Volatility Modelling of Asset Prices

using GARCH Models

Jens Näsström

Reg nr: LiTH-ISY-EX-3364-2003

February 11, 2003
 Volatility Modelling of Asset Prices
using GARCH Models
Master’s Thesis
Division of Automatic Control
Department of Electrical Engineering
Linköping University, Sweden

Jens Näsström

Reg nr: LiTH-ISY-EX-3364-2003

Supervisor: O.Prof. Manfred Deistler

Prof. Lennart Ljung

Examiner: Prof. Lennart Ljung

February 11, 2003

 Avdelning, Institution Datum
Division, Department Date
2003-02-11

Institutionen för Systemteknik

581 83 LINKÖPING

Språk Rapporttyp ISBN

Language Report category
Svenska/Swedish Licentiatavhandling
ISRN LITH-ISY-EX-3364-2003
X Engelska/English X Examensarbete
C-uppsats
Serietitel och serienummer ISSN
D-uppsats
Title of series, numbering
Övrig rapport
____

URL för elektronisk version

http://www.ep.liu.se/exjobb/isy/2003/3364/

Titel Volatilitets prediktering av finansiella tillgångar - med GARCH modeller som ansats
Title
Volatility Modelling of Asset Prices using GARCH Models

Författare Jens Näsström

Author

Sammanfattning
Abstract
The objective for this master thesis is to investigate the possibility to predict the risk of stocks in
financial markets. The data used for model estimation has been gathered from different branches
and different European countries. The four data series that are used in the estimation are price
series from: Münchner Rück, Suez-Lyonnaise des Eaux, Volkswagen and OMX, a Swedish stock
index. The risk prediction is done with univariate GARCH models. GARCH models are estimated
and validated for these four data series.

Conclusions are drawn regarding different GARCH models, their numbers of lags and
distributions. The model that performs best, out-of-sample, is the APARCH model but the standard
GARCH is also a good choice. The use of non-normal distributions is not clearly supported. The
result from this master thesis could be used in option pricing, hedging strategies and portfolio
selection.

Nyckelord
Keyword
GARCH models, risk prediction, system identification and econometrics
Für meine Eltern
Abstract

The objective for this master thesis is to investigate the possibility to predict
the risk of stocks in financial markets. The data used for model estimation
has been gathered from different branches and different European coun-
tries. The four data series that are used in the estimation are price series
from: Münchner Rück, Suez-Lyonnaise des Eaux, Volkswagen and OMX, a
Swedish stock index. The risk prediction is done with univariate GARCH
models. GARCH models are estimated and validated for these four data
series.
Conclusions are drawn regarding different GARCH models, their num-
bers of lags and distributions. The model that performs best, out-of-sample,
is the APARCH model but the standard GARCH is also a good choice. The
use of non-normal distributions is not clearly supported. The result from
this master thesis could be used in option pricing, hedging strategies and
portfolio selection.

Keywords: GARCH models, risk prediction, system identification and eco-

nometrics
Contents

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Presenting the companies . . . . . . . . . . . . . . . . . . . . 1
1.3 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Problem specification . . . . . . . . . . . . . . . . . . . . . . . 2
1.5 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.6 Reader’s guide . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Method 5
2.1 Econometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 System Identification . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 The data set . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Volatility . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.3 Selected models for estimation . . . . . . . . . . . . . 7
2.2.4 Criterion of fit . . . . . . . . . . . . . . . . . . . . . . 7
2.2.5 Validation . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Methods of estimation, toolboxes . . . . . . . . . . . . . . . . 9
2.3.1 The Ox GARCH package . . . . . . . . . . . . . . . . 10

3 Theory 13
3.1 Economic theory . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.1 Financial data . . . . . . . . . . . . . . . . . . . . . . 13
3.1.2 An introduction to risk . . . . . . . . . . . . . . . . . 14
3.1.3 Efficient market theory . . . . . . . . . . . . . . . . . 14
3.1.4 Leverage effect . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.1 Correlation . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2 Autocorrelation . . . . . . . . . . . . . . . . . . . . . . 15
3.2.3 Partial autocorrelation . . . . . . . . . . . . . . . . . . 16
3.3 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1 Jarque-Bera test . . . . . . . . . . . . . . . . . . . . . 16
3.3.2 Ljung-Box test . . . . . . . . . . . . . . . . . . . . . . 17
3.3.3 ARCH test . . . . . . . . . . . . . . . . . . . . . . . . 17
iv Contents

3.3.4 Parameter output and t-test . . . . . . . . . . . . . . . 17

3.3.5 Nelson test . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4 Stochastic processes . . . . . . . . . . . . . . . . . . . . . . . 18
3.4.1 White noise . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4.2 Stationarity . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4.3 Moving average process . . . . . . . . . . . . . . . . . 19
3.4.4 Autoregressive process . . . . . . . . . . . . . . . . . . 19
3.5 Computing volatility . . . . . . . . . . . . . . . . . . . . . . . 19
3.5.1 EWMA - Exponentially weighted moving average . . . 21
3.6 ARCH - Autoregressive conditional heteroskedasticity . . . . 21
3.7 Generalized ARCH - GARCH . . . . . . . . . . . . . . . . . . 22
3.7.1 GARCH(1,1) . . . . . . . . . . . . . . . . . . . . . . . 23
3.7.2 IGARCH . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.7.3 EGARCH . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.7.4 APARCH . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.8 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4 Volkswagen 27
4.1 Presenting data . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.1 Introducing Volkswagen . . . . . . . . . . . . . . . . . 27
4.2 Preestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.1 Data formatting . . . . . . . . . . . . . . . . . . . . . 27
4.2.2 Moments for the return series . . . . . . . . . . . . . . 28
4.2.3 Plots for Volkswagen . . . . . . . . . . . . . . . . . . . 28
4.2.4 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . 29
4.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.1 Choosing p and q . . . . . . . . . . . . . . . . . . . . . 31
4.3.2 Choosing model . . . . . . . . . . . . . . . . . . . . . 32
4.3.3 Choosing distribution . . . . . . . . . . . . . . . . . . 32
4.3.4 Estimated parameters . . . . . . . . . . . . . . . . . . 33
4.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4.1 Parameter time dependence . . . . . . . . . . . . . . . 34
4.4.2 Autocorrelation . . . . . . . . . . . . . . . . . . . . . . 34
4.4.3 Ljung-Box-Pierce Q-test . . . . . . . . . . . . . . . . . 35
4.4.4 Engle’s ARCH-test . . . . . . . . . . . . . . . . . . . . 35
4.4.5 Is zt normally distributed? . . . . . . . . . . . . . . . 36
4.4.6 Skewness of zt . . . . . . . . . . . . . . . . . . . . . . 37
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.5.1 Estimation . . . . . . . . . . . . . . . . . . . . . . . . 37

5 Münchner Rück 39
5.1 Presenting data . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.1 Introducing Münchner Rück . . . . . . . . . . . . . . . 39
5.2 Preestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Contents v

5.2.1 Data formatting . . . . . . . . . . . . . . . . . . . . . 40

5.2.2 Moments for the return series . . . . . . . . . . . . . . 40
5.2.3 Plots for Münchner Rück . . . . . . . . . . . . . . . . 40
5.2.4 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . 40
5.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3.1 Choosing p and q . . . . . . . . . . . . . . . . . . . . . 44
5.3.2 Choosing model . . . . . . . . . . . . . . . . . . . . . 44
5.3.3 Choosing distribution . . . . . . . . . . . . . . . . . . 44
5.3.4 Estimated parameters . . . . . . . . . . . . . . . . . . 45
5.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.4.1 Time dependent parameters . . . . . . . . . . . . . . . 46
5.4.2 Ljung-Box-Pierce Q-test . . . . . . . . . . . . . . . . . 47
5.4.3 Engle’s ARCH-test . . . . . . . . . . . . . . . . . . . . 47
5.4.4 How is zt distributed? . . . . . . . . . . . . . . . . . . 48
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.5.1 Estimation . . . . . . . . . . . . . . . . . . . . . . . . 48

6 Suez-Lyonnaise des Eaux 51

6.1 Presenting data . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.1.1 Introducing Suez-Lyonnaise des Eaux . . . . . . . . . 51
6.2 Preestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.2.1 Data formatting . . . . . . . . . . . . . . . . . . . . . 51
6.2.2 Moments for the return series . . . . . . . . . . . . . . 52
6.2.3 Correlation plots for lyoe . . . . . . . . . . . . . . . . 53
6.2.4 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . 54
6.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.3.1 Choosing p and q . . . . . . . . . . . . . . . . . . . . . 56
6.3.2 Choosing model . . . . . . . . . . . . . . . . . . . . . 57
6.3.3 Choosing distribution . . . . . . . . . . . . . . . . . . 57
6.3.4 Estimated parameters . . . . . . . . . . . . . . . . . . 58
6.3.5 Are the coefficients time dependent? . . . . . . . . . . 58
6.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.4.1 Parameter time dependency . . . . . . . . . . . . . . . 60
6.4.2 Autocorrelation . . . . . . . . . . . . . . . . . . . . . . 60
6.4.3 How is zt distributed? . . . . . . . . . . . . . . . . . . 61
6.4.4 Ljung-Box-Pierce Q-test . . . . . . . . . . . . . . . . . 61
6.4.5 Engle’s ARCH-test . . . . . . . . . . . . . . . . . . . . 63
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.5.1 Estimation . . . . . . . . . . . . . . . . . . . . . . . . 63

7 OMX index 65
7.1 Presenting data . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.1.1 Introducing OMX . . . . . . . . . . . . . . . . . . . . 65
7.2 Preestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
vi Contents

7.2.1 Data formatting . . . . . . . . . . . . . . . . . . . . . 65

7.2.2 Moments for the return series . . . . . . . . . . . . . . 65
7.2.3 Correlation plots for OMX . . . . . . . . . . . . . . . 66
7.2.4 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . 66
7.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.3.1 Choosing p and q . . . . . . . . . . . . . . . . . . . . . 69
7.3.2 Choosing model . . . . . . . . . . . . . . . . . . . . . 70
7.3.3 Choosing distribution . . . . . . . . . . . . . . . . . . 71
7.3.4 Estimated parameters . . . . . . . . . . . . . . . . . . 71
7.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.4.1 Parameter time dependence . . . . . . . . . . . . . . . 71
7.4.2 Autocorrelation . . . . . . . . . . . . . . . . . . . . . . 72
7.4.3 Ljung-Box-Pierce Q-test . . . . . . . . . . . . . . . . . 72
7.4.4 Engle’s ARCH-test . . . . . . . . . . . . . . . . . . . . 73
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.5.1 Estimation . . . . . . . . . . . . . . . . . . . . . . . . 75

8 Conclusions 77
8.1 GARCH modelling discussion . . . . . . . . . . . . . . . . . . 77
8.1.1 Testing for different (p,q) . . . . . . . . . . . . . . . . 77
8.1.2 Testing different models . . . . . . . . . . . . . . . . . 77
8.1.3 Testing different distributions . . . . . . . . . . . . . . 78
8.1.4 Parameter instability . . . . . . . . . . . . . . . . . . . 78
8.1.5 User’s choice . . . . . . . . . . . . . . . . . . . . . . . 78
8.1.6 The OMX index . . . . . . . . . . . . . . . . . . . . . 78
8.1.7 Parameter values . . . . . . . . . . . . . . . . . . . . . 78
8.2 Further Studies . . . . . . . . . . . . . . . . . . . . . . . . . . 79
8.2.1 Different regimes . . . . . . . . . . . . . . . . . . . . . 79
8.3 Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . 80
8.4 How good are the estimated models? . . . . . . . . . . . . . . 80

References 83
List of Figures

4.1 Price plot for Volkswagen. . . . . . . . . . . . . . . . . . . . . 29

4.2 Squared return for Volkswagen. . . . . . . . . . . . . . . . . . 29
4.3 Autocorrelation of the squared return for Volkswagen. . . . . 30
4.4 Partial autocorrelation of the squared return for Volkswagen. 30
4.5 Autocorrelation for Volkswagen with residuals after estimation. 35

5.1 Price plot for Münchner Rück . . . . . . . . . . . . . . . . . . 41

5.2 Return plot for Münchner Rück . . . . . . . . . . . . . . . . . 41
5.3 Squared returns for Münchner Rück . . . . . . . . . . . . . . 42
5.4 Autocorrelation on the squared return for muvg. . . . . . . . 42
5.5 Partial autocorrelation on the squared return for muvg. . . . 43

6.1 Price plot for Suez-Lyonnaise des Eaux . . . . . . . . . . . . . 53

6.2 Return plot for Suez-Lyonnaise des Eaux . . . . . . . . . . . 53
6.3 Squared return for lyoe. . . . . . . . . . . . . . . . . . . . . . 54
6.4 Autocorrelation on the squared return for lyoe. . . . . . . . . 54
6.5 Partial autocorrelation on the squared return for lyoe. . . . . 55
6.6 Time dependent coefficient plot for lyoe. Window size = 200. 59
6.7 Time dependent coefficient plot for lyoe. Window size = 700. 59
6.8 Autocorrelation for lyoe after estimation. . . . . . . . . . . . 61
6.9 Histogram for lyoe after estimation. . . . . . . . . . . . . . . 62

7.1 Price plot for OMX. . . . . . . . . . . . . . . . . . . . . . . . 67

7.2 Return plot for OMX. . . . . . . . . . . . . . . . . . . . . . . 67
7.3 Squared return for OMX. . . . . . . . . . . . . . . . . . . . . 68
7.4 Autocorrelation on the squared return for OMX. . . . . . . . 68
7.5 Partial autocorrelation on the squared return for OMX. . . . 69
7.6 Autocorrelation for OMX with normalised residuals after es-
timation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7.7 Histogram for OMX after estimation. . . . . . . . . . . . . . . 74
List of Tables

2.1 GARCH features comparison. . . . . . . . . . . . . . . . . . . 11

3.1 Ljung-Box-Pierce Q-test example. . . . . . . . . . . . . . . . . 17

3.2 Example table for parameter test. . . . . . . . . . . . . . . . . 18
3.3 Time series identification with the ACF and PACF. . . . . . 20

4.1 Days with high return in the Volkswagen series. . . . . . . . . 28

4.2 Moments, skewness and kurtosis for the return series of Volk-
swagen. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Ljung-Box-Pierce Q-test for Volkswagen. . . . . . . . . . . . . 31
4.4 Engle’s ARCH-test for Volkswagen. . . . . . . . . . . . . . . . 31
4.5 Selecting p,q for Volkswagen. . . . . . . . . . . . . . . . . . . 32
4.6 GARCH model selection for Volkswagen. . . . . . . . . . . . . 32
4.7 Distribution selection for Volkswagen. . . . . . . . . . . . . . 33
4.8 Estimated parameters for the chosen model for Volkswagen. . 33
4.9 Estimated parameters for the chosen model for Volkswagen. . 34
4.10 Ljung-Box-Pierce Q-test, validation for Volkswagen. . . . . . 35
4.11 Engle’s ARCH-test, validation for Volkswagen. . . . . . . . . 36
4.12 Moments, skewness and kurtosis for zt of Volkswagen. . . . . 36
4.13 Results from the jb-test. Normality test for zt . . . . . . . . . 37
4.14 Skewness for zt for Volkswagen. . . . . . . . . . . . . . . . . . 37

5.1 Days with high return in the muvg series. . . . . . . . . . . . 39

5.2 Moments, skewness and kurtosis for the return series of muvg. 40
5.3 Ljung-Box-Pierce Q-test for muvg. . . . . . . . . . . . . . . . 42
5.4 Engle’s ARCH-test for muvg. . . . . . . . . . . . . . . . . . . 43
5.5 Selecting p,q for muvg. . . . . . . . . . . . . . . . . . . . . . . 44
5.6 GARCH model selection for muvg. . . . . . . . . . . . . . . . 45
5.7 APARCH model selection for muvg. . . . . . . . . . . . . . . 45
5.8 Distribution selection for muvg. . . . . . . . . . . . . . . . . . 46
5.9 Estimated parameters for the chosen model for muvg. . . . . 46
5.10 Estimated parameters for muvg on different data sets. . . . . 47
5.11 The Ljung-Box-Pierce Q-test in the validation for muvg. . . . 48
5.12 Engle’s ARCH-test in validation for muvg. . . . . . . . . . . . 48
x List of Tables

