Ngailo Edward

MODELLING AND FORECASTING USING TIME
SERIES GARCH MODELS: AN APPLICATION OF

TANZANIA INFLATION RATE DATA
Ngailo Edward
M.Sc. (Mathematical Modelling) Dissertation
University of Dar es Salaam
July 2011
MODELLING AND FORECASTING USING TIME
SERIES GARCH MODELS: AN APPLICATION OF
TANZANIA INFLATION RATE DATA
By
Ngailo Edward
A Dissertation Submitted in (Partial) Fulllment of the Requirements for the
Degree of Master of Science(Mathematical Modelling) of the University of
DaresSalaam
University of Dar es Salaam
July 2011
i
CERTIFICATION
The undersigned certify that they have read and hereby recommend for accep-
tance by the University of Dar es Salaam a dissertation entitled: Modelling
and Forecasting Using Time Series GARCH Models: An Application
of Tanzania Ination Rate Data, in fullment of the requirements for the
degree of Master of Science (Mathematical Modelling) of the University of Dar
es Salaam.
Dr. E.S.Massawe
(Supervisor).
Date:. . . . . . . . . . . . . . . . . . . . . . . .
Dr. E.Luvanda
(Supervisor)
Date:. . . . . . . . . . . . . . . . . . . . . . . .
ii
DECLARATION AND COPYRIGHT
I, Ngailo Edward, declare that this dissertation is my own original work and
that it has not been presented and will not be presented to any other University
for a similar or any other degree award.
Signature:
This thesis is copyright material protected under the Berne Convention, the Copy-
right Act 1999 and other international and national enactments, in that behalf,
on intellectual property. It may not be reproduced by any means, in full or in
part, except for short extracts in fair dealings, for research or private study, crit-
ical scholarly review or discourse with an acknowledgement, without the written
permission of the School of Graduate Studies, on behalf of both the author and
the University of Dar es Salaam.
iii
ACKNOWLEDGEMENT
I would like to oer my sincere gratitude to several individuals who facilitated
the carrying out of this study. Although it is dicult to mention all of them, the
following needs special tributes.
I am so much honoured to thank Almighty God for keeping me healthy and taking
me through my course successfully. My sincere gratitude goes to my supervisors,
Dr. E.S.Massawe and Dr. E.Luvanda for support, guidance, encouragement and
constructive ideas towards the accomplishment of this work regardless of their
many responsibilities in academic and social aairs in and out of their universities.
I would like also to express my sincere thanks to Prof.Matti Heilio who taught
me time series analysis and stochastic models.
I am also indebted to sta of Tanzania National Bureau of Statistics for provid-
ing me with data, which really helped so much to accomplish this study. I am
indebted to Deputy Principal (Academic) of Dar es Salaam University College of
Education for sponsorship through STHEP project, which enabled me to pursue
the course.
I also extend my gratitude to the sta of the Department of Mathematics at the
University of Dar es Salaam who taught M.Sc(Mathematical Modelling) degree
programme from 20092011 for their dedicated commitment. I would like to say
thanks to my fellow masters students at the Department of Mathematics during
my studies. Their cooperative spirit and contribution during the whole period of
my study is appreciated.
Lastly my special thanks goes to my brothers for their prayers and support.
iv
DEDICATION
To my parents the late Sipriana Mwalongo and Kanuti Ngailo. May our almighty
God rest them in peace.
v
ABSTRACT
The research study is based on nancial time series modelling with special ap-
plication to modelling ination data for Tanzania. In particular the theory of
univariate non linear time series analysis is explored and applied to the ina-
tion data spanning from January 1997 to December 2010. The data was obtained
from Tanzania National Bureau of Statistics. Time series models namely, the
autoregressive conditional heteroscedastic (ARCH) (with their extensions to the
generalized ARCH (GARCH)) models were tted to the data. The stages in
the model building namely, identication, estimation and checking has been ex-
plored and applied to the data. A best tting model was selected based on how
well the model captures the stochastic variation in the data (goodness of t).
