions with unbounded support, such as the normal, may still be used as an
approximation for discrete returns, since, when fitted to return data, the implied probability of a loss larger
than 100% will be negligible. The same reasoning of course applies to basically any application of such
distributions to real world data, which are usually restricted to a finite-length interval.

14
 Asset Return Definitions

• Discrete returns have a very convenient property in portfolio analysis not

shared by continuous returns:

• Let
– Pit be the price of asset i at time t, and
– Rit be the return of asset i,
i = 1, . . . , N , where N is the number of assets in the portfolio.

• At time t − 1, we have ni units of asset i in the portfolio.

15
 Asset Return Definitions

• The value of the portfolio at time t − 1, Pp,t−1, is thus

N
X
Pp,t−1 = ni Pi,t−1, (10)
i=1

and asset i’s portfolio weight (i.e., the fraction of wealth invested in
asset i) is
niPi,t−1
xi = N , i = 1, . . . , N, (11)
P
nj Pj,t−1
j=1
with
N
X
xi = 1. (12)
i=1

16
 Asset Return Definitions

• The portfolio return Rp,t is given by

Pp,t − Pp,t−1
Rt,p =
Pp,t−1
P P
n
i iP P it − i ni Pi,t−1
=
i ni Pi,t−1
P
inPi (Pit − Pi,t−1)
=
i ni Pi,t−1

XN
niPi,t−1 Pit − Pi,t−1
= P
i=1 i ni Pi,t−1 Pi,t−1
N
X
= xiRi,t.
i=1

• That is, the portfolio return is a linear combination of the components’

returns.

17
 Asset Return Definitions

• For the continuous return, in general,

 
Pp,t
rp,t = log = log(1 + Rp,t)
Pp,t−1
!
X
= log xi(1 + Rit)
i
X
6= xi log (1 + Rit)
i
N
X
= xirit,
i=1

i.e., the continuously compounded portfolio return is not a linear

combination of continuously compounded asset returns.

18
 Asset Return Definitions

• However, for small returns, the difference is again moderate, and the
approximation
N
X
rp,t ≈ xirit (13)
i=1
is also frequently used.

19
 Basic Statistical Properties of Returns: Marginal
(Unconditional) Return Distribution
• Due to the uncertain nature of the returns of speculative assets, they are
best treated as random variables.

• The traditional assumption that has long dominated empirical finance

was that log–returns over “longer” time intervals (i.e., daily or longer)
are approximately normally distributed.

• The rationale behind this view is as follows:

• Daily log–returns, for example, are the sum of a large number of intraday
returns.

• Appealing to the central limit theorem, Osborne (1959) argued in a

classical article that3
“under fairly general conditions [...] we can expect that the distribution
function of [rt] will be normal.”
3
Brownian Motion in the Stock Market, Operations Research 7, 145-173.

20
 Marginal (Unconditional) Return Distribution

• To quickly (and informally) assess the appropriateness of the normal

distribution, a kernel density estimate can be used.

• The kernel density estimator is a nonparametric estimator of the density

and can be viewed as a “smoothed” form of the histogram.

• It looks more like a probability density function than a histogram and

does not depend on the number and location of bins.

• This smoothed histogram can then be used to compare the empirical

distribution with a fitted normal distribution (or any other hypothesized
distribution).

21
 Kernel Density Estimation

• Want to estimate the density of random variable X.

• The density of X, f (x), is the derivative of the cdf, F (x),

F (x) = Pr(X ≤ x), (14)

that is,
1
f (x) = lim Pr(x − h < X < x + h). (15)
h→0 2h

• Thus we may attempt to estimate the density by choosing a small h and

setting

b 1
f (x) = {number of observations in (x − h, x + h)} , (16)
2hT

where T is the size of our sample.

22
 Kernel Density Estimation

• For a sample X1, X2, . . . , XT , estimator (16) can be written as

XT  
1 1 x − Xt
fb(x) = w , (17)
T t=1 h h

where the weight function

(
1
2 if |x| < 1
w(x) = (18)
0 otherwise

• That is, the estimator is obtained by placing a box of width 2h and

height 2T1 h on each observation and then adding them up.

• How does this look like?

23
• This has been produced using a (pseudo) random sample of size T = 100
from a N(0,1) distribution and h = 12 .

0.45
estimate (h = 0.5)
0.4
N(0,1) density

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−4 −3 −2 −1 0 1 2 3 4

24
 Kernel Density Estimation

• The density estimator (16) has jumps at points Xt ± h, t = 1, . . . , T and

zero derivative everywhere else, leading to a ragged picture which may
be undesirable.