5.13 Moments, skewness and kurtosis for zt of muvg. . . . . . . . . 48

6.1 Days with high absolute returns in the lyoe series. . . . . . . 52

6.2 Moments, skewness and kurtosis for the return series of lyoe. 52
6.3 Ljung-Box-Pierce Q-test for lyoe. . . . . . . . . . . . . . . . . 55
6.4 Engle’s ARCH-test for lyoe. . . . . . . . . . . . . . . . . . . . 56
6.5 Selecting p,q for lyoe. . . . . . . . . . . . . . . . . . . . . . . 56
6.6 GARCH model selection for lyoe. . . . . . . . . . . . . . . . . 57
6.7 Distribution selection for lyoe. . . . . . . . . . . . . . . . . . . 57
6.8 Estimated parameters for the chosen model for lyoe. . . . . . 58
6.9 Estimated parameters for lyoe on two different data sets. . . 60
6.10 Estimated parameters for lyoe on two different samples of the
symmetric GARCH(1,1) model. . . . . . . . . . . . . . . . . . 61
6.11 Moments, skewness and kurtosis for zt of lyoe. . . . . . . . . 62
6.12 Ljung-Box-Pierce Q-test on validation data for lyoe. . . . . . 62
6.13 Engle’s ARCH-test, validation for lyoe. . . . . . . . . . . . . . 63

7.1 Days with high return in the OMX index. . . . . . . . . . . . 66

7.2 Moments, skewness and kurtosis for the return series of OMX. 66
7.3 Ljung-Box-Pierce Q-test for the OMX. . . . . . . . . . . . . . 68
7.4 Engle’s ARCH-test for OMX. . . . . . . . . . . . . . . . . . . 69
7.5 Selecting p,q for OMX. . . . . . . . . . . . . . . . . . . . . . . 70
7.6 GARCH model selection for OMX. . . . . . . . . . . . . . . . 70
7.7 Distribution selection for OMX. . . . . . . . . . . . . . . . . . 71
7.8 Estimated parameters for the chosen model for OMX. . . . . 72
7.9 Estimated parameters for the chosen model for OMX. . . . . 72
7.10 Ljung-Box-Pierce Q-test, validation for OMX. . . . . . . . . . 73
7.11 Engle’s ARCH-test, validation for OMX. . . . . . . . . . . . . 73
7.12 Moments, skewness and kurtosis for zt of OMX. . . . . . . . . 74

8.1 Parameter comparison. . . . . . . . . . . . . . . . . . . . . . . 80

8.2 Model explanation power overview for Volkswagen models. . . 81
8.3 Model explanation power overview for muvg models. . . . . . 81
8.4 Model explanation power overview for lyoe models. . . . . . . 82
8.5 Model explanation power overview for OMX models. . . . . . 82
Acknowledgements

First I would like to thank my supervisors and examiner Prof. Manfred

Deistler and Prof. Lennart Ljung for letting me have the opportunity to
write my master thesis here in Vienna. I would also like to thank the staff
at the Department of Econometrics and System Theory for making this a
pleasant stay.
I am also grateful to Eva Hamann and Dietmar Bauer, at the depart-
ment, for interesting discussions about econometrics in general and GARCH
models in particular.
Finally my thanks goes to my opponent Emil Tirén for useful comment
and suggestions, to Anders Blomqvist and Claes Wallin for proofreading and
to Thomas Schön for letting me use parts of his LATEX framework.

Vienna, December 2002

Jens Näsström
Notation

In this chapter symbols, operators and functions are explained. Abbrevia-

tions, both technical and economical, are printed out.

Symbols
X A discrete stochastic variable.
xt Stochastic process.
µ Mean value of a stochastic variable.
ψt The information set available at time t.

Operators and functions

L The
Pq lag operator
A(L) i
i=1 αi L
E(.) The expectation operator
xiv Notation

Abbreviations
ACF AutoCorrelation Function.
AIC Akaike’s Information Criterion.
AMAPE Adjusted Mean Absolute Percentage Error.
APARCH Asymmetric Power ARCH.
AR AutoRegressive.
ARMA AutoRegressive Moving Average.
ARCH AutoRegressive Conditional Heteroskedasticity.
ARX AutoRegressive with eXternal input.
DF Degrees of Freedom.
EGARCH Exponential GARCH.
EWMA Exponentially Weighted Moving Average.
GARCH Generalized AutoRegressive Conditional Heteroskedasticity.
IGARCH Integrated GARCH.
i.i.d. Identically Independently Distributed.
jb Jarque-Bera.
ks Kolmogorov-Smirnov.
Lbq-test Ljung-Box-Pierce Q-test.
Log Log-likelihood function value.
MA Moving Average.
MedSE Median Squared Error.
MAE Mean Absolute Error.
MLE Maximum Likelihood Estimation.
MSE Mean Squared Error.
OLS Ordinary Least Squares.
PACF Partial AutoCorrelation Function.
TIC Theil Inequality Coefficient.
w.s.s. Wide-Sense Stationary.

DE XETRA - Germany Stock Exchange.

LYOE Suez-Lyonnaise des Eaux.
MUVG Münchener Rückversicherungs-Gesellschaft AG.
PA Paris Stock Exchange.
VOWG Volkswagen.
Xetra Exchange Electronic Trading.
Chapter 1

Introduction

”October. This is one of the peculiarly dangerous months to speculate in

stocks in. The others are July, January, September, April, November, May,
March, June, December, August and February.” (Twain 1894)

1.1 Background
Measuring the risk on specific assets has become increasingly important
during the last decades. In a broad sense companies want to have good
control of their risk profile. In the financial market this is of even greater
importance. A number of large financial companies have had large losses
during the last years. One example is the failure of Long Term Capital
Management.
Risk can be divided into different subcategories: Strategic risks, oper-
ational risks and financial risks. In this thesis we focus on risks on stock
markets. Calculating risks is important for pricing derivatives, portfolio
selection and hedging strategies.

1.2 Presenting the companies

The data used for estimation is taken from different branches and differ-
ent European countries. Just to give an overview of the companies a brief
presentation is given below.

Münchner Rück’s stock price is taken from the Germany stock market
Xetra, (DE). The data was sampled from the 25th of February 1999 to
the 21st of October 2002. Münchner Rück is one of the world’s largest
reinsurance companies.

Suez-Lyonnaise des Eaux is traded at the Paris Stock Exchange (PA).

The data was sampled from the 9th of June 1999 to the 17th of Septem-
2 Introduction

ber 2002. Suez-Lyonnaise des Eaux is a big water power plant com-
pany.

Volkswagen’s stock price is taken from the Germany stock market (DE).
The data was sampled from the 25th of February 1999 to the 21st of
October 2002. Volkswagen is a large producer of cars and vans.

OMX is an index that consists of 30 large Swedish companies. The data

was sampled from the 5th of January 1998 to the 21st of March 2002.
The turnover in the Swedish stock market is 109 Euro/day.

The companies and the data will be further presented in the corresponding
preestimation section for each stock or index.

1.3 Objective
The objective for this thesis is to investigate to what extent it is possible to
predict the risk in stocks in financial markets.

1.4 Problem specification

The specific GARCH modelling questions that are treated in this thesis are:

• Do these data sets exhibit conditional heteroskedasticity?

• What number of lags is a good model choice?

• Which GARCH model does a good job in forecasting volatility for a

specific choice of p and q?

• Do the parameters change in time?

1.5 Limitations
To be able to give reliable results it is necessary to limit the problem and
focus on a few specific topics. In this thesis we focus exclusively on univariate
modelling of volatility. The multivariate case is not treated at all. Another
limitation is the mean equation, which is modelled with just a constant.
This is consistent with the efficient market theory, except for the fact that
investors also require extra return for taking risk. We are working with daily
data and are only interested in one-step-ahead prediction. The model set
only consists of different GARCH models.
Reader’s guide 3

1.6 Reader’s guide

Each chapter in the thesis is briefly presented below. The estimation chap-
ters, number 4-7, can be read independently of each other. Conventions and
abbreviations are presented in a special notation chapter.

Chapter 1 is the introduction chapter. It briefly presents the thesis and

the companies for which the GARCH models are estimated.

Chapter 2 is the methodology chapter. Here the model set and the estima-
tion procedure is presented. A brief overview of the different toolboxes
available is also given.

Chapter 3 presents the theory for the rest of the thesis. Statistics, stochas-
tic processes and GARCH model theory are the main focuses.

Chapter 4-7 are the estimation chapters for the different data series. The
structure for these chapters is basically the same: preestimation, es-
timation and validation. This chapter can be reed independently of
each other.

Chapter 8 is the conclusion chapter for all the data series.

The reader is assumed to have some knowledge on stochastic processes,

statistic theory and system identification. General knowledge about stock
market and economic theory for these markets is also assumed.
With this introduction we are ready to look into the system identification
method, concerning risk prediction, for stock markets.
Chapter 2

Method

This chapter presents the method used in this thesis. The chapter can roughly
be divided into three different parts. First econometrics and system identi-
fication is presented. Second the system identification procedure is given.
Finally the methodology of estimation in GARCH models is introduced.

2.1 Econometrics
Econometrics as a discipline has existed for about 100 years. It can be sepa-
rated into three large subgroups: First microeconometrics dealing with, for
example, consumer behaviour: second financial econometrics dealing with fi-
nancial data and finally macroeconometrics which focuses on macroeconomic
phenomena such as inflation, unemployment and economic growth. (Deistler
1999) This thesis is concerned with financial econometrics. We are focusing
on stock markets and we are interested in prediction of the first and the
second moments; In particular our main focus is on the second moments.

2.2 System Identification

Each chapter which contains estimation of GARCH models is presented
in the same way. First the company is presented with plots and tables
containing interesting and important company-dependent facts. Some tests
are performed to see whether or not there exists correlation in the squared
return and to see if there exists heteroskedasticity.
The estimation is only done in the univariate case. The procedure is
to first choose the number of lags (p,q) for the symmetric GARCH with a
normal distribution and then to test, for this specific (p,q), different mod-
els. Finally, when the model is chosen, three different distributions for this
model are tested. In the validation part we test to see if the correlation
and heteroskedasticity has been removed from the data series and perform
other tests to validate the model. The system identification procedure that
6 Method

is given in this section is not the only way to do it. A general approach to
system identification is presented by Ljung (1999).

2.2.1 The data set

Outliers are not handled in the raw estimation since it may be hard to know
what an outlier is in a financial market. After the estimation for the whole
sample, the models are estimated for different sub-samples. By doing this
we can get a sense of how outliers affect the model coefficients and how the
parameters in the model are changing over time.
One obvious feature when estimating and predicting in the financial
market is that we have to use the given data and cannot redo any experiment.
Another limitation is the number of data samples that are used. In all our
series we have about 1000 data samples, the data is then split into estimation
part (70%) and validation part (30%). The reason for the unequal split
is that the GARCH estimation procedure requires large sample sizes to
produce good estimates.

2.2.2 Volatility
Modelling volatility can be done in many different ways. Most volatility
models are from these model classes:

GARCH-methods are a way of investigating how a function of past returns,

in a specific financial series, should be constructed and mapped onto
the second moment (Hull 2000). For a specific mapping, the data
series can also be forecast with this method. This is the approach that
will be used in this thesis and will be further explained in Section 3.5.

Stochastic volatility methods are models where the volatility follows a

stochastic differential equation. One way of doing this is presented
by Hull (2000):

dS √
= rdt + V dzs
S (2.1)
dV = a(b − V )dt + ξV α dzv

where a, b, ξ and α are constants, dzs and dzv are Wiener processes
and S and r are the stock price and the risk free rate respectively. V
is the asset’s variance rate, the square of its volatility. This is a time
continuous stochastic differential and can be seen as one possible way
of handling the problem that volatility is not constant over time. This
approach is not used in this thesis and will not be presented further.

Regime switching models are based on the assumption that economic and
financial markets seem to behave differently in different periods of
2.2.3 Selected models for estimation 7

time. This is used to specify different models for different periods.

Markov models can be used for switching between these different mod-
els (Hamilton 1994). This approach will not be used and is not further
explained in this thesis.

Methods from these classes can also be combined. For example Klaasen
(2002) is using regime switching models in combination with GARCH
models.

2.2.3 Selected models for estimation

First moment modelling is not the main focus in this thesis. However a
reasonable model for the first moment has to be used. A misspecification in
this equation could lead to wrong conclusions about which GARCH model
to support. This is because the squared error term from the mean equation
is used as the ”true” volatility when different models are measured on their
forecastability.
As described by Ljung (1999), system identification is, in many cases, a
iterative procedure. It is therefore impossible to come up with the correct
model in one try. Evenso, here a quite general model set is presented, which
is used on all the four time series.
In an efficient market, all information available up until today is already
included in the price. This would lead to the conclusion that there is no
need for any ARX/ARMAX models in the mean equation. Efficient market
theory also states that an investor is expecting the risk free rate as return
and an addition return for taking risk. (Hamilton 1994) Based on this dis-
cussion the model for the return would be a constant plus a risk term and an
unpredictable part εt . However, coupling of equations, between the mean
equation and the second moment, is not possible in the toolbox that we
use for estimation. Therefore the model for the mean that we use, is just
a constant plus the unpredictable part yt = c + εt , here yt is the return on
that specific day and c is the mean value over the values, known so far.
In the model for the second moment GARCH models are used. The
main question here is which GARCH models to use among all those that
exist. Here the choice is based on previous GARCH research (Pagan &
Schwert 1990) and (Peters 2001). We will use four different GARCH models:
GARCH, IGARCH, EGARCH and APARCH, with a small number of lags
0, 1, 2. These are all presented in the theory chapter (see section 3.7).

2.2.4 Criterion of fit

A good performance measure of the second moment can be hard to find
since the volatility is not directly observable. One way of dealing with
this problem is to not rely on one specific measure but rather use several
8 Method

measures. In this thesis six different measures are used for evaluating the
performance of volatility forecasts from different GARCH models:

1. Mean Squared Error (MSE)

2. Median Squared Error (MedSE)

3. Mean Absolute Error (MAE)

4. Adjusted Mean Absolute Percentage Error (AMAPE)

5. Theil Inequality Coefficient (TIC)

6. Mincer-Zarnowitz R2 (R2 )

The measures above need some explanation. The MSE is calculated by:

1 X 2
S+h
(σ̂ t − σt2 )2 (2.2)
h+1
t=S

h is the number of steps ahead that we want to predict, in our case h is

always equal to 1, σ̂ 2t is the forecast volatility, σt2 is the ”true” volatility (in
our case ε2t ) and S is the total sample size.
The MedSE is:

1 X
S+h
(σ̂med 2 − σmed 2 )2 (2.3)
h+1
t=S

The MAE is:

1 X 2
S+h
|σ̂ t − σt2 | (2.4)
h+1
t=S

The AMAPE is:

1 X σ̂ 2t − σt2
S+h
(2.5)
h+1 σ̂ 2t + σt2
t=S

The TIC defined by:

1 PS+h
h+1 (σ̂ 2t − σ 2t )2
t=S
q q (2.6)
1 PS+h 2 1 PS+h 2
h+1 σ̂
t=S t + h+1 t=S σt

is a scale invariant measure that lies between 0 and 1 where 0 indicates

perfect fit.
2.2.5 Validation 9

The R2 is calculated by regressing σ̂t2 on the ε2t , which can be formulated

with this equation

ε2t = α + β σ̂t2 + ut (2.7)

These measures are presented by Brooks (1997) and are implemented in the
GARCH toolbox for Ox by Laurent & Peters (2002).
The first issue to address when using these measures is: What is the true
volatility? This can be done in different ways but in this thesis we will use
the squared residuals ε2t from the equation yt = c + εt , as the true volatility
when measuring the fit. The reason for choosing ε2t as the true volatility is
easily seen in this equation

Eε2t = E(zt σt ) = 1Eσt = σ̂t (2.8)

zt is i.i.d. N (0, 1) and σt is the variance for εt conditioning on the informa-

tion available at time t. This is further explained in Section 3.7.
Another ’true’ volatility is to use intra-day measures and then for a
number PKof different time periods during the day calculate the volatility,
2 2
σt = k=1 r(k),t , where r(k),t2 is the return of the kth intra-day interval of
the tth day. K is the number of intervals per day (Peters 2001). This is not
used since intra-day data is not used for the data series.

2.2.5 Validation
In the validation part, tests are performed to judge whether ARCH effects
and autocorrelation have been removed or not. Tests are also performed
to see if the given assumptions about the model are fulfilled i.e. are the
normalised residuals distributed in the way that was assumed in the model:
normal, student-t or skewed student-t.
A good way of testing a model is to see how it is performing on data not
used in the estimation part. Data points (700 − 1000), depending on the
data set, have been saved for this reason. The out-of-sample measures are
computed with one step ahead prediction (not reestimating the coefficients)
and then for the next day, when new information is available, do the predic-
tion again. We are of course only interested in the model performance for
the second moment (the first moment is just a constant).

2.3 Methods of estimation, toolboxes

When estimating GARCH models, some kind of computer-based software
has to be used. Different software has different functionality, drawbacks and
features. Brooks, Burke & Persand (2001) presents nine different GARCH
estimating software packages and compares them. He finds that there could
10 Method

be large differences between estimated GARCH parameter values from dif-

ferent toolboxes. Therefore it is important to be careful in choosing the
toolbox to use. In this thesis a toolbox1 compatible with the Ox program-
ming language is used (Laurent & Peters 2002). Unfortunately this toolbox
is not one of the nine which are compared by Brooks et al. (2001). Still,
in the tutorial by Laurent & Peters (2002) they compare their own toolbox
with the same benchmarks that Brooks uses and come up with good results.