The goodness of t is assessed through the Akaike information criteria (AIC) ,
Bayesian information criteria (BIC) and standard error (SE). Based on mini-
mum AIC and BIC values, the best t GARCH models tend to be GARCH(1, 1)
and GARCH(1, 2). After estimation of the parameters of selected models, a series
of diagnostic and forecast accuracy test were performed. Having satised with
all the model assumptions, GARCH(1, 1) model was judged to be the best model
for forecasting. Based on the selected model, we forecasted twelve (12) months
ination rates of Tanzania in-sample period (that is from January 2010 to De-
cember 2010). From the results, it has been observed that the forecasted series
are close to the actual series.
vi
TABLE OF CONTENTS
Certication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Declaration and Copyright . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
1 CHAPTER ONE
INTRODUCTION 1
1.1 General Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Ination performance in Tanzania . . . . . . . . . . . . . . 3
1.2 Statement of the problem . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Research objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 General objective . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Specic objectives . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Signicance of the study . . . . . . . . . . . . . . . . . . . . . . . 6
vii
2 CHAPTER TWO
LITERATURE REVIEW 7
2.1 Overview of Related Literature . . . . . . . . . . . . . . . . . . . 7
3 CHAPTER THREE
CONDITIONAL HETEROSCEDASTICITY: ARCH-GARCHMOD-
ELS 14
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 ARCH(q) Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.1 ARCH(1) Model . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.2 Estimation of the ARCH(1) and the ARCH(q) Models . . 17
3.2.3 Forecasting with the ARCH model . . . . . . . . . . . . . 20
3.3 The GARCH Model . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.1 Estimation of the GARCH(1, 1) . . . . . . . . . . . . . . . 24
3.3.2 The GARCH(p, q) Model . . . . . . . . . . . . . . . . . . . 27
3.3.3 Estimation of GARCH(p, q) Model . . . . . . . . . . . . . 31
3.3.4 Model Checking . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.5 Forecasting with GARCH(p, q) models . . . . . . . . . . . 32
3.4 Model selection criteria . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.1 The Akaike Information Criterion . . . . . . . . . . . . . . 33
viii
3.4.2 The Bayesian information criterion . . . . . . . . . . . . . 34
3.5 Forecasting performance . . . . . . . . . . . . . . . . . . . . . . . 35
4 CHAPTER FOUR
DATA ANALYSIS AND METHODOLOGY 37
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Data description and basic statistics . . . . . . . . . . . . . . . . . 37
4.3 Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.4 Testing for ARCH eects and Serial correlation in the return series 41
4.5 Model estimation and evaluation . . . . . . . . . . . . . . . . . . 44
4.5.1 Model selection and analysis . . . . . . . . . . . . . . . . . 44
4.5.2 A Comparison of the tted GARCH models . . . . . . . . 48
4.6 Diagnostic checking of the GARCH(1, 1) model . . . . . . . . . . 49
4.7 Forecasting with the GARCH(1, 1) model . . . . . . . . . . . . . . 52
4.7.1 Forecast Evaluation and Accuracy Criteria . . . . . . . . . 53
5 CHAPTER FIVE
DISCUSSION, CONCLUSION AND FUTURE WORK 57
5.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.3 Suggestions for future work . . . . . . . . . . . . . . . . . . . . . 58
ix
6 REFERENCES 59
7 APPENDIX 64
x
LIST OF FIGURES
4.1 Time plot of monthly ination in Tanzania . . . . . . . . . . . . . 39
4.2 First Dierence of Log of CPI . . . . . . . . . . . . . . . . . . . . 40
4.3 ACF with Bounds for Raw Return Series . . . . . . . . . . . . . . 43
4.4 PACF with Bounds for Raw Return Series . . . . . . . . . . . . . 43
4.5 ACF of the squared Returns . . . . . . . . . . . . . . . . . . . . . 44
4.6 Plot of Return, Estimated Volatility and Innovations (residuals) . 49
4.7 Time plot of residuals from GARCH(1,1) . . . . . . . . . . . . . . 50
4.8 ACF of the squared standardized residuals . . . . . . . . . . . . . 51
4.9 (a)Histogram of residuals(top) and (b)Normal probability(bottom)
plot from GARCH(1,1) . . . . . . . . . . . . . . . . . . . . . . . 55
4.10 Time plot of Ination rate and one year forecasts from GARCH(1,1) 56
xi
LIST OF TABLES