• An intuitive motivation for the kernel estimator is thus to modify the

estimator above by replacing the weight function (18) with a “nicer”
kernel function K(x) which leads to a smooth estimator.

• E.g., a Gaussian kernel is often used, such that

K(x) = φ(x) = (2π)−1/2 exp{−x2/2}. (19)

R
• In general, the kernel must be a density, i.e., K(x) ≥ 0 and K(x)dx =
1.

25
 Kernel Density Estimation

• The kernel density estimator at a point x is

XT  
1 1 x − X t
fb(x) = K , (20)
T t=1 h h

where K is the kernel function.

• Parameter h is the bandwidth or smoothing parameter which determines

the smoothness of the estimator.

• Larger h ⇒ smoother estimate.

• For example, consider daily log–returns of the DAX 30 index from January
2000 to May 2013 (T = 3477 observations), with h = 0.05 and h = 0.5.

26
 density of daily DAX returns
0.6
kernel (h = 0.05)

0.5

0.4

0.3

0.2

0.1

0
−8 −6 −4 −2 0 2 4 6 8

27
 density of daily DAX returns
0.6
kernel (h = 0.05)
kernel (h = 0.5)
0.5

0.4

0.3

0.2

0.1

0
−8 −6 −4 −2 0 2 4 6 8

28
 Kernel Density Estimation

• How to choose h?

• A small h allows the density estimator to capture the fine structure of

the true density, but it also increases random variation.

• On the other hand, a large h reduces random variation but may hide
important details of the density.

• Thus, as is often the case in statistics, there is a bias/variance trade–off.

29
 Kernel Density Estimation

• Trying to choose h optimally (in the MSE4 sense) shows that the optimal
h depends on T (the sample size), the kernel, and, unfortunately but not
surprisingly, the true density (which is unknown).

• However, it turns out5 that, with a Gaussian kernel,6 choosing

h = 0.9AT −1/5 (21)

gives good results for a wide range of densities, where

A = min{sample standard deviation, interquartile range/1.34}, (22)

and the interquartile range is the difference between the 0.75–quantile

and the 0.25–quantile.
4
Mean squared error.
5
See the discussion in Silverman (1986): Density Estimation for Statistics and Data Analysis, London:
Chapman and Hall, Chapter 3.
6
And a random (iid) sample, which will not hold for most financial data.

30
 S&P 500 index level (daily), January 2000 to October 2011
130

120

110

100

90

80

70

60

50
2000 2002 2004 2006 2008 2010 2012

31
 S&P 500 index returns (daily), January 2000 to October 2011
15

10

−5

−10
2000 2002 2004 2006 2008 2010 2012

32
 DAX 30 index returns (daily), January 2000 to October 2011
15

10

−5

−10
2000 2002 2004 2006 2008 2010 2012

33
 Kenrel density estimate of S&P 500 return density
0.5

0.45

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−8 −6 −4 −2 0 2 4 6 8

34
 Kernel Density Estimation

• We compare the kernel estimate with a fitted normal, i.e., with the
density  
2
1 (x − µb)
f (x) = √ exp − 2
,
2πb
σ 2bσ
where
T
1X
µ
b=r= rt
T t=1
and
XT
1
b 2 = s2 =
σ (rt − r)2
T − 1 t=1
are the sample mean and variance, respectively.

35
 Density of S&P 500 returns
0.5
empirical (kernel)
0.45 fitted normal

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
−8 −6 −4 −2 0 2 4 6 8

36
The log–density helps to better detect differences in the tails:

0
Log−Density of S&P 500 returns
10
empirical (kernel)
10
−1
fitted normal

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10

−8
10

−9
10
−8 −6 −4 −2 0 2 4 6 8

37
 Density of DAX 30 returns
0.4
empirical (kernel)
0.35
fitted normal

0.3

0.25

0.2

0.15

0.1

0.05

0
−8 −6 −4 −2 0 2 4 6 8

38
 0
Log−Density of DAX 30 returns
10
empirical (kernel)
fitted normal
−1
10

−2
10

−3
10

−4
10

−5
10

−6
10
−8 −6 −4 −2 0 2 4 6 8

39
Basic Statistical Properties of Returns: Excess Kurtosis
(Thick Tails)

• Financial returns at higher frequencies (higher than a month at least)

are not normally distributed.

• In particular, they have much more probability mass in the center and
the tails (fat tails) than a normal distribution with the same variance.

• The empirical density exhibits excess kurtosis relative to the normal

distribution.

• This implies, among other things, that the probability of large losses is
much higher than under the Gaussian assumption, which is of considerable
interest for many financial applications.

40