2.3.1 The Ox GARCH package

In this thesis the Ox GARCH package is used for estimating and forecast-
ing. The package has a variety of different features. Four different distri-
butions are included: normal, student-t, skewed student-t and generalized
error distribution. A number of different models are available, both for the
conditional mean and the conditional variance, as well as a number of tests
and forecast possibilities (Laurent & Peters 2002). In table 2.1 some of the
available toolboxes and their features are presented. Most of these model
features are not used in this thesis so the table is just to be seen as a general
comparison. The reason for choosing the G@RCH package is on the one
hand that the light version is a free version and on the other hand that it
has the GARCH model features that we want to use.

1
G@RCH 2.3
2.3.1 The Ox GARCH package 11

Table 2.1. GARCH features comparison. The table is a modified version of the
table in Laurent & Peters (2002).

G@RCH PcGive Eviews S-Plus SAS Stata

Version 2.3 10 4.0 6 8.2 7
Cond. mean
Expl. var. + + + + + +
ARMA + + + + + +
ARFIMA + - - - - -
ARCH-m - + + + + +
Cond. var.
Expl. var. + + + + + +
GARCH + + + + + +
IGARCH + - - - + -
EGARCH + + + + + +
GJR + + + + - +
APARCH + - - + - +
C-GARCH - - + + - -
FIGARCH + - - + - -
FIEGARCH + - - + - -
FIAPARCH + - - - - -
HYGARCH + - - - - -
Distr.
Normal + + + + + +
Student-t + + - + + -
GED + + - + - -
Skewed-t + - - - - -
Double Exp. - - - + - -
Estimation
ML + + + + + +
QML + + + - - +
Chapter 3

Theory

This chapter contains all the theory used in this thesis. First a background
is given by presenting economic theory, then the necessary statistical tools
and theory are presented. Finally the theory for volatility estimation and the
theory for GARCH models are introduced.

3.1 Economic theory

This first theory section presents the economic theory, which is used in this
thesis. The main highlight from this section is the efficient market theory.
A good book that presents financial theory is Hull (2000). This book is
recommended to the reader who wants a more detailed description.

3.1.1 Financial data

In this thesis we are using a continuous approach when converting between
stock prices and return. Let the stock price be denoted S and µ be the rate
of return. If changes in stock price is denoted ∆S and changes in time are
denoted ∆t the model can be described as

∆S = µS∆t (3.1)

If we let ∆t approach zero and then solve the upcoming differential equation,
the stock price is given by:

ST = S0 eµT (3.2)

Equation 3.2 is the one that will be used in this thesis when converting
between return and stock prices. (Hull 2000)
14 Theory

3.1.2 An introduction to risk

Volatility of a stock is a measure of the uncertainty of the returns from
that specific stock. When returns are calculated from equation (3.2), the
volatility of a stock price becomes the standard deviation of the return.
In stock markets, periods of high risk seem to be followed by periods
of high risk and for low risk periods it is the other way around. This phe-
nomenon is often referred to as volatility clustering. (Hull 2000) For example
when a shock in a stock market occurs, the volatility is high for the subse-
quent period.

3.1.3 Efficient market theory

The theory of efficient markets can be divided into three different levels:
weak, semi-weak and strong efficiency. Capital markets are weak form effi-
cient if it is impossible to form a trading strategy that performs better than
average, using the information of historical prices. This means that so called
technical analysis is fruitless. Semi weak efficiency means that it is impossi-
ble to perform better than average, even when including exogenous variables
(X), where X includes all public information. Strong efficiency means that
private information is also included by exogenous variables. This would
mean, that not even with inside information would it be possible to perform
better than average. (LeRoy 1989)

3.1.4 Leverage effect

One phenomenon in equity markets is that volatility tends to increase more
when the return decreases, than it does, when the return increases. One
reason for this is referred to as the leverage effect. As a company’s equity
reduces in value the leverage for the company increases. When the stock
price for a company decreases their equity declines in value and therefore
their leverage increases. The company has become more risky. On the other
hand, when the company’s stock price increases, the leverage is decreases.
The risk of an equity is therefore dependent on the sign of the return of the
stock. (Hull 2000) This will be further discussed in Section 3.7.4.

3.2 Statistics
A brief description of the statistics that will be used is presented. In all
cases below, X is a discrete valued stochastic variable, k is the summation
index and px (k) is the probability that X is taking value k. A more detailed
description is presented in Hamilton (1994).
3.2.1 Correlation 15

The first moment is the population mean and is defined as

X
E(X) = kpx (k) = µ (3.3)
k

The noncentral second moment is then defined as

X
E(X 2 ) = k 2 px (k) = V ar(X) + E 2 (X) (3.4)
k

Noncentral moments are then in the general case defined as

X
E(X r ) = k r px (k) r = 1, 2, 3 . . . (3.5)
k

Skewness is defined as
E((X − µ)3 )
(3.6)
(V ar(X))3/2
A variable with positive skewness is more likely to have is values far above
the mean value than far below. For a normal distribution the skewness is
zero.
Kurtosis is defined as
E((X − µ)4 )
(3.7)
(V ar(X))2

For a normal distribution the kurtosis is 3. A distribution with a kurtosis

greater than 3 has more probability mass in the tails, so called ”fat tails”,
or leptokurtic.

3.2.1 Correlation
The population correlation between two different random variables X and
Y is defined by
Cov(X, Y )
Corr(X, Y ) ≡ p (3.8)
V ar(X)V ar(Y )
(Hamilton 1994)

3.2.2 Autocorrelation
The jth autocorrelation is defined as the jth autocovariance divided by the
variance:
Cov(Xt , Xt−j )
Corr(Xt , Xt−j ) ≡ p (3.9)
V ar(Xt )V ar(Xt−j )
(Hamilton 1994)
16 Theory

3.2.3 Partial autocorrelation

The partial autocorrelation is also a useful tool in the identification of a time
series. It is defined as the last coefficient in a linear projection of Y on the
(m)
m most recent values. Letting this be denoted αm , if the constant for the
process is zero then the equation becomes:

(m) (m) (m)

Yt+1 = α1 Yt + α2 Yt−1 + . . . + αm Yt−m+1 (3.10)

If the process were a true AR(p) process the coefficients with lags greater
than m would be zero. (Hamilton 1994)

3.3 Hypothesis tests

A hypothesis test is a procedure for analysing data to address questions
whether a certain criterion is fulfilled or not. This can be tested in a number
of different ways and this section presents the hypothesis test that will be
used in this thesis.
All tests have a corresponding p-value. This p-value, under the assump-
tion of the null hypothesis, is the probability of observing the given sample
result.
At the significance level of 95%, which in our notation corresponds with
a critical value of 0.05, then we reject the null hypothesis if the p-value is
lower than this critical value. A p-value greater than 0.05 corresponds to
insufficient evidence for rejecting the null hypothesis. (MathWorks 2002)

3.3.1 Jarque-Bera test

The idea behind the Jarque-Bera test is to test whether a specific distribu-
tion is normal or not. The jb-value is calculated as:
 
T −k 2 (K − 3)2
jb = S + (3.11)
6 4

where T is the number of observations, k is the number of estimated pa-

rameters, S is the skewness and K is the kurtosis. In some programs the
excess kurtosis instead of the kurtosis is calculated. The kurtosis is then 3
plus the excess value. The intuitive feeling about this test is that the larger
the jb-value is, the lower the probability is that the given series is drawn
from a normal distribution. The test statistic of the Jarque-Bera test is
χ2 -distributed with 2 degrees of freedom under the null hypothesis, that the
series is normally distributed. (Hamann 2001)
3.3.2 Ljung-Box test 17

Table 3.1. Ljung-Box-Pierce Q-test example. P-values are given in angle brackets.
With this low p-value we have to reject the hypothesis that no serial correlation
exists.

Q(20) = 157.24 [2.55698e-023 ]

3.3.2 Ljung-Box test

The Ljung-Box test is performed to test whether a series has significant
autocorrelation or not. The Lbq-value is calculated by:
X
k
ri2
Qk = T (T + 2) (3.12)
T −i
i=1

where T is the number of samples, k is the number of lags and ri the ith
autocorrelation. If Qk is large then the probability that the process has
uncorrelated data decreases. The null hypothesis for the test is that there
exists no correlation and under that hypothesis, Qk is χ2 with k degrees of
freedom. Table 3.1 presents an example of how this can be done (the table is
taken from the estimation procedure for Volkswagen in Section 4.2.4). The
Lbq-value is in this case calculated for twenty number of lags. The p-value
corresponding to this Qk value is presented in angle brackets. (MathWorks
2002)

3.3.3 ARCH test

It is fairly easy to test whether the residuals from a regression have con-
ditional heteroskedasticity or not. The test is based on OLS regression,
where the OLS residuals ût from the regression are saved. û2t is thereafter
regressed on a constant and its own m-lagged values. This is done for all
samples t = 1...T . This regression has a corresponding R2 -value. T R2 is
then asymptotically χ2 -distributed with m degrees of freedom under the null
hypothesis that ut is i.i.d. N (0, σ 2 ). (Engle 1982)
This ARCH-test can also be performed as a test for GARCH-effects. The
ARCH-test for lag (p + q) is locally equivalent to a test for GARCH effects
with lags (p, q). (MathWorks 2002)
The null hypothesis, H0 , is that no ARCH effects exists. This is tested
for lags up to T . The OX-package presents this as an F-test with the value
from the F-distribution and the p-value in angle brackets.

3.3.4 Parameter output and t-test

For each selected model from the estimation part, the maximum likelihood
estimates of the selected parameters are calculated. An example of this is
18 Theory

Table 3.2. Example table for parameter test. This output is from Volkswagen and
model is an APARCH(1,1). Maximum likelihood estimation is used.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) −0.004560 0.068735 −0.06634 0.9471
Cst(V) 0.654696 0.413332 1.584 0.1136
GARCH(Beta1) 0.710141 0.075228 9.440 0.0000
ARCH(Alpha1) 0.133134 0.047949 2.777 0.0056
APARCH(Gamma1) −0.133827 0.120061 −1.115 0.2653
APARCH(Delta) 1.957050 0.780077 2.509 0.0123

presented in table 3.2. The first column is the name of each parameter,
Cst(M) stands for the constant in the mean equation. Cst(V) is the esti-
mated variance that does not depend on the time t. The other parameters
are specific for the chosen models. The second column contains the coeffi-
cients for each parameter. In the third column, the estimated standard error
for this coefficient value is presented. The t-value is just the coefficient value
divided with the standard error. T-prob is the p-value given from a t-test.
All these t-tests are tested against zero. Depending on whether the value
can take negative values or not the t-test is either single or double sided.

3.3.5 Nelson test

The Nelson test is a test for time dependency of the parameters that are
estimated with a maximum likelihood estimation. The null hypothesis is
that the parameters are constant and the alternative is that the parameter
Θ follows a martingale process. (Hansen 1994)

3.4 Stochastic processes

This section for stochastic processes is treated in more detail by (Hamilton
1994).

3.4.1 White noise

One of the basic blocks when modelling stochastic processes is the white
noise

E(εt ) = 0
 2
σ for t = s (3.13)
E(εt εs ) =
0 for t 6= s
3.4.2 Stationarity 19

3.4.2 Stationarity
A process Yt is said to be covariance-stationary or weakly stationary if nei-
ther the mean ut nor the autocovariance γjt depend on the time t and if the
given moments exists.

E(Yt ) = µ (for all t)

(3.14)
E((Yt − µ)(Yt−j − µ) = γj (for all t and any j)

3.4.3 Moving average process

A moving average process of order one MA(1) is described as

Yt = α0 + α1 εt−1 + εt (3.15)

where α0 and α1 could be any real constants, and εt is a white noise process
described in equation 3.13. This can of course also be considered in the
general MA(q) case. Then the equation becomes

X
q
Yt = α0 + αj εt−j + εt (3.16)
j=1

3.4.4 Autoregressive process

First we consider an AR(1) process

Yt = α0 + β1 Yt−1 + εt (3.17)

where α0 and β1 can be any real constants, and εt is a white noise process
described in equation 3.13. An AR(1) process can also be generalized to an
AR(p) process.

X
p
Yt = α0 + βj Yt−j + εt (3.18)
j=1

Autocorrelation and the partial autocorrelation that were introduced in

the statistical section is a great help in deciding which AR- and ARMA-
models to be considered. Table 3.3 can be used in this part of the identifi-
cation process. The table is presented at (www.itl.nist.gov 2002).

3.5 Computing volatility

Volatility is a measure of the uncertainty of the return for an asset.
All volatility estimation used in this thesis is calculated from return
series only. This means that the trend from the price series already has
been removed. For further details on this procedure see section 3.1.1.
20 Theory

One obvious way to compute the volatility in a specific market on a

specific day is a linearly weighted moving average, where a specific number
of previous observations is used to calculate the volatility.

1 X
p
σn2 = (un−i − ū)2 (3.19)
p−1
i=1

where ū is the mean value for the process of the period where σn2 is calculated
and ui is the daily return for a specific day. σn2 is here the volatility on day
n for the variable.
If something unexpected happens to a specific market, a shock of some
kind, then it would be more reasonable to assign different weights to different
periods in time. This can be done by
X
p
σn2 = αi u2n−i (3.20)
i=1

where ū is assumed to be equal to zero, which can be a good approximation

if we are dealing with daily data. The problem now is of course to find these
αi . One reasonable constraint is that αi > αj when day i is more recent
than j. Other constraints should be that
X
p
αi = 1 (3.21)
i=1

Table 3.3. Time series identification with the ACF and PACF.
Shape of ACF Indicated model
Exponential, decaying to zero Autoregressive model.
Use the partial
autocorrelation plot
to identify the order
of the autoregressive model.
Alternating positive and Autoregressive model.
negative, decaying to zero Use the partial autocorrelation
plot to help identify the order.
One or more spikes, Moving average model.
rest are essentially zero order identified by
where plot becomes zero.
Decay, starting after a few lags Mixed autoregressive
and moving average model.
All zero or close to zero Data is essentially random.
High values at fixed intervals Include seasonal autoregressive term.
No decay to zero Series is not stationary.
3.5.1 EWMA - Exponentially weighted moving average 21

and that αi > 0. This approach can be generalized in a number of different

ways. One of the more popular ones, is to assign some weight to a long run
volatility. This leads to an equation like

X
p
σn2 = γV + αi u2n−i (3.22)
i=1

where V is the long run volatility and γ is the weight assigned to this volatil-
ity. The constraint from equation (3.21) now looks like

X
p
γ+ αi = 1 (3.23)
i=1

This was first suggested by Engle (1982) and is known as the ARCH(p)
model. (Hull 2000)

3.5.1 EWMA - Exponentially weighted moving average

The weights in equation 3.21 can also be calculated in a different way. Let
αi = λαi−1 , where λ is a constant (0 < λ < 1). This restrictions for λ
is consistent with equation 3.21. This weighting scheme has some good
characteristics. The volatility formula can now be written as

σn2 = λσn−1
2
+ (1 − λ)u2n−1 (3.24)

Here only two values have to be stored for each volatility estimate. It also
adapts to market changes because of the functional form of the equation.
Large changes in un−1 will directly effect the volatility estimate. (Hull 2000)

3.6 ARCH - Autoregressive conditional heteroske-

dasticity
The ARCH-model was first presented by Engle (1982) and has since then
received a lot of attention. First consider an ordinary AR(p) model of the
stochastic process yt .

yt = c + α1 yt−1 + . . . + αp yt−p + ut (3.25)

where ut is white noise. The basic AR(p)-model is now extended so that

the conditional variance of ut could change over time. One extension could
be that u2t itself follows an AR(m)-process.

u2t = θ0 + θ1 u2t−1 + . . . + θm u2t−m + wt (3.26)

22 Theory

where wt is a new white noise process and ut is the error in forecasting yt .

This is the general ARCH(m)- process. (Engle 1982) For easier calculations
and for estimation, a stronger assumption about the process is added.
1/2
u t = ht υ t (3.27)

where υt is an i.i.d. Gaussian process with zero mean and a variance equal
iid
to one υt ∼ N (0, 1) and the whole model for the variance is now

εt |ψt−1 ∼ N (0, ht )
X
q
(3.28)
ht = α0 + αi ε2t−i
i=1

where
α0 > 0 αi > 0,
i = 1, . . . , q

ψt−1 is the information available at time t − 1. Now, when the process for
the variance is defined, we add an additional equation for modelling yt . The
return price is modelled with a constant.

yt = c + εt (3.29)

this means that εt is innovations from a linear regression.