4.1 Descriptive Statistics for the Ination rate series. . . . . . . . . . 38
4.2 Ljung-Box-Pierce Q-test for autocorrelation: (at 95% condence). 41
4.3 Ljung-Box Q-test for squared returns: (at 95% condence). . . . . 42
4.4 Engle ARCH test for heteroscedasticity: (at 95% condence). . . . 42
4.5 Comparison of suggested GARCH models. . . . . . . . . . . . . . 45
4.6 Parameter estimates for GARCH(1,1). . . . . . . . . . . . . . . . 46
4.9 Ljung-Box-Pierce Q-test on standardized innovations. . . . . . . . 52
4.10 Engle ARCH test on standardized innovations. . . . . . . . . . . . 52
4.11 Forecast Accuracy Test on the most likely suggested GARCH models. 53
4.12 One year in-sample forecasts of Ination obtained from the GARCH(1,1). 54
xii
ABBREVIATIONS
ACF Aucorrelation Function
ADF Augumented Dickey Fuller
AIC Akaike Information Criteria
AR AutoRegressive
ARCH AutoRegressive Conditional Heteroscedastic
ARIMA AutoRegressive Integrated Moving Average
BIC Bayesian Information Criteria
CPI Consumer Price Index
EGARCH Exponential Generalized AutoRegressive Conditional Heteroscedastic
GARCH Generalized AutoRegressive Conditional Heteroscedastic
IGARCH Integated Generalized AutoRegressive Conditional Heteroscedastic
MAE Mean Absolute Error
RMSE Root Mean Square Error
TNBS Tanzania National Bureau of Statistics
PACF Partial Autocorrelation Function
QTM Quantum Theory of Money
SARIMA Seasonal AutoRegressive Integrated Moving Average
CHAPTER ONE
INTRODUCTION
1.1 General Introduction
Time, in terms of years, months, days or hours is a device that enables one to
relate phenomena to a set of common, stable reference points. In making con-
scious decisions under uncertainty, we all make forecasts. Almost all managerial
decisions are based on some form of forecast. Essentially the concept of time
series is based on the historical observations. It involves explaining past obser-
vations in order to try to predict those in the future (Ahiati, 2007 ). A time
series is a collection of observations measured sequentially through time. These
measurements may be made continuously through time or be taken at a discrete
set of time points (Chateld, 2000 ).
In recent years, ination has become one of the major economic challenges fac-
ing most countries in the world especially those in Africa including Tanzania.
Ination is a major focus of economic policy worldwide as described by David
(2001). Ination dynamics and evolution can be studied using a stochastic mod-
elling approach that captures the time dependent structure embedded in the time
series ination data. The autoregressive conditional heteroscedasticity (ARCH)
models, with its extension to generalized autoregressive conditional heteroscedas-
ticity (GARCH) models as introduced by Engle (1982) and Bollerslev (1986)
respectively accommodates the dynamics of conditional heteroscedasticity (the
changing variance nature of the data). Heteroscedasticity aects the accuracy of
forecast condence limits and thus has to be handled properly by constructing
1
2
appropriate non-constant variance models (Amos, 2009).
Ination as described by Websters (2000), is the persistent increase in the level of
consumer prices or persistent decline in the purchasing power of money. Ination
can also be expressed as a situation when the demand for goods and services
exceeds their supply in the economy (Hall, 1982). In reality ination means that
your money can not buy as much as what it could have bought yesterday.
The most common way of measuring ination is the Consumer Price Index
(CPI) over monthly, quarterly or yearly. The formula for calculating the ination
data can be dened as follows; Let P
t
be the current average price level of an
economic basket of goods and services and P
t1
be the average price level of the
basket a year ago, then the ination rate I
t
at time t is calculated as ;
I
t
=
P
t
P
t1
P
t1
100%.
An alternative measure is the gross domestic product (GDP) which is usually
taken annually. In Tanzania the CPI is calculated by the National Bureau of
Statistics under the auspices of the Ministry of Finance. The national CPI
covers prices collected in twenty towns in Tanzania mainland. All prices are the
prevailing market prices and are gathered for two hundred and seven items. The
CPI is a dynamic gure obtained by measuring the price variations of numerous
goods and services consumed by typical Tanzanian households on a monthly basis.