3.7 Generalized ARCH - GARCH

This section is describing a generalization of the ordinary ARCH-model. The
model structure was introduced by Bollerslev (1986). The generalization is
a similar to the extension of an AR(p) to an ARMA(p,q). Formally the
process can be written as

εt |ψt−1 ∼ N (0, ht )
X
q X
p
(3.30)
ht = α0 + αi ε2t−i + βi ht−i
i=1 i=1

where
p integer, q integer
α0 > 0, αi ≥ 0, i = 1, . . . , q
β ≥ 0, i = 1, . . . , p

thus the additional feature is that the process now also includes lagged ht−i
values. For p = 0 the process is an ARCH(q). For p = q = 0 (an extension
allowing q = 0 if p = 0), εt is white noise. (Bollerslev 1986)
3.7.1 GARCH(1,1) 23

Theorem 1. The GARCH(p,q) process as defined in equation 3.30 is wide-

sense stationary with E(εt ) = 0, var(εt ) = α0 (1 − A(1) − B(1))−1 and
cov(εt , εs ) = 0 for t 6= s if and only if A(1) + B(1) < 1.

Proof. see Bollerslev (1986) page 323.

P P
In theorem 1 A(1)= qi=1 αi and B(1)= pi=1 βi . In most cases the num-
ber of parameters is rather small, for instance GARCH(1,1). Next we will
focus in on the model where p = q = 1, a GARCH(1,1) process.

3.7.1 GARCH(1,1)
In the case where p = q = 1 the model becomes

εt |ψt−1 ∼ N (0, ht )
(3.31)
ht = α0 + α1 ε2t−1 + β1 ht−1

where

α0 > 0, α1 ≥ 0, β1 ≥ 0
α1 + β1 < 1

Interesting properties for εt (unconditioned) would be its moments and

they are given by:

Theorem 2. For the GARCH(1,1) process as defined in equation 3.31 a

necessary and sufficient condition for existence of the 2mth moment is
m  
X m
µ(α1 , β1 , m) = aj α1j β1m−j < 1 (3.32)
j
j=0

where
Y
j
a0 = 1, aj = (2i − 1),
i=1

The 2mth moment can be described by the recursive formula

E(ε2m
t ) = am
  !
X
m−1
m
· a−1 2n m−n
n E(εt )α0 µ(α1 , β1 , n) (3.33)
m−n
n=0
· (1 − µ(α1 , β1 , m))−1
24 Theory

Proof. Bollerslev (1986) page 325.

Due to the symmetry properties of the normal distribution, it follows that

if the 2mth moment exists, the 2(m − 1)th moment is zero. It can also be
shown that the distribution of εt is heavily tailed. (Bollerslev 1986)
The existence of the 2mth moment for GARCH models with higher order
than (1, 1) can also be interesting to investigate. The general case is not
known, but for a specific order sufficient and necessary conditions can be
derived. This is done by Bollerslev (1986) page 313 and for the GARCH(1,2)
this is found to be

α2 + 3α12 + 3α22 + β12 + 2α1 β1 − 3α23 + 3α12 α2 + 6α1 α2 β1 + α2 β12 < 1

(3.34)

3.7.2 IGARCH
One special case of GARCH-models is when the sum of estimated parameters
(except for α0 ) equals one:

X
q X
p
αi + βj = 1 (3.35)
i=1 j=1

This is known as integrated GARCH or IGARCH. This is similar to the

EWMA-procedure if α0 is set to zero. (Hamilton 1994)

3.7.3 EGARCH
Nelson (1991) presents a model which is known as Exponential-GARCH or
EGARCH. The idea is to loosen the positivity constraints from the standard
GARCH but still keep the nonnegativity constraint on the volatility for the
conditional variance. A suitable way of doing this is by
∞
X
ln(σt2 ) = αt + βi g(zt−i ) β1 ≡ 1 (3.36)
i=1

where the function g(zt ) can be formulated in different ways. Nelson (1991)
is suggesting

g(zt ) = θzt + γ(|zt | − E|zt |) (3.37)

to handle both the sign and the magnitude of zt .

A slightly different model of the EGARCH is implemented by Laurent
& Peters (2002)

ln(σt2 ) = ω + (1 − β(L))−1 (1 + α(L))g(zt−i ) (3.38)

3.7.4 APARCH 25

3.7.4 APARCH
One of the more general GARCH models is the APARCH model (Asymmet-
ric Power ARCH) which is also implemented by Laurent & Peters (2002) in
the GARCH toolbox. The model structure is
X
q X
p
σtδ =ω+ αi (|εt−i | − γεt−i ) +
δ
βj σtδ (3.39)
i=1 j=1

Which includes many other GARCH models as special cases.

3.8 Distributions
In this thesis three different probability distributions are used. The standard
normal distribution, the student-t and the skewed student-t. For the esti-
mation of the parameters the log-likelihood functions of these distributions
are used. The reason to use different distributions (other than the normal)
is that the third and fourth moment of the normal distribution could be
too restrictive. The student-t has one parameter for modelling the fourth
moment and this parameter is also estimated in the maximum likelihood
estimation. For the skewed student-t there is also an additional parameter
for having skewness not equal to zero in the distribution. This is the most
general of the three distributions.
Chapter 4

Volkswagen

In this chapter, model estimation and model selection for the Volkswagen
stock is performed. The procedure is following the structure that is pre-
sented in the methodology chapter. Volkswagen as a company is presented,
preestimation and estimation is performed and finally some conclusions are
drawn.

4.1 Presenting data

4.1.1 Introducing Volkswagen
The stock price for Volkswagen is taken from the Germany stock market
(DE). The data is sampled from the 25th of February 1999 to the 21st of
October 2002. Particularly interesting days are presented in table 4.1. The
certain criterion is that the squared return should be above 60 in this scaled
measure presented in Section 4.2.1. One thing notable, is that the 11th of
September does not meet the criterion but that three out of eight days with
extremely high return occurred in September 2001. The dates are presented
in the form (dd.mm.yy).

4.2 Preestimation
4.2.1 Data formatting
For the data series of Volkswagen 952 data points is the whole data set. The
data is daily data and the return is calculated by:
returnt = 100(log(pricet ) − log(pricet−1 )) (4.1)
The reason for multiplying by 100 is due to numerical problems in the esti-
mation part. This will not affect the structure of the model since it is just
a linear scaling. Data with numbers from 1-800 are used in the estimation
part and data with numbers 801-952 are used as cross validation.
28 Volkswagen

Table 4.1. Days with high return in the Volkswagen series.

number return date

210 −8.4920 15.12.99
540 9.6572 21.03.01
666 8.2731 13.09.01
670 7.8758 19.09.01
672 −7.9494 21.09.01
892 −9.1398 26.07.02
935 −7.9203 25.09.02
946 −8.8166 10.10.02

Table 4.2. Moments, skewness and kurtosis for the return series of Volkswagen.

Volkswagen
First moment 0
Second moment 5.1369
Third moment −0.4209
Fourth moment 126.4742
Skewness −0.0362
Kurtosis 4.7930

4.2.2 Moments for the return series

In the preestimation stage it would be interesting to get a picture of what
the distribution of the raw return series is like. The moments, skewness and
kurtosis are presented in table 4.2. For a theoretical introduction to these
measures see section 3.2. It is notable that the kurtosis is considerably larger
than three, which would have been the kurtosis if the return series were
drawn from a normal distribution. This behaviour is known to frequently
occur in financial markets.

4.2.3 Plots for Volkswagen

Before starting the estimation procedure it is interesting to see the behaviour
of the specific data set presented in graphical terms. A price plot is pre-
sented in figure 4.1. The squared returns of the data series are presented in
figure 4.2. In the diagram we see clustering of high volatility. A graphical
presentation of the autocorrelation is presented in figure 4.3. Here we can
see that there is a high persistence in the second moment for at least 10
lags. The partial autocorrelation is presented in figure 4.4 and it also shows
4.2.4 Hypothesis tests 29

persistence in the second moment.

75

70

65

60

55

50

45

40

35

30
0 100 200 300 400 500 600 700 800 900 1000

Figure 4.1. Price plot for Volkswagen.

100

90

80

70

60

50

40

30

20

10

0
0 100 200 300 400 500 600 700 800 900 1000

Figure 4.2. Squared return for Volkswagen.

4.2.4 Hypothesis tests

The indication from the graphical plot that there exists correlation between
data plots is in this subsection tested with the Ljung-Box-test. The test for
heteroskedasticity is performed with an ARCH test. See section 3.3 for a
theoretical explanation for these tests.
30 Volkswagen

Sample Autocorrelation Function (ACF)

0.8

0.6
Sample Autocorrelation

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 4.3. Autocorrelation of the squared return for Volkswagen.

Sample Partial Autocorrelation Function

0.8
Sample Partial Autocorrelations

0.6

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 4.4. Partial autocorrelation of the squared return for Volkswagen. They
represent a measure of partial autocorrelation in the volatility.

Ljung-Box-Pierce Q-test

The computed statistical values for the certain lags are presented together
with the corresponding p-value in angle brackets. Result from the test is
presented in table 4.3 for Volkswagen. The null hypothesis is that no serial
correlation exists and this hypothesis is accepted when the p-value is high.
In this case we reject that hypothesis.
Estimation 31

Table 4.3. Ljung-Box-Pierce Q-test for Volkswagen. P-values are given in angle
brackets. With these low p-values we reject the hypothesis that no serial correlation
exists.

Q(10) = 149.508 [4.70469e-027 ]

Q(15) = 155.131 [2.30051e-025 ]
Q(20) = 157.24 [2.55698e-023 ]

Table 4.4. Engle’s ARCH-test for Volkswagen on the raw return series. H0 is that
no ARCH effect exists and in this case we have p-values equal to zero which means
that we can reject the hypothesis.

ARCH 1-5 test: F(5,789) = 14.027 [0.0000]**

ARCH 1-10 test: F(10,779) = 10.222 [0.0000]**
ARCH 1-15 test: F(15,769) = 7.2884 [0.0000]**

Engle’s ARCH-test

Engle’s ARCH-test performed before estimation. This test supports that

there exists heteroskedasticity. P-values equal to zero for all of the three
series. The result for Volkswagen is presented in table 4.4. ARCH-effects
are tested on lags 1-5, 1-10 and 1-15. With these p-values we reject the
hypothesis that no ARCH-effects exist.

4.3 Estimation
This section is presenting the whole estimation procedure for the Volkswa-
gen GARCH models. First p and q for the standard GARCH are chosen
and then for this specific p and q different GARCH models are tested, esti-
mated and finally one GARCH model with specific p and q is chosen. For
this GARCH(p,q)-model different distributions are evaluated. Finally the
parameters for the chosen GARCH models are presented.

4.3.1 Choosing p and q

Different p and q values for the standard symmetric GARCH model are
tested. The selection procedure is based on each model’s ability to produce
forecasts. Six different forecasts measures are used. See section 2.2.4 for
a description of these measures. The Volkswagen table 4.5 shows that the
GARCH(1,1) performs best in three out of six out-of-sample measures. In
two of the measures of which the GARCH(1,1) does not score the best the
32 Volkswagen

Table 4.5. Selection for symmetric GARCH models for Volkswagen. Log-likelihood
is presented as well as Akaike’s information criterion for different p and q. Six
different forecasts measures are also presented to determine which p and q that
performs best on out-of-sample data for the symmetric GARCH.

(p,q) (0,0) (1,1) (1,2) (2,1) (2,2)

Log −1726.53 −1693.55 −1693.36 −1693.15 −1692.75
Akaike 4.321326 4.243863 4.245887 4.245383 4.246866
R2 0.0000 0.09341 0.08100 0.08612 0.08247
MSE 237.4908 203.1438 205.1336 204.2901 204.7869
MedSE 15.7309 15.6039 15.4001 14.9537 15.4742
MAE 8.2124 8.0538 8.0759 8.0668 8.0755
AMAPE 0.5570 0.5356 0.5354 0.5355 0.5352
TIC 0.7123 0.5779 0.5787 0.5784 0.5777

Table 4.6. GARCH model selection for Volkswagen. All models with p = q = 1.

GARCH EGARCH APARCH IGARCH

Log −1693.545 −1701.674 −1692.588 −1706.267
Akaike 4.243863 4.269186 4.246470 4.273168
R2 0.0934135 0.116252 0.115478 0.11051
MSE 203.1438 200.1063 199.5250 195.0135
MedSE 15.6039 16.8086 15.5668 22.1922
MAE 8.0538 8.1821 7.9957 8.8270
AMAPE 0.5356 0.5420 0.5339 0.5450
TIC 0.5779 0.5765 0.5732 0.4865

difference is in the fourth decimal. Choice for future model selection is

p = 1, q = 1.

4.3.2 Choosing model

Four different GARCH models are tested. These models are described in
Section 3.7. In table 4.6 the model APARCH(1,1) is performing best on
out-of-sample measures. Therefore the APARCH(1,1) is used in the future
model selection.

4.3.3 Choosing distribution

Three different distributions are tested. A description of these models is
made in Section 3.8. The distributions are: normal, student-t and skewed
4.3.4 Estimated parameters 33

Table 4.7. Distribution selection for Volkswagen. The distribution is a Student

distribution, with 6.0162 degrees of freedom. The distribution is a Skewed Stu-
dent distribution, with a tail coefficient of 5.9079 and an asymmetry coefficient of
−0.0515656.

Normal Student-t Skewed Stud.-t

Log −1692.588 −1677.872 −1677.220
Akaike 4.246470 4.212179 4.213049
R2 0.115478 0.113103 0.113761
MSE 199.5250 198.6129 198.7152
MedSE 15.5668 16.5073 16.5735
MAE 7.9957 8.0435 8.0369
AMAPE 0.5339 0.5347 0.5346
TIC 0.5732 0.5667 0.5677

Table 4.8. Estimated parameters for the chosen model for Volkswagen. The model
is an APARCH(1,1) with normal distribution.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) −0.004560 0.068735 −0.06634 0.9471
Cst(V) 0.654696 0.413332 1.584 0.1136
GARCH(Beta1) 0.710141 0.075228 9.440 0.0000
ARCH(Alpha1) 0.133134 0.047949 2.777 0.0056
APARCH(Gamma1) −0.133827 0.120061 −1.115 0.2653
APARCH(Delta) 1.957050 0.780077 2.509 0.0123

student-t distribution. The result for the distribution comparison is pre-

sented in table 4.7. Here we will use the normal distribution as our choice.

4.3.4 Estimated parameters

The best GARCH model in out-of-sample data is the APARCH(1,1). The

parameters for the chosen model is presented in table 4.8. The APARCH-
(Gamma1) is the coefficient, which takes the leverage effect into account.
This parameter in this case is not significant. For a more detailed discus-
sion about the leverage effect, see section 3.1.4. This table is explained in
detail in Section 3.3.4. One thing that is also interesting to note is that
the APARCH(Gamma1) is shown to be time dependent in the Nelson test,
section 3.3.5.
34 Volkswagen

Table 4.9. Estimated parameters for the chosen model for Volkswagen. The model
is a GARCH(1,1) with normal distribution. Data set divided in two parts (1-470)
and (471-953) to see if the parameters is time dependent.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) 0.050629 0.086188 0.5874 0.5572
Cst(V) 2.060946 0.690373 2.985 0.0030
GARCH(Beta1) 0.277634 0.182906 1.518 0.1297
ARCH(Alpha1) 0.199341 0.066573 2.994 0.0029
Cst(M) −0.014238 0.093011 −0.1531 0.8784
Cst(V) 0.208289 0.094871 2.196 0.0286
GARCH(Beta1) 0.832004 0.037223 22.35 0.0000
ARCH(Alpha1) 0.140909 0.033285 4.233 0.0000

4.4 Validation
In this section the chosen GARCH model is validated. One part of the val-
idation procedure has already been taking part when the model was chosen
on its forecasting ability. Other tests are performed to see if the autocorrela-
tion in the squared return has successfully been removed and to see whether
ε2
or not the distribution of the normalised residuals ht2 = zt are normally
t
distributed. All these tests and measures will be presented in the following
section.

4.4.1 Parameter time dependence

To check for time dependence of the parameters the data set is divided in
two parts (1-470) and (471-953). As seen in table 4.9 the estimated co-
efficients are heavily dependent on the data set. The data estimation is
for the standard GARCH model with normal distribution. The reason is
due to convergence problems in the estimation of the APARCH(1,1) coef-
ficients. Cst(V) is 2.06 in the first period and 0.208 in the other period.
GARCH(Beta1) is 0.28 in the first data set and 0.83 in the other. It is rea-
sonable that the GARCH(Beta1) parameter is changing over time. When
beta is low then the value of Cst(V) is going up and vice versa. It certainly
looks like the GARCH-effects in this stock are changing over time.

4.4.2 Autocorrelation
As seen in figure 4.5 most of the significant autocorrelation has been re-
moved. Compared this result to the result from the preestimation (see fig-
ure 4.3).
4.4.3 Ljung-Box-Pierce Q-test 35

Sample Autocorrelation Function (ACF)

0.8

0.6
Sample Autocorrelation

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 4.5. Autocorrelation for Volkswagen with residuals after estimation.

Table 4.10. Ljung-Box-Pierce Q-test on the squared residuals, validation for Volk-
swagen. H0 is that no significant correlation exists. H0 should be accepted when
the probability is high. The test is performed on lags 10, 15 and 20.

Q(10) = 7.94418 [0.438942 ]

Q(15) = 12.2437 [0.507767 ]
Q(20) = 14.7842 [0.676728 ]

4.4.3 Ljung-Box-Pierce Q-test

The Lbq-test in the validation part is to test whether there still is a signif-
icant correlation in the standardised innovations or not. The standardised
ε2
innovations are ht2 = zt and according to the theory for GARCH models
t
and our assumptions in Section 3.7, zt is i.i.d. N (0, 1). For results on the
Lbq-test see table 4.10. The test is made on zt with lags 10, 15 and 20.
Each row in the table represents one test with a specific lag. H0 is accepted,
p-values are given within angle brackets. The theory for the test is presented
in Section 3.3.2.

4.4.4 Engle’s ARCH-test

The result from Engle’s ARCH-test is presented in table 4.11. This indicates
that we have successfully removed the conditional heteroskedasticity that ex-
isted when the test was performed on the pure return series in Section 4.2.4.
36 Volkswagen

Table 4.11. Engle’s ARCH-test, validation for Volkswagen. H0 is that no ARCH

effect exists.

ARCH 1-5 test: F(5,787) = 0.57168 [0.7218]

ARCH 1-10 test: F(10,777) = 0.97848 [0.4606]
ARCH 1-15 test: F(15,767) = 0.99426 [0.4590]

Table 4.12. Moments, skewness and kurtosis for zt of Volkswagen, where zt is the
standardised residuals.

Volkswagen
First moment 0.00089
Second moment 1.00539
Skewness 0.31870
Excess kurtosis 1.1851

4.4.5 Is zt normally distributed?

According to our assumptions in Section 3.7, zt is supposed to be normally

distributed. In this section we will analyse if this really is the case. There
is a number of ways to do this. The discussion here is based on the most
common one, which is the Jarque-Bera test, and the calculated moments for
the distribution. For a more detailed description see section 3.3.

Moments of zt

For the calculated first two moments, skewness and excess kurtosis see ta-
ble 4.12. It is quite clear that this distribution at least has first and second
moments close to their corresponding normal values. Noticeable here is that
the skewness has increased with a factor 10 compared with the raw return
series. The reason for this will be treated in Section 4.4.6.

Jarque-Bera test

To test whether the zt is normally distributed or not a Jarque-Bera test is

performed. The result from the jb-test is presented in table 4.13. Theory for
the Jarque-Bera test is presented in Section 3.3.1. A p-value close to zero
indicates that we must reject the hypothesis that zt is normally distributed.
4.4.6 Skewness of zt 37

Table 4.13. Results from the jb-test. Normality test for zt .

Statistic p-value
Jarque-Bera 46.917 6.4877e-011

Table 4.14. Skewness for zt for Volkswagen. With standardised residuals for
the GARCH(1,1). APARCH(1,1) not presented because of numerical convergence
problems.

Data interval Skewness value t-test p-value

(1-450) −0.13855 1.2039 0.22864
(451-952) 0.20687 1.8979 0.057715

4.4.6 Skewness of zt
For a detailed analysis of the reason for the high skewness in zt four dif-
ferent estimates are made. The data is divided into two different parts,
1 − 450 and 451 − 952, and then two different models are estimated. One
is the APARCH(1,1), which is our chosen model, and one is the symmet-
ric GARCH(1,1). The skewness result from this estimation is presented in
table 4.14. It is quite obvious that the skewness for the raw return is de-
pendent on which time part the estimation is done over. The result for the
APARCH(1,1) is not presented because of numerical convergence problems
in the estimation procedure.

4.5 Summary
4.5.1 Estimation
Different (p, q) values are tested in the symmetric case. These are then used
as fixed p and q for different GARCH-models. In this case p = q = 1
appears to be the best out-of-sample choice.

Different models are tested for these specific p and q. The asymmetric-
power-ARCH is the model that performs best among the models that
are tested.

Different distribution for this specific APARCH(1,1) are then tested and
estimated. There is no evidence that any other distribution than the
normal should be used.

The estimated coefficient values are fairly close to the standard symmetric
GARCH(1,1), with APARCH(Gamma1) close to zero and APARCH-
38 Volkswagen

(Delta) close to two.1 Still, the APARCH(1,1) outperforms the GA-

RCH(1,1) in out-of-sample measures.

The standardised residuals zt are not normally distributed and this is be-
cause of the skewness and the kurtosis. However the standardised
mean and the standardised variance are close to there excepted val-
ues.

Convergence problems arise when the data set is too small (< 500) data
samples. Ideally more than 700 data points should be used for the
estimation of GARCH parameters.

Time dependent coefficients is a problem within this data set. Even for
a standard GARCH(1,1) the parameters are different when estimated
over two data samples in the same series.

The final decision for which model is up to the user. The model which
perform best on out-of-sample measures is the APARCH(1,1) but this
model has convergence problems and the coefficients seem to be time
dependent. If this model is used, then the coefficient has to be reesti-
mated when new data (information) is available. The other possibility
is to use the more robust symmetric GARCH(1,1) which is not the best
on out-of-sample measures but has less time dependent parameters.

1
This is easily seen when these parameters are put into the APARCH-equation.
Chapter 5

Münchner Rück

In this chapter model estimation for the Münchner Rück stock is performed.
The procedure is following the structure presented in the methodology chap-
ter. The structure is the same as for Volkswagen.

5.1 Presenting data

5.1.1 Introducing Münchner Rück
The stock price for Münchner Rück is taken from the Germany stock market
(DE). The data is sampled daily from the 25th of February 1999 to the 21st
of October 2002. Especially interesting days are presented in table 5.1.
There are three days with return larger than 10% over the whole period.
The days before and after these large returns are also presented. The date
is in the format (dd.mm.yy). Instead of typing Münchner Rück we will from
here on use the abbreviation muvg.

Table 5.1. Days with high return in the muvg series.

date nr muvg
22.12.99 119 0.0422
23.12.99 120 0.1568
24.12.99 121 0
10.09.01 567 0.0123
11.09.01 568 −0.1704
12.09.01 569 0.0508
26.07.02 796 0.0156
29.07.02 797 0.1067
30.07.02 798 0.0075
40 Münchner Rück

Table 5.2. Moments, skewness and kurtosis for the return series of muvg.

muvg
First moment 0
Second moment 6.6197
Third moment 0.0867
Fourth moment 329.6006
Skewness 0.0051
Kurtosis 7.5215

5.2 Preestimation
5.2.1 Data formatting

The data series for muvg consists of 852 data points. The data is daily and
the return is calculated with equation (4.1). Data points (1-700) are used in
the estimation part and data points (701-832) are used for cross-validation.

5.2.2 Moments for the return series

The moments, skewness and kurtosis are presented in table 5.2. The data
series has a high kurtosis which is usual for stock price return series.

5.2.3 Plots for Münchner Rück

In this section the most important plots are presented. Before starting the
estimation procedure it is interesting to see the behaviour of the specific data
set presented in graphical terms. The price and return plots are presented
in figure 5.1 and figure 5.2 respectively. The squared return and the auto-
correlation is presented in figure 5.3 and figure 5.4. Here we observe that
there is some persistence in the second moment for at least five lags. The
partial autocorrelation is plotted in figure 5.5 and it also shows persistence
in the second moment.

5.2.4 Hypothesis tests

The feeling from the graphical plot that there exists correlation between
data plots is in this subsection tested with the Ljung-Box-test. The test
for heteroskedasticity is performed with an ARCH-test (see section 3.3 for
a theoretical explanation for these tests).
5.2.4 Hypothesis tests 41

Münchner Rück

400
350
300
stock price

250
200
150

0 200 400 600 800

date

Figure 5.1. Price plot for Münchner Rück displaying daily closing prices. The
time axis ranges from the 9th July 1999 to the 17th of September 2002.

Münchner Rück
−0.05 0.00 0.05 0.10 0.15
return

−0.15

0 200 400 600 800

Figure 5.2. Return plot for Münchner Rück, continuously compounded. The data
is daily and starts at the 9th July 1999 and ends the 17th of September 2002.

Ljung-Box-Pierce Q-test

We compute the statistical value for specific lags and present them together
with the p-values in angle brackets; The result from the test is presented in
table 5.3. The null hypothesis is that no serial correlation exists and this
hypothesis is accepted when the p-value is high. In this case we have to
reject that hypothesis.
42 Münchner Rück

300

250

200

150

100

50

0
0 100 200 300 400 500 600 700 800 900

Figure 5.3. Squared returns for Münchner Rück, for daily data starts at the 9th
July 1999 and ends the 17th of September 2002.

muvg

0.8

0.6
Sample Autocorrelation

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 5.4. Autocorrelation for muvg on the squared return.

Table 5.3. Ljung-Box-Pierce Q-test for muvg. P-values are given in angle brackets.
With these low p-values we have to reject the hypothesis that no serial correlation
exists.

Q(10) = 54.0457 [6.76393e-009 ]

Q(15) = 56.1805 [2.50501e-007 ]
Q(20) = 65.5801 [2.49021e-007 ]
Estimation 43

muvg

0.8
Sample Partial Autocorrelations

0.6

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 5.5. The partial autocorrelation on the squared return for muvg, a measure
of partial autocorrelation, in the volatility.

Table 5.4. Engle’s ARCH-test for muvg on the raw return series. H0 is that no
ARCH effect exists and in this case we have p-values equal to zero which means
that we can reject that hypothesis.

ARCH 1-2 test: F(2,693) = 3.2265 [0.0403]*

Engle’s ARCH-test

Engle’s ARCH-test is performed before estimation. This test supports that

there exists heteroskedasticity since the p-value is less than 0.05. The full
result for muvg is presented in table 5.4. ARCH-effects are tested on lags
1-2. With this p-value we reject the hypothesis that no ARCH effects exist,
but only on the level of 95%.

5.3 Estimation

This section is presenting the whole estimation procedure for muvg GARCH
models. First p and q for the standard GARCH model is chosen and then
for these specific p and q different GARCH models are tested, estimated
and finally one GARCH model with a specific p and q is chosen. For this
GARCH(p,q)-model different distributions are evaluated. Finally the pa-
rameters for the chosen GARCH model are presented.
44 Münchner Rück

Table 5.5. Selection for symmetric GARCH for muvg. Log-likelihood is presented
as well as Akaike’s information criterion for different p and q. Six different mea-
sures are also presented to determine which p and q that perform best on out-of-
sample data.

GARCH(p,q) (0,0) (1,1) (1,2) (2,1) (2,2)

Log −1595.2 −1548.66 −1548.54 −1548.41 −1548
Akaike 4.563421 4.436184 4.438688 4.438313 4.439988
R2 0 0.177904 0.180966 0.18553 0.175453
MSE 405.4252 302.5713 301.3332 299.5203 303.6363
MedSE 28.2512 24.8449 26.2407 27.5363 28.2426
MAE 10.9956 10.6565 10.6410 10.6349 10.7037
AMAPE 0.5834 0.5355 0.5351 0.5351 0.5375
TIC 0.7169 0.4731 0.4717 0.4702 0.4752

5.3.1 Choosing p and q

Different p and q for the standard symmetric GARCH models are tested.
The selection procedure is based on each model’s ability to produce forecasts.
The six different forecasts measures are discussed in Section 2.2.4 are used.
The muvg table 5.5 shows that the GARCH(2,1) performs best in three out
of six out-of-sample measures. Choice for future model selection is p = 2
and q = 1.

5.3.2 Choosing model

Four different GARCH models are tested. These models are theoretically
described in Section 3.7. In table 5.6 the model APARCH(2,1) is performing
best on out-of-sample measures. Therefore the APARCH(2,1) is used in the
future model selection. The EGARCH model did not converge so the Log
and Akaike values are presented for the initial estimate instead.
The APARCH(1,1) has been a good choice in the other data series,
therefore the APARCH(2,1) will be tested against the APARCH(1,1) to
verify that it really performs best. The result is presented in table 5.7. The
result confirms that APARCH(2,1) is our choice for further estimation.

5.3.3 Choosing distribution

Three different distributions are tested on APARCH(2,1): normal, student-t
and skewed student-t distribution. The result for the distribution compari-
son is presented in table 5.8. It is not an obvious choice which distribution
to choose. The student-t and the skewed student-t are performing quite
equally in the out-of-sample measures. But looking at the Log value, we
5.3.4 Estimated parameters 45

Table 5.6. GARCH model selection for muvg. The EGARCH model did not con-
verge during the estimation, due to numerical problems. All models are estimated
with p = 2 and q = 1.

GARCH EGARCH APARCH IGARCH

Log −1548.41 −1556.41 −1546.59 −1549.35
Akaike 4.438313 4.466900 4.438823 4.438140
R2 0.18553 - 0.210067 0.185916
MSE 299.5203 - 289.8320 304.4905
MedSE 27.5363 - 30.7050 32.5273
MAE 10.6349 - 10.6297 10.9585
AMAPE 0.5351 - 0.5315 0.5344
TIC 0.4702 - 0.4541 0.4480

Table 5.7. Verification of the APARCH(2,1) model for muvg.

APARCH(1,1) APARCH (2,1)

Log −1546.94 −1546.59
Akaike 4.436977 4.438823
R2 0.204106 0.210067
MSE 292.2124 289.8320
MedSE 32.0056 30.7050
MAE 10.6204 10.6297
AMAPE 0.5317 0.5315
TIC 0.4567 0.4541

notice that it has not changed when adding an extra parameter. This is an
indication that we should choose the student-t as our distribution choice.
This is also the AIC choice and it will be our final model choice.

5.3.4 Estimated parameters

The best GARCH model in out-of-sample data is the APARCH(2,1). The
parameters for the chosen model is presented in table 5.9. The APARCH-
(Gamma1) takes the leverage effect into account, but the parameter in this
case is not significant. For a more detailed discussion about the leverage
effect see section 3.1.4. This table is explained in detail in Section 3.3.4.
Three coefficients are time dependent on the level of 95%, these are: Cst(V),
GARCH(Beta1) and GARCH(Beta2) (see the Nelson test in Section 3.3.5).
46 Münchner Rück

Table 5.8. Distribution selection for muvg. The student distribution has 7.0056
degrees of freedom and the skewed student distribution has a tail coefficient of
7.00123 and an asymmetry coefficient of −0.00283962. All three are tested on
an APARCH(2,1) model.

Normal Student-t Skewed Stud.-t

Log −1546.59 −1520.52 −1520.52
Akaike 4.438823 4.367206 4.370059
R2 0.210067 0.219198 0.219242
MSE 289.8320 286.8509 286.8081
MedSE 30.7050 28.8216 28.8743
MAE 10.6297 10.5239 10.5264
AMAPE 0.5315 0.5297 0.5297
TIC 0.4541 0.4641 0.4638

Table 5.9. Estimated parameters for the chosen APARCH(2,1) model for muvg.
The model is estimated with a maximum likelihood estimation.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) −0.019815 0.072585 −0.2730 0.7849
Cst(V) 0.068825 0.047832 1.439 0.1506
GARCH(Beta1) 1.423870 0.300096 4.745 0.0000
GARCH(Beta2) −0.483291 0.277752 −1.740 0.0823
ARCH(Alpha1) 0.046485 0.024051 1.933 0.0537
APARCH(Gamma1) 0.298389 0.140086 2.130 0.0335
APARCH(Delta) 1.899521 0.447514 4.245 0.0000
Student(DF) 7.005608 1.575725 4.446 0.0000

5.4 Validation
In this section the chosen GARCH model is validated. One part of the val-
idation procedure has already been taking part when the model was chosen
on its forecasting ability. Other tests are performed to see if the autocorre-
lation in the squared return has successfully been removed. Both these tests
and measures will be presented in the following sections.