Specic weights are assigned to categories of expenditure based on consumer
spending statistics. The dierent types of expenditure are grouped together in a
basket and the high volume spending items carry the most weight and therefore
have the most material impact on the calculated index. Once the CPI is known,
the rate of ination is the rate of change in the CPI over a period for example
3
month-on-month ination rate and usually its units is in percentages. Ination
can be caused by either too much money in circulation in the country or too
few goods oered for sale. People who are living with xed income suer most
when prices of commodities rise, as they cannot buy as much as they could buy
previously (Aidoo, 2010). Unanticipated ination makes future prices unknown,
as this situation discourages savings due to the fact that the money is worth
more at present than in the future. It leads to inecient allocation of economic
resources. Therefore it reduces economic growth, because economy needs a certain
level of savings to nance investments which boosts economic growth.
1.1.1 Ination performance in Tanzania
The historical trend of ination in Tanzania shows that ination has always been
a two-digit gure starting in 1966. In the 1970s, it has been limited to uctuate
between 10% and 20%. At the end of the 1970s and the beginning of 1980, a
radical increase was recorded. Ination rose to the level of 30%. It stabilized at
this level, only dropping to 20% at the end of 1980s. The government of Tan-
zanias strategy for reducing ination has since 1986 focused on tight monetary
policy and increased output production. Ination remained high during these
periods, although at a slightly lower level than the pre-reform level of 22.3% in
1985 (Rutayisire, 1986) The signicant decline in the ination rate since 1994
reects the impact of tight monetary and scal policies pursued by the central
bank and government. Ination declined from 15.5% in December 1996 to 5.5%
in December 2000; nonetheless, it rose to 6.7% in December 2006. The period
between December 2000 and December 2005, the ination went down to 6.2%
and 4.4% respectively. From 6.7% of 2006 ination has been rising and it reached
to 11.6% in 2010 (Tanzania National Bureau of Statistics, 2010)
4
Within time series modelling, there are two approaches available for forecasting;
the univariate and multivariate. In particular this research study has forecasted
in-sample future values of ination time series data using the univariate forecast-
ing approach; in which forecasts depend only on present and past values of a single
series being forecasted. An alternative approach available is a multivariate fore-
casting; in which the forecasts of a given variable depend, on more additional time
series data. The autoregressive conditional heteroscedasticity models are specic
subset of univariate modelling in nancial time series and statistics, in which a
current conditional variance depends not only on the previous observations, but
also on the previous conditional variance. The stochastic models; the autore-
gressive integrated moving average (ARIMA) and the seasonal autoregressive
integrated moving average (SARIMA) models are also other approaches under
univariate modelling. The main dierences from GARCH models come about due
to the assumptions and characteristics of the models, particularly the assumption
of constant variance.
This research study is therefore intended to model long term behaviour of monthly
ination rate data of Tanzania from 1997 to 2010 and predict future values. The
best way to study and understand the underlying phenomenon of the ination
process is to construct practical mathematical stochastic model that accounts for
variations in the process of ination.
1.2 Statement of the problem
The assumption of constant variance over some period when the resulting se-
ries moves or progress through time, statistically is inecient and inconsistent
(Campbell et al., 1997). In real life nancial data for instance ination data,
5
variance changes with time (a phenomenon dened as heteroscedasticity), hence
there is a need of studying models which accommodates this possible variation
in variance. In considering the issue ination modelling and forecasting in Tan-
zania, it has been found that there are many researches done, but no previous
mathematical study has used Tanzania monthly ination rate data to model the
conditional variance and forecast future values. Consequently, this work intends
to use the autoregressive conditional heteroscedasticity (ARCH) models with its
extension to generalized ARCH (GARCH) models to model and accommodate
the dynamics of conditional heteroscedasticity in ination data. Moreover by
nding an appropriate model to represent the data, the study intends to use it
to predict future values based on the past observations.
1.3 Research objectives
1.3.1 General objective
The general objective of the study is to establish the generalized autoregressive
conditional heteroscedasticity time series model to analyse ination data based
on Tanzania.
1.3.2 Specic objectives
The specic objectives of this study are:
To develop time series model (GARCH model) for ination data in Tanza-
nia.
6
To determine the accuracy of the model.
To determine future ination rate.
1.4 Signicance of the study
The research study is of signicant to several parties as follows:
To Academicians and Researchers;
This research study is expected to provide knowledge and basis for further
studies in the same area or other related elds of study.
To policy makers;
The study is intended to contribute to policy making through developing a
model which can be used to forecast ination thus guide policy makers in
formulating macroeconomic policies.
CHAPTER TWO
LITERATURE REVIEW
2.1 Overview of Related Literature
A brief survey on previous works provides the context of the proposed study.
Amos (2009) studied nancial time series modelling using ination data spanning
from January 1994 to December 2008 for South Africa. In the study two time
series models which are the seasonal autoregressive integrated moving average
(SARIMA) model and the generalized autoregressive conditional heteroscedastic-
ity (GARCH) model were tted to the data for encountering trend and seasonal
terms and accomodating time varying variance respectively. A best tting model
for each family of models oering an optimal balance between goodness of t and
parsimony was selected. The SARIMA model of order (1,1,0)(0,1,1) and GARCH
model of order (1,1) were chosen to be the best tting models for determining
the two years forecasts of ination rate of South Africa. However GARCH model
of order (1,1) was observed to be superior in producing future forecasts because
of its ability to capture variations in the data.
Perniagaan, Ekonomi and Nor (2004) explored the varying volatility dynamics
of ination rate data in Malaysia for the period from August 1980 to December
2004. The Generalized autoregressive conditional heteroscedasticity (GARCH)
models and the exponential generalized autoregressive conditional heteroscedas-
ticity (EGARCH) models were used to capture the stochastic variations and
asymmetries in the data. An in-sample evaluation of the sub-periods volatility
7
8
was done. Based from the results using both models the GARCH model of order
(1, 1) and EGARCH model of order (1, 1) produced good estimates of sub-periods
volatility. But due to nature of the data of having highly irregular uctuation
EGARCH model was selected to be the best.
Dijk and Franses (1996) studied the performance of the GARCH model and two of
its non-linear modications to forecast weekly stock market volatility. The models
were the Quadratic GARCH and the Glosten, Jagannathan and Runkle (1992)
or (GJR) models which had been proposed to describe the negative skewness in
stock market indices. Their forecasting results for ve weekly stock market indices
showed that the QGARCH model could signicantly be improved on the linear
GARCH model so that it could be good at calibrating data including extreme
events. Based on the results they concluded that the forecasting of GARCH
type models appeared sensitive to extreme within sample observations. The GJR
model on the other hand was not recommended for forecasting.
Garcia et al., (2003) pointed out that, the generalized autoregressive conditional
heteroscedasticity (GARCH) models consider, the price series as not invariant
(does not have zero mean and constant variance) as it happens in an autore-
gressive integrated moving average (ARIMA) process. In their study GARCH
process measured the implied volatility of a series due to price spikes. Their pa-
per focused on day-ahead forecast of daily electricity markets with high volatility
periods using a GARCH methodology. The GARCH model in their study was to
provide the twenty four (24) forecasts of the clearing prices for the next day based
on historical data. They observed a good performance of the prediction method.
The daily mean errors were around 4%, where the lowest mean error was 2.60%
and the highest one was 7.60%. In general, they concluded that the results ob-
tained by the model were quite reasonable, as the errors obtained were not larger
than 10%. Nevertheless to verify the predictive accuracy of the GARCH model,
9
dierent statistical measures were utilized such as mean week error (MWE) and
forecast mean square error (FMSE).
Assis, Amran and Remali (2006) compared the forecasting performance of dier-
ent time series methods for forecasting cocoa bean prices at Bagan Datoh cocoa
bean. The monthly average data of Bagan Datoh cocoa bean prices graded for
the period of January 1992 to December 2006 was used. Four dierent types of
univariate time series models were compared namely the exponential smoothing,
autoregressive integrated moving average (ARIMA), generalized autoregressive
conditional heteroscedasticity (GARCH) and the mixed ARIMA/ GARCH mod-
els. Root mean squared error (RMSE), mean absolute percentage error (MAPE),
mean absolute error (MAE) and Theils inequality coecient (Ustatistics) were
used as the selection criteria to determine the best forecasting model. The study
revealed that the time series data were inuenced by a positive linear trend fac-
tor while a regression test results showed the non-existence of seasonal factors.