5.4.1 Time dependent parameters

The data set is divided into two parts: One with data points (1-399) and
the other with data points (400-832). As one can see in table 5.10 the values
of the parameters depend on the data set. Either large samples should be
used to get accurate parameter estimation or the model has highly time
5.4.2 Ljung-Box-Pierce Q-test 47

Table 5.10. Estimated parameters for muvg on different data sets for the
APARCH(2,1) model with a skewed student-t distribution. The data set is divided
in two parts, (1-399) and (400-832) respectively.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) 0.105616 0.113493 0.9306 0.3526
Cst(V) 0.066571 0.065363 1.018 0.3091
GARCH(Beta1) 0.074981 0.079091 0.9480 0.3437
GARCH(Beta2) 0.865831 0.079229 10.93 0.0000
ARCH(Alpha1) 0.038896 0.015634 2.488 0.0133
APARCH(Gamma1) −0.964388 0.544377 −1.772 0.0772
APARCH(Delta) 1.230629 0.406636 3.026 0.0026
Student(DF) 8.756206 3.208301 2.729 0.0066
Cst(M) −0.203426 0.078598 −2.588 0.0100
Cst(V) 0.063072 0.030381 2.076 0.0385
GARCH(Beta1) 1.144989 0.205460 5.573 0.0000
GARCH(Beta2) −0.237451 0.185655 −1.279 0.2016
ARCH(Alpha1) 0.078962 0.023409 3.373 0.0008
APARCH(Gamma1) 0.996783 0.133186 7.484 0.0000
APARCH(Delta) 1.207820 0.258940 4.664 0.0000
Student(DF) 24.694184 22.765973 1.085 0.2787

dependent coefficients. These parameter values can be compared also with

the values in table 5.9. Notice also how the APARCH(Gamma1) and the
GARCH(Beta2) even have changed sign between the different periods.

5.4.2 Ljung-Box-Pierce Q-test

The Lbq-test in the validation part tests if there still is significant correlation
in2 the standardised innovations or not. The standardised innovations are
εt
h2t
= zt . For results on the Lbq-test see table 5.11. The test is made on
zt with lags 10, 15 and 20. Each row in the table represents a test with a
specific lag. H0 has to be accepted with these p-values given within angle
brackets. The theory for the test is presented in Section 3.3.2.

5.4.3 Engle’s ARCH-test

Result from Engle’s ARCH-test is presented in table 5.12. This result indi-
cates that we have successfully removed the conditional heteroskedasticity
that existed when the test was performed on the pure return series in Sec-
tion 5.2.4.
48 Münchner Rück

Table 5.11. The Ljung-Box-Pierce Q-test on the squared residuals in the validation
for muvg. H0 is that no significant correlation exists. H0 should be accepted when
probability is high. The test is performed on lags 10, 15, 20.

Q(10) = 5.47597 [0.602084 ]

Q(15) = 7.37838 [0.83163 ]
Q(20) = 10.4553 [0.883439 ]

Table 5.12. Engle’s ARCH-test in validation for muvg. H0 is that no ARCH

effect exists.

ARCH 1-5 test: F(5,686) = 0.31077 [0.9066]

ARCH 1-10 test: F(10,676)= 0.54581 [0.8578]
ARCH 1-15 test: F(15,666)= 0.46504 [0.9575]

5.4.4 How is zt distributed?

Moments of zt

For the calculated first two moments, skewness and excess kurtosis, see
table 5.13. Here at least the first moment and the second moment are close
to their expected values.

5.5 Summary
5.5.1 Estimation
Different (p,q) values are tested in the symmetric case. These are then used
as fixed p and q for different GARCH models. In this case p = 2, q = 1
is the best out-of-sample choice.

Table 5.13. Moments, skewness and kurtosis for zt of muvg, where zt is the
standardised residuals.

muvg
First moment 0.01913
Second moment 1.04194
Skewness 0.45008
Excess kurtosis 5.9992
5.5.1 Estimation 49

Different models are tested with these specific p and q. The asymmetric-
power-ARCH (APARCH(2,1) performs best among the models that
are tested.

Different distributions for this specific APARCH(2,1) are then tested and
estimated. The distribution we use is best on out-of-sample measures
together with the AIC choice and our final choice is the student-t
distribution.

Numerical problems for convergence occurred in the estimation part of the

EGARCH(2,1). Therefore this model has not been fully tested.
Chapter 6

Suez-Lyonnaise des Eaux

In this chapter model estimation for the volatility in Suez-Lyonnaise des

Eaux stock is performed. The procedure is following the structure that was
presented in the methodology chapter. The structure is the same as for
Volkswagen.

6.1 Presenting data

6.1.1 Introducing Suez-Lyonnaise des Eaux
The stock price for Suez-Lyonnaise des Eaux is taken from the Paris Stock
Exchange (PA). The data is daily from the 9th of June 1999 to the 17th of
September 2002. Especially interesting days are presented in table 6.1. The
selection criterion is that the absolute return should be above 10%. The
11th of September is one of those days. Days before and after interesting
days are also presented. The data is presented in the form (dd.mm.yy).
During this chapter, instead of typing Suez-Lyonnaise des Eaux, we will use
the abbreviation lyoe. The stock price from date 00.06.06 is missing. This
is handled by taking the mean of the price the day before and the day after.

6.2 Preestimation
6.2.1 Data formatting
The data series of lyoe consists of 832 data points. The return is calculated
by:

returnt = 100(ln(pricet ) − ln(pricet−1 ) (6.1)

The reason for multiplying by 100 is due to numerical problems in the es-
timation part. This will not affect the structure of the model, since this is
52 Suez-Lyonnaise des Eaux

Table 6.1. Days with high absolute returns in the lyoe series. The day before and
the day after interesting dates are also printed. Date format is (dd.mm.yy).

date nr lyoe
01.02.00 148 0.0228
02.02.00 149 0.1026
03.02.00 150 0.0456
10.09.01 567 −0.0092
11.09.01 568 −0.1013
12.09.01 569 0.0795
12.07.02 786 −0.0074
15.07.02 787 −0.1219
16.07.02 788 0.0762
24.07.02 794 −0.0500
25.07.02 795 0.1043
26.07.02 796 0.0141
05.08.02 802 −0.0801
06.08.02 803 0.1241
07.08.02 804 −0.0637

Table 6.2. Moments, skewness and kurtosis for the return series of lyoe.

lyoe
First moment 0
Second moment 5.0779
Third moment 1.4272
Fourth moment 205.0037
Skewness 0.1247
Kurtosis 7.9506

just a linear scaling. Data points (1-700) are used in the estimation part
and data points (701-832) are used as cross validation.

6.2.2 Moments for the return series

In the preestimation part it would be interesting to get a picture of what
the distribution of the raw return series is like. The moments, skewness and
kurtosis is presented in table 6.2. For a theoretical introduction to these
measures see section 3.2. Noteworthy is the kurtosis which is considerably
larger than three.
6.2.3 Correlation plots for lyoe 53

6.2.3 Correlation plots for lyoe

Three interesting preestimation plots for the price, return and the squared
return are presented in figure 6.1, figure 6.2 and figure 6.3 respectively. The
autocorrelation and partial autocorrelation is presented in figure 6.4 and in
figure 6.5 respectively.

Suez−Lyonnaise des Eaux

40
35
stock price

30
25
20

0 200 400 600 800

date

Figure 6.1. Price plot for Suez-Lyonnaise des Eaux, showing the daily closing
prices. The data series starts at the 9th July 1999 and ends at the 17th of September
2002.

Suez−Lyonnaise des Eaux

0.10
0.05
return

0.00
−0.10 −0.05

0 200 400 600 800

Figure 6.2. Return plot for Suez-Lyonnaise des Eaux, continuously compounded.
Daily data starts at the 9th July 1999 and end the 17th of September 2002.
54 Suez-Lyonnaise des Eaux

Lyoe
0.016

0.014

0.012

0.01

0.008

0.006

0.004

0.002

0
0 100 200 300 400 500 600 700 800 900

Figure 6.3. Squared return for lyoe.

Lyoe

0.8

0.6
Sample Autocorrelation

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 6.4. Autocorrelation for lyoe. This is the autocorrelation on the squared
return, thus this is a measure of autocorrelation in the volatility.

6.2.4 Hypothesis tests

The indication from the graphical plot that there exists correlation between
data plots is in this subsection tested with the Ljung-Box-test. The test for
heteroskedasticity is performed with an ARCH-test.
6.2.4 Hypothesis tests 55

Lyoe

0.8
Sample Partial Autocorrelations

0.6

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 6.5. Partial autocorrelation for lyoe on the squared return.

Table 6.3. Ljung-Box-Pierce Q-test for lyoe. P-values are given in angle brackets.
With these low p-values we have to reject the hypothesis that no serial correlation
exists.

Q(10) = 98.323 [9.39552e-018 ]

Q(15) = 99.1289 [2.44655e-015 ]
Q(20) = 101.958 [9.68458e-014 ]

Ljung-Box-Pierce Q-test

The computed statistical value for the specific lag is presented together
with the p-value in angle brackets. Results from the test are presented in
table 6.3 for lyoe. The null hypothesis is that no serial correlation exists and
this hypothesis is accepted when the p-value is high. In this case we have
to reject that hypothesis. The correlation is tested for the squared return.

Engle’s ARCH-test

Engle’s ARCH-test performed before estimation. This test supports that

there exists heteroskedasticity. The result for lyoe is presented in table 6.4.
ARCH-effects are tested on lags 1-2. With these p-values we reject the
hypothesis that no ARCH-effects exist.
56 Suez-Lyonnaise des Eaux

Table 6.4. Engle’s ARCH-test for lyoe on the raw return series. H0 is that no
ARCH effect exists and in this case we have p-values equal to zero which means
that we can reject the hypothesis.

ARCH 1-2 test: F(2,693) = 20.045 [0.0000]**

Table 6.5. Selection for symmetric GARCH for lyoe. The log-likelihood is pre-
sented as well as Akaike’s information criterion for different p and q. Six different
forecasts measures are also presented to determine what p and q that perform best
on out-of-sample data for the symmetric GARCH.

GARCH(p,q) (0,0) (1,1) (1,2) (2,1) (2,2)

Log −1404.37 −1371.08 −1368.17 −1369.1 −1368.35
Akaike 4.018192 3.928801 3.923349 3.926012 3.926722
R2 0.0000 0.203696 0.164474 0.182692 0.165933
MSE 820.9542 611.5326 620.8024 622.5953 621.2564
MedSE 10.0241 8.4617 8.5447 8.6772 8.6531
MAE 14.1338 11.7340 11.7006 11.7149 11.6941
AMAPE 0.6600 0.5618 0.5565 0.5594 0.5566
TIC 0.8595 0.6010 0.5937 0.6078 0.5955

6.3 Estimation

This section is presenting the whole estimation procedure for lyoe GARCH
models. First p and q for the standard GARCH are chosen and then for these
specific p and q different GARCH models are tested, estimated and finally
one GARCH model with a specific p and q is chosen. For this GARCH(p,q)-
model different distributions are evaluated. Finally the parameters for the
chosen GARCH model are presented.

6.3.1 Choosing p and q

Different p and q for the standard symmetric GARCH model are tested. The
selection procedure is based on each model’s ability to produce forecasts. Six
different forecasts measures are used.
The lyoe table 6.5 shows that the GARCH(1,1) performs best in three
out of six out-of-sample measures. The choice for future model selection is
p = 1, q = 1.
6.3.2 Choosing model 57

Table 6.6. GARCH model selection for lyoe. All models are estimated for p = 1,
q = 1.

GARCH EGARCH APARCH IGARCH

Log −1371.08 −1367.3 −1366.16 −1384.88
Akaike 3.928801 3.923722 3.920469 3.965381
R2 0.203696 0.111235 0.260001 0.221219
MSE 611.5326 684.4780 549.2374 557.8387
MedSE 8.4617 12.0363 18.0764 26.4158
MAE 11.7340 12.5614 11.4671 12.8462
AMAPE 0.5618 0.5742 0.5580 0.5455
TIC 0.6010 0.6687 0.5351 0.4451

Table 6.7. Distribution selection for lyoe.

Normal Student-t Skewed Stud.-t

Log −1366.16 −1341.01 −1340.99
Akaike 3.920469 3.851465 3.854258
R2 0.260001 0.275744 0.276033
MSE 549.2374 537.4875 536.3566
MedSE 18.0764 18.5647 19.2230
MAE 11.4671 11.4685 11.4730
AMAPE 0.5580 0.5521 0.5520
TIC 0.5351 0.5259 0.5244

6.3.2 Choosing model

Four different GARCH models are tested. In table 6.6 the model AP-
ARCH(1,1) is performing best on out-of-sample measures and therefore the
APARCH(1,1) is used for further model selection.

6.3.3 Choosing distribution

Three different distributions are tested: normal, student-t and skewed stud-
ent-t distribution. The result for the distribution comparison is presented
in table 6.7. The estimated degrees of freedom for the student-t distribution
is 5.610786. For skewed student distribution, the tail coefficient is 5.60036
and the asymmetry coefficient is −0.0116698.
58 Suez-Lyonnaise des Eaux

Table 6.8. Estimated parameters for the chosen model for lyoe. The model is an
APARCH(1,1) estimated with a maximum likelihood estimation.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) −0.077537 0.061286 −1.265 0.2062
Cst(V) 0.460606 0.269931 1.706 0.0884
GARCH(Beta1) 0.689094 0.093886 7.340 0.0000
ARCH(Alpha1) 0.160329 0.048998 3.272 0.0011
APARCH(Gamma1) 0.376513 0.163371 2.305 0.0215
APARCH(Delta) 1.862487 0.578710 3.218 0.0013
Asymmetry −0.011670 0.055067 −0.2119 0.8322
Tail 5.600366 1.110430 5.043 0.0000

6.3.4 Estimated parameters

The best GARCH model in out-of-sample data is the APARCH(1,1). The
parameters for the chosen model is presented in table 6.8. The APAR-
CH(Gamma1) is the coefficient which takes the leverage effect into ac-
count. This parameter in this case is not significant. For a description
about the APARCH-model see section 3.7.4. For a more detailed discus-
sion about the leverage effect see section 3.1.4. This table is explained in
detail in Section 3.3.4. One thing that is also interesting to note is that the
APARCH(Gamma1) and the asymmetry coefficient are shown to be time
dependent in the Nelson test.

6.3.5 Are the coefficients time dependent?

When the asymmetric coefficient and the APARCH(Gamma1) are time de-
pendent it is natural to ask whether or not the model can be improved if a
floating window with fixed size is used.
The symmetric GARCH(1,1) coefficients are presented in figure 6.6 and
figure 6.7. The plot is from data, where data points are numbered with an
offset of 700 points for the ARCH and GARCH coefficients including ± one
standard deviation.
A closer look at figure 6.6 shows that something interesting is happening
in days (60-70). The Nelson test does not reject the hypothesis that the
coefficients are time independent, when tested over the whole data set (952
data points). An interesting thing could be to estimate over this specific
period with different window sizes to see if this can improve the forecasting
ability of the model. The forecast period is set to 132 days, out-of-sample
data. First all data (700 data points) are used for the estimation then
the window size is shrunken in steps, with window sizes 700, 500, 300 and
150. The result from this estimating does not support the use of smaller
Validation 59

estimation windows.
1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140

Figure 6.6. Time dependent coefficient plot for lyoe. Window size = 200.

0.8

0.7

0.6
coefficient value

0.5

0.4

0.3

0.2

0.1
0 20 40 60 80 100 120 140

Figure 6.7. Time dependent coefficient plot for lyoe. Window size = 700.

6.4 Validation
In this section the chosen GARCH model is validated. One part of the val-
idation procedure has already been taking part when the model was chosen
on its forecasting ability. Other tests are performed to see if the autocorre-
lation in the squared return has successfully been removed. All these tests
60 Suez-Lyonnaise des Eaux

Table 6.9. Estimated parameters for lyoe on two different data sets. Data set
(1-299) has a skewed student distribution, with a tail coefficient of 4.77673 and
an asymmetry coefficient of 0.190183. Data set (300-832) has a skewed student
distribution, with a tail coefficient of 8.0258 and an asymmetry coefficient of -
0.200555.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) −0.033879 0.100848 −0.3359 0.7372
Cst(V) 0.172707 0.167201 1.033 0.3025
GARCH(Beta1) 0.790681 0.149073 5.304 0.0000
ARCH(Alpha1) 0.111767 0.053624 2.084 0.0380
APARCH(Gamma1) −0.238023 0.529053 −0.4499 0.6531
APARCH(Delta) 0.578116 0.536007 1.079 0.2817
Asymmetry 0.190183 0.087422 2.175 0.0304
Tail 4.776735 1.264920 3.776 0.0002
Cst(M) −0.183533 0.069957 −2.624 0.0090
Cst(V) 0.172982 0.062135 2.784 0.0056
GARCH(Beta1) 0.809172 0.040062 20.20 0.0000
ARCH(Alpha1) 0.135793 0.029269 4.639 0.0000
APARCH(Gamma1) 0.998766 0.097638 10.23 0.0000
APARCH(Delta) 1.350461 0.252826 5.341 0.0000
Asymmetry −0.200555 0.063555 −3.156 0.0017
Tail 8.025805 2.685321 2.989 0.0029

and measures will be presented in the following sections.

6.4.1 Parameter time dependency

In table 6.9, where coefficient values for the APARCH model is displayed, it
is obvious that the parameter values are changing over time. The Asymme-
try coefficient is in the first period 0.19 and in the other one −0.20; in both
cases significantly unequal to zero, so either we take this model, meaning
that we have to reestimate the model when new information is available,
or, if we are interested in a more robust model, we could use the standard
symmetric GARCH(1,1) with normal distribution. It is not that bad con-
sidering its simplicity. Table 6.10 presents the estimates for two different
samples. Here at least the parameters are a bit more stable.