Moreover, the autocorrelation function (ACF) and the Augmented Dickey-Fuller
(ADF) test showed that the time series data was not stationary but became sta-
tionary after the rst dierentiating process was carried out. In their paper they
pointed out that forecasting the future prices of cocoa bean through the most
accurate time series model could help the Malaysia government as well as the
buyers (exporters and millers) and sellers (farmers and dealers) in cocoa bean in-
dustry to perform better strategic planning and also to help them in maximizing
revenue and minimizing the cost. Based on the results of the ex-post forecasting
(starting from January until December 2006) GARCH model performed better
compared to exponential smoothing, ARIMA and the mixed ARIMA/GARCH
model for forecasting Bagan Datoh cocoa bean prices.
Kontonikas (2004) analyzed the relationship between ination and ination un-
certainty in the United Kingdom from 1973 to 2003 with monthly and quarterly
10
data. Dierent types of GARCH Mean (M)-Level (L) models that allow for si-
multaneous feedback between the conditional mean and variance of ination were
used to test the relationship. They found that there was a positive relationship
between past ination and uncertainty about future ination, in line with Fried-
man and Ball (2006) to which in their study of testing for rate of dependence and
asymmetric in ination uncertainty they concluded that there was a link between
ination rate and ination uncertainty.
Jehovanes (2007) studied a time lag between a change in money supply and
the ination rate response. A modied generalized autoregressive conditional
heteroscedasticity (GARCH) model was employed to monthly ination data for
the period 1994 to 2006. In the study he used the maximum likelihood estimation
technique to estimate parameters of the model and to determine signicance of
the lagged value. GARCH model produced good results which indicated that a
change in money supply would aect ination rate considerably in seven months
ahead.
Chong, Loo and Muhammad (2002) used seven years of daily observed Sterling
exchange rate. The GARCH model including the family of GARCH Mean model
were used to explain the commonly observed characteristics of the unconditional
distribution of daily rate of returns series. The results indicated that the hy-
potheses of constant variance model could be rejected, at least within sample,
since almost all the parameter estimates of the ARCH and GARCH models were
signicant at 5 percent level. The Q statistics and the Lagrange Multiplier test
revealed that the use of the long memory GARCH model was preferable to the
short memory and high order ARCH model. The results from various goodness
of t statistics were not consistent for Sterling exchange rates. It appeared that
the BIC and AIC test proposed GARCH models to be the best for within sample
modelling while the mean square error (MSE) test suggested the GARCH mean
11
model to be the best for modelling the heteroscedasticity of daily exchange rates.
Engle (1982) studied on ARCH and GARCH models, and revealed that, these
models were designed to deal with the assumption of non-stationarity found in
real life nancial data. He further pointed out that these models have become
widespread tools for dealing with time series heteroscedastic. The ARCH and
GARCH models treat heteroscedasticity as a variance to be modelled. The goal
of such models is to provide a volatility measure like a standard deviation that
can be used in nancial decisions concerning risk analysis, portfolio selection and
derivative pricing.
Hamilton (1994) carried out a study on the importance of forecasting the condi-
tional variance and pointed out that sometimes we might be interested in forecast-
ing not only the level of the series, r
t
but also its changing variance. He further
described that changes in the variance are quite important for understanding -
nancial markets, since investors require higher expected returns as compensation
for holding riskier assets. A variance that changes over time also has implications
for the validity and eciency of statistical inference about the parameters that
describe dynamics of the level of the series, r
t
.
Chateld (2000), explored in his book that, the idea behind a GARCH model
is similar to that behind ARMA model in the sense that a higher order AR or
MA model may often be approximated by a mixed ARMA model, with fewer
parameters, using a rational polynomial approximation. Thus a GARCH model
can be thought of as an approximation to a higher-order ARCH model. The
GARCH(1, 1) model has become the standard model for describing changing vari-
ance for no obvious reason other than relative simplicity. In practice, if such a
model is tted to data, it is often found that (+) < 1 so that the stationarity
condition may be satised. If + = 1,, then the process does not have nite
12
variance, although it can be shown that the squared observations are station-
ary after taking rst dierences leading what is called an integrated GARCH or
IGARCH model.
Brooks (2008) studied stochastic volatility models and found that most time series
models such as GARCH, will have forecasts that tend towards the unconditional
variance of the series as the prediction horizon increases. This is a good property
for a volatility forecasting model to have, since it is well known that volatility
series are mean-reverting. This implies that if they are at a low level relative to
their historic average they will have a tendency to rise back towards the average.
This feature is accounted for in GARCH volatility forecasting models.
Mugume and Kasende (2009) investigated ination and ination forecasting in
Uganda from 1993 to 2009. They employed various ination forecasting models
like Philips curve, P-star model based on Quantum Theory of Money (QTM),
and the price equation and ARIMA model. In addition, M3 was employed and
the results of both short-run dynamics and long-run equilibrium indicated that
ination had not been a result of money growth. The long-run ination equation
seemed to indicate that exchange rate depreciation could have had a stronger
impact in driving ination upwards than money supply, although it had no short-
run impact.
Igogo (2010) did a study on the eect of real exchange rate volatility on trade
ows in Tanzania for period of 1968 to 2007. The study employed recent ARCH
famly models to measure volatility. Firstly the GARCH(1, 1) model was employed
and found to violate the non-negativity condition. However the study employed
EGARCH(1, 1) model proposed by Nelson(1991) to resolve the problem. The
adequacy of the EGARCH(1, 1) model to measure the real exchange rate volatility
was conrmed by testing for ARCH eect after running the model. Furthermore,
13
the study reveals that it is the real exchange rate rather than its volatility that
is found to have a signicant eect on trade ows although the eect is larger on
exports than on imports. Therefore in the short-run imports are mainly aected
by the domestic income while exports are mainly aected by the real exchange
rate.
CHAPTER THREE
CONDITIONAL HETEROSCEDASTICITY:
ARCH-GARCH MODELS
3.1 Introduction
The analysis of nancial data has received considerable attention in the literature
over the last 20 years. Several models have been suggested for capturing special
features of nancial data, and most of these models have the property that the
conditional variance (or the conditional scaling) depends on the past. One of the
best known and most often used is the autoregressive conditionally heteroscedas-
tic (ARCH) process introduced by Engle, (1982). The ARCH model has been
investigated and generalized by several authors, including Bollerslev (1986) and
Gourieroux (1997). The theoretical results on ARCH and related properties have
played a special role in empirical work in the analysis of data on exchange rates,
stock prices and ination rate data to mention but a few.
In real life nancial data, variance may change with time, therefore there is a need
of studying models which accommodate this possible variation in variance. The
autoregressive conditional heteroscedasticity (ARCH) models, with its extension
to generalized ARCH, (GARCH) models as introduced by Engle and Bollerslev
respectively, accommodates the dynamics of conditional heteroscedasticity. This
is done by relating the current variance errors to the previous errors in the case
of ARCH where as in GARCH, the previous conditional variances are included.
Heteroscedasticity aects the accuracy of forecast condence limits and thus has
14
15
to be handled properly by constructing appropriate non-constant variance models.
The ARCH-GARCH modelling considers the conditional error variance as a func-
tion of the past realization of the series. Campbell et al., (1997) argued that it
is both logically inconsistent and statistically inecient to use and model volatil-
ity measures that are based on the assumption of constant variance over some
period when the resulting series moves or progress through time. In the case of
nancial data for example, large and small errors occur in clusters, which imply
that large returns are followed by more large returns and small returns are further
followed by small returns. In this context of the current study this is equivalent
to saying that periods of high ination are usually followed by further periods of
high ination, while low ination is likely to be followed by much low ination.
The theory of the ARCH and GARCH models are detailed in sections 3.2 and
3.3 respectively.
3.2 ARCH(q) Model
Let {r
t
} be the mean-corrected return or rate of ination,
t
be the Gaussian
white noise with zero mean and unit variance and H
t
be the information set at
time t given by H
t
= {r
1
, r
2
, . . . , r
t1
}. Then the process {r
t
} is ARCH(q) (Engle,
1982) if
r
t
=
t
t
, (3.1)
where
E(r
t|H
t
) = 0, (3.2)
V ar(r
t|H
t
) =
2
t
=
0
+
q
i=1
i
r
2
ti
, (3.3)
16
and the error term
t
is such that
E(
t|H
t
) = 0, (3.4)
V ar(
t|H
t
) = 1. (3.5)
Equations (3.4) and (3.5) show that the error term
t
is a conditionally stan-
dardized martingale dierence dened as follows: A stochastic series {r
t
} is a
martingale dierence if its expectation with respect to past values of another
stochastic series X
i
is zero, that is,
E(r
t+i|X
i
,X
i1
,...,
) = 0
for i = 1, 2, . . . . In this type of the impact of the past on the present volatil-
ity is assumed to be a quadratic function of lagged innovations. The coecient
(
0
,
1
, . . . ,
q
) can consistently be estimated by regressing {r
2
t
} on r
2
t1
, r
2
t2
, . . . , r
2
tq
.