6.4.2 Autocorrelation
As seen in figure 6.8 most of the significant autocorrelation has been re-
moved. Compared this to the result from the preestimation (see figure 6.4).
6.4.3 How is zt distributed? 61

Table 6.10. Estimated parameters for lyoe on two different samples of the sym-
metric GARCH(1,1) model. Data set one is data points (1-416) and the other data
set is (417-832), both with a normal distribution.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) −0.018987 0.072974 −0.2602 0.7949
Cst(V) 0.608372 0.210834 2.886 0.0041
GARCH(Beta1) 0.606972 0.099067 6.127 0.0000
ARCH(Alpha1) 0.190916 0.063013 3.030 0.0026
Cst(M) −0.069178 0.085035 −0.8135 0.4164
Cst(V) 0.257725 0.116921 2.204 0.0281
GARCH(Beta1) 0.798054 0.058151 13.72 0.0000
ARCH(Alpha1) 0.163000 0.047141 3.458 0.0006

Sample Autocorrelation Function (ACF)

0.8

0.6
Sample Autocorrelation

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 6.8. Autocorrelation for lyoe with standardised residuals after estimation.

6.4.3 How is zt distributed?

A graphical presentation of the standardised distribution is presented in
figure 6.9 and the calculated first two moments, skewness and excess kurtosis
is displayed in table 6.11. The first and the second moment are close to their
expected values.

6.4.4 Ljung-Box-Pierce Q-test

The Lbq-test in the validation part tests if there still is or is not a significant
correlation in the standardised innovations. Standardised innovations are
ε2t
h2
= zt . For results on the Lbq-test, see table 6.12. The test is made on
t
62 Suez-Lyonnaise des Eaux

Table 6.11. Moments, skewness and kurtosis for the standardised residuals in the
data set of lyoe.

lyoe
First moment 0.01464
Second moment 1.00107
Skewness -0.065131
Excess kurtosis 2.6564

60

50

40

30

20

10

0
−3 −2 −1 0 1 2 3

Figure 6.9. Histogram for lyoe after estimation with standardised residuals. The
histogram is divided in 100 bins.

zt with lags 10, 15 and 20. Each row in the table represents one test with
a specific lag, p-values are given within angle brackets i.e., H0 has to be
accepted. The theory for the test is presented in Section 3.3.2.

Table 6.12. Ljung-Box-Pierce Q-test on the squared residuals, validation for lyoe.
H0 is that no significant correlation exists. H0 should be accepted when probability
is high. The test is performed on lags 10, 15, 20.

Q(10) = 9.59469 [0.294633 ]

Q(15) = 11.2472 [0.590121 ]
Q(20) = 13.1222 [0.784262 ]
6.4.5 Engle’s ARCH-test 63

Table 6.13. Engle’s ARCH-test, validation for lyoe. H0 is that no ARCH effect
exists.

ARCH 1-5 test: F(5,687) = 1.1681 [0.3233]

ARCH 1-10 test: F(10,677)= 1.0342 [0.4125]
ARCH 1-15 test: F(15,667)= 0.82823 [0.6463]

6.4.5 Engle’s ARCH-test

Result from Engle’s ARCH-test is presented in table 6.13. This indicates
that we have successfully removed the conditional heteroskedasticity that ex-
isted when the test was performed on the pure return series in Section 6.2.4.
The null hypothesis is that no ARCH effect exists and we cannot reject that
hypothesis.

6.5 Summary
6.5.1 Estimation
Different p,q are tested in the symmetric case. These are then used as fixed
p and q for different GARCH-models. In this case p = q = 1 is the
best out-of-sample choice.
Different models are tested for this specific p and q. The asymmetric-
power-ARCH is the model which performs best among the models
that are tested.
Different distributions for this specific APARCH(1,1) are then tested and
estimated. The skewed student-t distribution is the one that performs
best on out-of-sample measures.
The estimated coefficient APARCH(Gamma1) is significantly different fr-
om zero but the asymmetry coefficient in the distribution is not sig-
nificant.
The autocorrelation of the standardised residuals zt exhibits no evidence
that correlation still exists. The heteroskedasticity has also success-
fully been removed.
The dynamic features for the model have not been fully investigated. Either
the APARCH(1,1) model with skewed student-t distribution is used
and then we have to be aware of the fact that the model has to be
reestimated when new information is available, or, if we want a more
robust model, we can use the standard symmetric GARCH(1,1) with
normal distribution.
Chapter 7

OMX index

In this chapter model estimation for the Swedish OMX index is performed.
The procedure is following the structure presented in the methodology chap-
ter. Preestimation and estimation is performed. Finally some conclusions
are drawn.

7.1 Presenting data

7.1.1 Introducing OMX
The Swedish index OMX consists of 30 large Swedish companies. The anal-
ysed data is from the the 5th of January 1998 to the 21st of March 2002.
The data consists of 1051 data points (price data). Daily closing data is then
1050 data points (return data). Especially interesting days are presented in
table 7.1. The selection criterion is that the absolute return should be above
5%, smaller than the selection criterion on the stocks since the OMX index
returns are more smoothed. The days from the 11th September to the 17th
September 2001 are not included in the index at all. The data is presented
in the form (dd.mm.yy).

7.2 Preestimation
7.2.1 Data formatting
For the data series of OMX, data points (1-800) are used in the estimation
part and data points (801-1050) are used as cross validation. The prices
series is converted to a return series via equation (4.1).

7.2.2 Moments for the return series

In the preestimation part it would be interesting to get a picture of what the
distribution of the raw return series looks like. The moments, skewness and
66 OMX index

Table 7.1. Days with high return in the OMX index.

data return (%) date

192 −6.8848 08.10.98
194 11.0228 12.10.98
237 −5.9640 10.12.98
485 5.6546 08.12.99
639 −5.2604 21.07.00
704 −5.1285 23.10.00
753 6.6036 04.01.01
800 −7.2006 12.03.01
818 5.4310 05.04.01
825 5.7336 18.04.01
929 5.3296 24.09.01

Table 7.2. Moments, skewness and kurtosis for the return series of OMX.

OMX
First moment 0
Second moment 3.3578
Third moment 0.7126
Fourth moment 52.7191
Skewness 0.1158
Kurtosis 4.6760

kurtosis are presented in table 7.2. For a theoretical introduction to these

measures see section 3.2. Noticeable is the kurtosis is much larger than
three, which would be the kurtosis if the series were normally distributed.

7.2.3 Correlation plots for OMX

The three interesting preestimation plots of the price, return and the squa-
red return are presented in figure 7.1, figure 7.2 and figure 7.3 respectivily.
The autocorrelation and partial autocorrelation is presented in figure 7.4
and figure 7.5.

7.2.4 Hypothesis tests

From the graphical plots we could assume that there exists correlation be-
tween data points. This is in this subsection tested with the Ljung-Box-test.
The test for heteroskedasticity is performed with an ARCH-test.
7.2.4 Hypothesis tests 67

1600

1400

1200

1000

800

600

400
0 200 400 600 800 1000 1200

Figure 7.1. Price plot for OMX with daily closing prices.

12

10

−2

−4

−6

−8
0 200 400 600 800 1000 1200

Figure 7.2. Return plot for OMX.

Ljung-Box-Pierce Q-test

The computed statistical value for a certain lag is presented together with
the p-values in angle brackets in table 7.3 for OMX. The null hypothesis is
that no serial correlation exists and this hypothesis is accepted when the
p-value is high. In this case we have to reject that hypothesis.
68 OMX index

140

120

100

80

60

40

20

0
0 200 400 600 800 1000 1200

Figure 7.3. Squared return for OMX. The large values are because of the linear
scaling in the transformation from price series to return series.

Sample Autocorrelation Function (ACF)

0.8

0.6
Sample Autocorrelation

0.4

0.2

−0.2
0 5 10 15 20 25 30 35 40 45 50
Lag

Figure 7.4. This is the autocorrelation of the squared return for OMX.

Table 7.3. Ljung-Box-Pierce Q-test for the OMX. P-values are given in angle
brackets. With these low p-values we reject the hypothesis that no serial correlation
exists.

Q(10) = 103.985 [8.65908e-018 ]

Q(15) = 131.561 [1.05139e-020 ]
Q(20) = 158.337 [1.57155e-023 ]
Estimation 69

Sample Partial Autocorrelation Function

0.8
Sample Partial Autocorrelations

0.6

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 7.5. This is the partial autocorrelation of the squared return for OMX.

Table 7.4. Engle’s ARCH-test for OMX on the raw return series. H0 is that no
ARCH effect exists and in this case we have p-values equal to zero which means
that we can reject the hypothesis that no ARCH effects exists.

ARCH 1-2 test: F(2,795) = 25.273 [0.0000]**

ARCH 1-5 test: F(5,789) = 11.652 [0.0000]**

Engle’s ARCH-test
Engle’s ARCH-test performed before estimation. This test supports that
there exists heteroskedasticity. The result for OMX is presented in table 7.4.
ARCH-effects are tested on lags 1-2 and 1-5.

7.3 Estimation
This section is presenting the whole estimation procedure for OMX GARCH
models. First p and q for the standard GARCH is chosen and then for this
specific p and q different GARCH models are tested, estimated and finally
one GARCH model with a specific p and q is chosen. For this GARCH(p,q)-
model different distributions are evaluated. Finally the parameters for the
chosen GARCH models are presented.

7.3.1 Choosing p and q

Different p and q for the standard symmetric GARCH model is tested.
The selection procedure is based on each model’s ability to produce good
70 OMX index

Table 7.5. Selection for symmetric GARCH for OMX. The log-likelihood value is
presented as well as Akaike’s information criterion for different p and q values. Six
different forecast measures are also displayed to determine for which p and q the
performance is best on out-of-sample data.

(p,q) (0,0) (1,1) (1,2) (2,1) (2,2)

Log −1609.06 −1546.45 −1545.88 −1543.49 −1543.357
Akaike 4.027640 3.876132 3.877211 3.871236 3.873393
R2 0.0000 0.0934158 0.0783754 0.0488753 0.0549111
MSE 26.4675 24.1874 24.6836 26.3069 25.9692
MedSE 6.9878 5.7996 5.8391 6.1702 6.1699
MAE 3.3548 3.3171 3.3435 3.4964 3.4665
AMAPE 0.5173 0.4962 0.4971 0.5053 0.5035
TIC 0.5384 0.4580 0.4635 0.4729 0.4706

Table 7.6. GARCH model selection for OMX. The result for EGARCH(1,1) is
omitted because of convergence problems.

GARCH (1,1) APARCH (1,1) IGARCH (1,1)

Log −1546.45 −1538.94 −1547.44
Akaike 3.876132 3.862339 3.876093
R2 0.0934158 0.113057 0.0941127
MSE 24.1874 23.9171 24.8089
MedSE 5.7996 6.6722 6.8080
MAE 3.3171 3.4406 3.4731
AMAPE 0.4962 0.5025 0.5021
TIC 0.4580 0.4429 0.4463

forecasts. Six different forecasts measures are used. Table 7.5 implies that
the GARCH(1,1) performs best in six out of six out-of-sample measures.
Choice for future model selection is p = 1, q = 1.

7.3.2 Choosing model

Four different GARCH models are tested. These models are theoretically
described in Section 3.7. Table 7.6 does not show a clear picture of which
model to use. The APARCH(1,1) model is performing best is three of the
measures and the GARCH(1,1) performs best in three others. Here we use
the AIC to determine that the APARCH(1,1) model is our choice for further
model selection. However GARCH(1,1) is better in three of the measure so
even the GARCH(1,1) seems to do a good job.
7.3.3 Choosing distribution 71

Table 7.7. Distribution selection for OMX. The distributions are normal dis-
tribution, a Student distribution, with 14.8614 degrees of freedom and a Skewed
Student distribution, with a tail coefficient of 13.1282 and an asymmetry coefficient
of −0.103472.

Normal Student-t Skewed Stud.-t

Log −1538.94 −1535.39 −1533.7
Akaike 3.862339 3.855985 3.854248
R2 0.113057 0.115848 0.116233
MSE 23.9171 23.9032 23.7695
MedSE 6.6722 6.9391 6.7925
MAE 3.4406 3.4567 3.4238
AMAPE 0.5025 0.5031 0.5017
TIC 0.4429 0.4407 0.4427

7.3.3 Choosing distribution

The three distributions are: normal, student-t and skewed student-t. The
result for the distribution comparison is presented in table 7.7. The values
measured change very little depending on the distribution. There is no
need for adding to extra parameters in the model when the measures just
differ in the decimals. Our choice for further estimation will be the normal
distribution.

7.3.4 Estimated parameters

The best GARCH model in out-of-sample data is the APARCH(1,1). The
parameters for the chosen model is presented in table 7.8. For a description
of the APARCH-model, see section 3.7.4. This table is explained in detail
in Section 3.3.4. No parameter is time dependent according to the Nelson
test. Noticeable is the APARCH(Delta) value which is lower than for the
other series. All parameters have nice t-values which could be an indication
that it is easier to find good coefficients estimates of GARCH models if the
series is a index instead of a stock.

7.4 Validation
7.4.1 Parameter time dependence
As one can see in table 7.9 the parameters are quite stable through time. The
data is split into two different parts and then the parameters are estimated.
The APARCH(1,1) model with a normal distribution is used. The reason
for the unequal split is due to convergence problems.
72 OMX index

Table 7.8. Estimated parameters for the chosen model for OMX. The model is an
APARCH(1,1) with normal distribution. The data set consists of samples (1-800),
which is the data set used in the estimation part.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) 0.079910 0.055018 1.452 0.1468
Cst(V) 0.041795 0.021145 1.977 0.0484
GARCH(Beta1) 0.903821 0.025832 34.99 0.0000
ARCH(Alpha1) 0.091222 0.022479 4.058 0.0001
APARCH(Gamma1) 0.317723 0.125141 2.539 0.0113
APARCH(Delta) 0.944506 0.295055 3.201 0.0014

Table 7.9. Estimated parameters for the chosen model for OMX. The model is an
APARCH(1,1) with normal distribution using MLE. The data set split in two parts
(1-449) and (450-1050) to see if the parameters are time dependent.

Parameter Coefficient Std.Error t-value t-prob

Cst(M) 0.092859 0.062296 1.491 0.1368
Cst(V) 0.040483 0.025550 1.584 0.1138
GARCH(Beta1) 0.916693 0.046367 19.77 0.0000
ARCH(Alpha1) 0.068163 0.040031 1.703 0.0893
APARCH(Gamma1) 0.711710 0.382442 1.861 0.0634
APARCH(Delta) 1.055108 0.345069 3.058 0.0024
Cst(M) −0.019698 0.073771 −0.2670 0.7896
Cst(V) 0.128395 0.075498 1.701 0.0895
GARCH(Beta1) 0.891609 0.040424 22.06 0.0000
ARCH(Alpha1) 0.058905 0.026495 2.223 0.0266
APARCH(Gamma1) 0.515656 0.307282 1.678 0.0938
APARCH(Delta) 1.163291 0.482020 2.413 0.0161

7.4.2 Autocorrelation
As seen in figure 7.6 most of the significant autocorrelation has been re-
moved. Compared this to the result from the preestimation (see figure 7.4).

7.4.3 Ljung-Box-Pierce Q-test

The Lbq-test in the validation part tests if there still is a significant corre-
lation in the standardised innovations or not. For results on the Lbq-test
see table 7.10. The test is made on zt with lags 5, 10 and 20. Each row in
the table represents one test with a specific lag and H0 has to be accepted.
7.4.4 Engle’s ARCH-test 73

Sample Autocorrelation Function (ACF)

0.8

0.6
Sample Autocorrelation

0.4

0.2

−0.2
0 2 4 6 8 10 12 14 16 18 20
Lag

Figure 7.6. Autocorrelation for OMX with normalised residuals after estimation.

Table 7.10. Ljung-Box-Pierce Q-test on the squared normalised residuals, valida-

tion for OMX. H0 is that no significant correlation exists. H0 should be accepted
when probability is high. The test is performed on lags 5, 10, 20. Here we accept
H0 .

Q(5) = 3.40302 [0.33356 ]

Q(10) = 6.24369 [0.619956 ]
Q(20) = 16.4683 [0.559901 ]

Which means that we successfully have removed the autocorrelation that

existed in the preestimation part.

7.4.4 Engle’s ARCH-test

Result from Engle’s ARCH-test is presented in table 7.11. This indicates
that we have successfully removed the conditional heteroskedasticity that
existed when the test was performed on the pure return series.

Table 7.11. Engle’s ARCH-test, validation for OMX. H0 is that no ARCH effect
exists.

ARCH 1-2 test: F(2,1043)= 1.0572 [0.3478]

ARCH 1-5 test: F(5,1037)= 0.64800 [0.6631]
ARCH 1-10 test: F(10,1027)= 0.64373 [0.7769]
74 OMX index

Table 7.12. Moments, skewness and kurtosis for zt of OMX, where zt is the
standardised residuals.

OMX
First moment −0.00710
Second moment 1.00129
Skewness 0.052673
Excess kurtosis 0.43265

Moments of zt

The calculated first two moments, skewness and kurtosis are given in ta-
ble 7.12. The distribution for zt is quite close to the normal distribution. A
more formal test is performed with the Jarque-Bera test. The Jarque-Bera
value is 8.6750 with a corresponding p-value of 0.013069 so we cannot reject
the hypothesis that zt is normally distributed. A graphical presentation of
this distribution is presented in figure 7.7 where the values are presented in
a histogram with 50 bins.

90

80

70

60

50

40

30

20

10

0
−3 −2 −1 0 1 2 3 4

Figure 7.7. Histogram for OMX after estimation. This is done on standardised
residuals. The histogram is split into 50 bins.
Summary 75

7.5 Summary
7.5.1 Estimation
Different (p,q) were tested in the symmetric case. The choice is p = q = 1
based on our out-of-sample measures.

Different models were tested using the specific p and q. The asymmetric-
power-ARCH is the model which performs best among the tested mod-
els. The GARCH(1,1) can also be a good choice since it performs
almost as good as the APARCH(1,1) but has fewer parameters.

Different distributions for this specific APARCH(1,1) are tested and esti-
mated. There is some indication for the use of skewed-t distribution
but the difference is so small that this is rejected in favour of the
normal distribution due to its simplicity.

The estimated coefficient value for the asymmetry parameter is significant.

This parameter is positive which means that the risk is increasing
more when the index is moving downwards compared to how much
the risk is increasing when the index moves upwards. This could then
be related to the leverage effect.

The data set of an index is more homogenous than a specific stock data
set. Since all effects are more smoothed out within an index. This is
seen in the stability of the parameters in the data set when it is split
and estimated as two separate parts.
Chapter 8

Conclusions

This chapter presents the general conclusions that can be drawn from the
estimation chapters. A discussion about of lags, different models and dis-
tributions is presented. Suggestions for further studies and some reflections
for the GARCH models are also given.

8.1 GARCH modelling discussion

8.1.1 Testing for different (p,q)
The whole estimation procedure for (p,q) is based on the symmetric standard
GARCH with normal distribution. Five different p and q values are tested:
(0,0), (1,1), (1,2), (2,1) and (2,2). p is the number of GARCH coefficients
and q is the number of ARCH coefficients. P and q equal (0,0) is just for
the naive model that no ARCH effect exists in the model. This is however
never the case in the studied data set; (0,0) is always the worst choice among
the models that are tested. p, q equal (1,1) is by far the most used values
in GARCH research today and our result is also consistent with this. In
three out of four models (1,1) is the best choice. However, in muvg the best
choice is (2,1), adding one more beta coefficient. All these results are based
on out-of-sample measures as the main selection criterion.

8.1.2 Testing different models

Four models are tested for the specific (p,q) chosen in the first estimation
part. These are GARCH, EGARCH, APARCH and IGARCH. In all cases
the APARCH does the best job. This indicates that asymmetric coefficient
and power coefficient are good to implement in the GARCH model. Of
course this cannot be taken as a general result. In some of the estimations
the asymmetric coefficient is not significant and the power parameter is close
to two. This is an indication that GARCH(1,1) can be a good choice in these
cases.
78 Conclusions

8.1.3 Testing different distributions

Three different distributions are tested: normal distribution, student-t and
skewed student-t. The result for the use of different distributions is not
pointing in a clear direction. It is clear that different distributions does
not affect the result to any large extent. Notice also that if one should
use distributions which are not normal, then at least 700 data points have
to be used in the estimation procedure to get accurate estimates for the
parameters that describe the skewness and the tail of the distributions.

8.1.4 Parameter instability

One problem that arises when using models with parameters for asymmetries
and other non-linear features is that parameters depend heavily on which
part of the data set they are estimated over, especially if the sample size is
small. Ideally more than 700 samples should be used. If models are esti-
mated over smaller sample sizes, then another problem that arises is to find
feasible solutions. Especially the EGARCH has problem with convergence.
The dynamic features for the models have not been fully investigated.
If we use the APARCH model then we have to be aware of the fact that
the model has to be reestimated when new information is available. If we
want a more robust model then we use the standard symmetric GARCH
with normal distribution, because it has more stable parameter values.

8.1.5 User’s choice

The final decision of what model to use is in the end the user’s choice. The
APARCH(1,1) is almost always a good choice for the data series that we have
tested, but this model has convergence problems and the coefficients might
be time dependent. If this model is used then the coefficients have to be
reestimated when new data (information) is available. The other possibility
is to use the more robust symmetric GARCH(1,1), which is not the best on
out-of-sample measures but has more time stable parameters.

8.1.6 The OMX index

The data set of the OMX index is more smoother than data for a specific
stock. This is because the index is determined by the behaviour of several
stocks. This is seen in the stability of the parameters in the data set, when
it is split and estimated in two separate parts.

8.1.7 Parameter values

This section contains a comparison between parameter estimates from dif-
ferent stocks. All series are best estimated with an APARCH-model, mostly
Further Studies 79

with lag values p = 1, q = 1. The parameter values are presented in ta-

ble 8.1 where only the common parameters are compared. GARCH(Beta1)
for muvg is not compared in this table because it has two Beta parameters
for the chosen model and a comparison for GARCH(Beta1) for this stock
model could lead to misleading result. The values from the different distri-
bution are not compared because it is only in two of the series that another
distribution than the normal is used. Cst(M) is close to zero as one could
expect and the value for all different models are in the same region. Cst(V)
should be positive since this is the variance of εt if we would have if p = 0,
q = 0. The two series with highest Cst(V) values also have large standard
deviation. The GARCH(Beta1) parameter has values between 0.7 and 0.9,
which are common values for this parameter. The ARCH(Alpha1) is in the
interval 0.05 − 0.16, all with quite low estimated standard deviations.
In three of the models the APARCH(Gamma1), which is the asymmetry
model coefficient, has a value in the region of 0.3. The fourth model’s
APARCH(Gamma1) value is negative which means that the effect of a shock
upwards affects the volatility more than an effect downwards. Bear in mind
though, that this coefficient value is not significantly unequal to zero. The
APARCH(Delta) coefficient, which in the symmetric GARCH model is equal
to two, has a value less than two in all of the models.

8.2 Further Studies

The model for the mean equation used in this thesis uses a constant for the
return series. This is consistent with the efficient market theory, despite
the fact that investors also require extra return for taking extra risk. One
improvement to this model could be to do further analysis of the model for
the mean equation. This could be done with a ARMAX-model for example.
The measures in this thesis are made with εt as the ”true” volatility. It
is not obvious that this is the best choice. For example intra-day data could
be used instead of daily data.

8.2.1 Different regimes

The problem with time dependent estimated coefficients could be an indi-
cation that the stock market behaves differently (with respect to volatility)
in different periods of time. One idea could be that one GARCH model
should be used when the stock market is in a bull period and that another
GARCH model should be used when there is a bear period. This could be
modelled with regime switching models together with GARCH models. It
has not been done in this thesis but could be an interesting model region
for further analysis.
80 Conclusions

Table 8.1. Parameter comparison for APARCH models. First row is Volkswagen,
second is muvg, third is lyoe and fourth is OMX.

Parameter Coefficient Std.Error

Cst(M) −0.004560 0.068735
Cst(M) −0.019815 0.072585
Cst(M) −0.077537 0.061286
Cst(M) 0.079910 0.055018
Cst(V) 0.654696 0.413332
Cst(V) 0.068825 0.047832
Cst(V) 0.460606 0.269931
Cst(V) 0.041795 0.021145
GARCH(Beta1) 0.710141 0.075228
GARCH(Beta1) - -
GARCH(Beta1) 0.689094 0.093886
GARCH(Beta1) 0.903821 0.025832
ARCH(Alpha1) 0.133134 0.047949
ARCH(Alpha1) 0.046485 0.024051
ARCH(Alpha1) 0.160329 0.048998
ARCH(Alpha1) 0.091222 0.022479
APARCH(Gamma1) −0.133827 0.120061
APARCH(Gamma1) 0.298389 0.140086
APARCH(Gamma1) 0.376513 0.163371
APARCH(Gamma1) 0.317723 0.125141
APARCH(Delta) 1.957050 0.780077
APARCH(Delta) 1.899521 0.447514
APARCH(Delta) 1.862487 0.578710
APARCH(Delta) 0.944506 0.295055

8.3 Generalization
To draw general conclusions from only four data samples is of course dub-
ious. Still it is fairly obvious that small p and q values, i.e. (1,1) are a
good choice. It is also notable that the APARCH model seems to do a good
job for all of the data series. The value of using different distributions is
not obvious; mostly the normal distribution does a good job in light of its
simplicity.

8.4 How good are the estimated models?

We will finish this thesis with a section that presents the explanation power
for the chosen GARCH models that we have estimated.
How good are the estimated models? 81

Table 8.2. Model explanation power overview for Volkswagen models.

GARCH(0,0) GARCH(1,1) APARCH(1,1)

Log −1726.53 −1693.55 −1692.588
Akaike 4.321326 4.243863 4.246470
R2 0.0000 0.09341 0.115478
MSE 237.4908 203.1438 199.5250
MedSE 15.7309 15.6039 15.5668
MAE 8.2124 8.0538 7.9957
AMAPE 0.5570 0.5356 0.5339
TIC 0.7123 0.5779 0.5732

Table 8.3. Model explanation power overview for muvg models. The distribution
for the APARCH(2,1) model is a student-t.

GARCH(0,0) GARCH(1,1) APARCH(2,1)

Log −1595.2 −1548.66 −1520.52
Akaike 4.563421 4.436184 4.367206
R2 0 0.177904 0.219198
MSE 405.4252 302.5713 286.8509
MedSE 28.2512 24.8449 28.8216
MAE 10.9956 10.6565 10.5239
AMAPE 0.5834 0.5355 0.5297
TIC 0.7169 0.4731 0.4641

In all our four stocks there exists heteroskedasticity and significant au-
tocorrelation in the squared return1 . Both the heteroskedasticity and the
autocorrelation are satisfactorily removed. This is seen in the validation
part for the GARCH models when zt is tested for heteroskedasticity and
autocorrelation.
In tables 8.2, 8.3, 8.4 and 8.5 an overview of the estimated models,
with respect to their explanation power, is given. In the first column the
GARCH(0,0) is presented. This is just as a benchmark that these GARCH
model are doing a good job. In the second column the GARCH(1,1) is
presented. In the third column the chosen model for the specific stock is
displayed. The model that has the best out-of-sample measure is marked in
bold numbers. As one can see, our model choice in the third column has a
high explanation power compared to the other two columns.

1
This is calculated when the constant return value has been removed from the series.
82 Conclusions

Table 8.4. Model explanation power overview for lyoe models. The APARCH(1,1)
model is model together with a skewed student-t distribution.

GARCH(0,0) GARCH(1,1) APARCH(1,1)

Log −1404.37 −1371.08 −1340.99
Akaike 4.018192 3.928801 3.854258
R2 0.0000 0.203696 0.276033
MSE 820.9542 611.5326 536.3566
MedSE 10.0241 8.4617 19.2230
MAE 14.1338 11.7340 11.4730
AMAPE 0.6600 0.5618 0.5520
TIC 0.8595 0.6010 0.5244

Table 8.5. Model explanation power overview for OMX models. The
APARCH(1,1) is with a normal distribution.

GARCH(0,0) GARCH(1,1) APARCH(1,1)

Log −1609.06 −1546.45 −1538.94
Akaike 4.027640 3.876132 3.862339
R2 0.0000 0.0934158 0.113057
MSE 26.4675 24.1874 23.9171
MedSE 6.9878 5.7996 6.6722
MAE 3.3548 3.3171 3.4406
AMAPE 0.5173 0.4962 0.5025
TIC 0.5384 0.4580 0.4429
References

Bollerslev, T. (1986), ‘Generalized autoregressive conditional heteroskedas-

ticity’, Journal of Econometrics 31, 307–327.

Brooks, C. (1997), ‘Linear and non-linear (non)-forecastability of high-

freqency exchange rates’, Journal of Forecasting 16, 125–145.

Brooks, C., Burke, S. P. & Persand, G. (2001), ‘Benchmarks and the accu-
racy of garch model estimation’, International Journal of Forecasting
17, 45–56.

Deistler, M. (1999), ‘Geschichte und Perspektiven der ökonometrie’, RWI-

Mitteilungen 50(3), 109–127.

Engle, R. F. (1982), ‘Autoregressive conditional heteroskedasticity with es-

timates of the variance of United Kingdom inflation’, Econometrica.

Hamann, E. (2001), Time series models for the West German economy,
Master’s thesis, Technische Universität Wien.

Hamilton, J. D. (1994), Time Series Analysis, 4 edn, Princeton University

press.

Hansen, B. E. (1994), ‘Autoregressive conditional density estimation’, In-

ternational Economic Review 35(3), 705–730.

Hull, J. C. (2000), Options, futures, & other derivatives, 4 edn, Prentice

Hall International Inc. International Edition.

Klaasen, F. (2002), ‘Improving GARCH volatility forecasts with regime

switching GARCH’, Empirical Economics 27, 363–394.

Laurent, S. & Peters, J. (2002), A tutorial for G@RCH 2.3, a complete

Ox Package for Estimating and Forecasting ARCH Models, http://-
www.egss.ulg.ac.be/garch.

LeRoy, S. F. (1989), ‘Efficient capital markets and martingales’, Journal of

Economic Leterature 27, 1583–1621.
84 References

Ljung, L. (1999), System Identification: Theory For the User, 2 edn,

Prentice-Hall, Upper Saddle River, N.J. USA.

MathWorks (2002), Statistics Toolbox User’s Guide, 4 edn, The MathWorks

Inc.

Nelson, D. B. (1991), ‘Conditional heteroskedasticity in asset returns - a

new approch’, Econometrica 59(2), 347–370.

Pagan, A. R. & Schwert, G. W. (1990), ‘Alternative models for conditional

stock volatility’, Journal of econometrics 45, 267–290.

Peters, J.-P. (2001), Estimating and forecasting volatility of stock indices

using asymmetric GARCH models and (skewed) student-t densities,
Technical report, University de Liège, Boulevard du Rectorat, 7 (Bat.31
/ Boı̂te 54) 4000 Liège Belgium.

Twain, M. (1894), The Tradgedy of Pudd´nhead Wilson,, Oxford University

Press, New York. Chapter 13. Cited in Dangers of data mining: The
case of calendar effects in stock returns, Journal of Econometrics 105
(201).

www.itl.nist.gov (2002), ‘http://www.itl.nist.gov/div898/handbook/

pmc/section4/pmc446.htm’. Visited 2002-09-10.
På svenska

Detta dokument hålls tillgängligt på Internet – eller dess framtida ersättare –

under en längre tid från publiceringsdatum under förutsättning att inga extra-
ordinära omständigheter uppstår.
Tillgång till dokumentet innebär tillstånd för var och en att läsa, ladda ner,
skriva ut enstaka kopior för enskilt bruk och att använda det oförändrat för
ickekommersiell forskning och för undervisning. Överföring av upphovsrätten
vid en senare tidpunkt kan inte upphäva detta tillstånd. All annan användning av
dokumentet kräver upphovsmannens medgivande. För att garantera äktheten,
säkerheten och tillgängligheten finns det lösningar av teknisk och administrativ
art.
Upphovsmannens ideella rätt innefattar rätt att bli nämnd som upphovsman i
den omfattning som god sed kräver vid användning av dokumentet på ovan
beskrivna sätt samt skydd mot att dokumentet ändras eller presenteras i sådan
form eller i sådant sammanhang som är kränkande för upphovsmannens litterära
eller konstnärliga anseende eller egenart.
För ytterligare information om Linköping University Electronic Press se
förlagets hemsida http://www.ep.liu.se/

In English

The publishers will keep this document online on the Internet - or its possible
replacement - for a considerable time from the date of publication barring
exceptional circumstances.
The online availability of the document implies a permanent permission for
anyone to read, to download, to print out single copies for your own use and to
use it unchanged for any non-commercial research and educational purpose.
Subsequent transfers of copyright cannot revoke this permission. All other uses
of the document are conditional on the consent of the copyright owner. The
publisher has taken technical and administrative measures to assure authenticity,
security and accessibility.
According to intellectual property law the author has the right to be
mentioned when his/her work is accessed as described above and to be protected
against infringement.
For additional information about the Linköping University Electronic Press
and its procedures for publication and for assurance of document integrity,
please refer to its WWW home page: http://www.ep.liu.se/

© [Jens Näsström]