To ensure non-negative volatility we require
0
0,
i
0 for all i = 1, 2, . . . , q.
3.2.1 ARCH(1) Model
The ARCH(1) model is a particular case of the general ARCH(q) model and is
dened as follows: Let {r
t
} be the mean-corrected return and
t
be a Gaussian
white noise with mean zero and unit variance. If H
t
is the information set available
at time t then the process {r
t
} is ARCH(1) if q = 1 given by
r
t
=
t
t
, (3.6)
and the conditional variance
2
t
is given by
2
t
=
0
+
1
r
2
t1
, (3.7)
17
where
0
and
1
are unknown parameters. Thus under the normality assump-
tion of
t
the process can be stated conditionally in terms of H
t
similar to the
variance
2
t
as given above. To ensure non-negativity condition for the condi-
tional variance, the constraints
0
0 and
1
0 must be satised. Equations
(3.6) and (3.7) suggest that a large past squared mean-corrected return implies
a large conditional variance
2
t
resulting in r
t
being large in absolute value. The
ARCH(1) is a special case of ARCH(q) and therefore what applies for ARCH(q)
also applies for ARCH(1) .
3.2.2 Estimation of the ARCH(1) and the ARCH(q) Mod-
els
Based on the assumption of the normality made on the
t
the method of maxi-
mum likelihood estimation is adopted. Let r
1
, r
2
, . . . , r
t
, be a realization from an
ARCH(1) process, then the likelihood of the data can be written as a product of
the conditionals as
f(r
1
, r
2
, . . . , r
t
|) = f(r
t
|r
t1
)f(r
t1
|r
t2
) . . . f(r
2
|r
1
)f(r
1
|), (3.8)
where = (
0
,
1
)
. It is more practical to set condition on r

1
since the form
f(r
1
|) is dicult to obtain. Usually r
1
is assumed to be known and equal to its
observed value. This allows us to use the conditional likelihood given by
f(r
1
, r
2
, . . . , r
t
|; r
1
) = f(r
t
|r
t1
)f(r
t1
|r
t2
) . . . f(r
2
|r
1
)f(r
1
|; r
1
). (3.9)
Since r
t
|H
t
N(0,
2
t
) it follows that
18
f(r
t
|H
t
) =
1
_
2
2
t
exp
_
r
2
t
2
2
t
_
(3.10)
where
2
t
=
0
+
1
r
2
t1
. The conditional log-likelihood is expressed as
=
c
(
0
,
1
|r
1
),
= nf(r
2
, . . . , r
1
; ),
=
1
2
t
i=2
n(2
2
i
)
1
2
t
i=2
_
r
2
t
2
t
_
. (3.11)
The maximum likelihood estimates are obtained by maximizing this function with
respect to
0
and
1
, (Tsay, 2002). Note that the function is non linear in these
parameters and thus its maximization must be done using appropriate non linear
optimization routine. Let a process [r
t
]
T
t=1
be a series generated by an ARCH(1)
process, where T is the sample size. Conditioning on the initial observation, the
joint density function can be written as
f(r) =
T
t=2
f(r
t
|H
t
). (3.12)
To nd the conditional maximum likelihood estimates of
0
and
1
, rst one
needs the derivatives of the conditional log-likelihood with respect to
0
and
1
given by
0
=
1
2
2
t
_
r
2
t
2
t
1
_
0
,
=
1
2
2
t
_
r
2
t
2
t
1
_
2
t

2
t
0
, (3.13)
and
1
=
1
2
2
t
_
r
2
t
2
t
1
_
1
,
19
=
1
2
2
t
_
r
2
t
2
t
1
_
2
t

2
t
1
. (3.14)
More generally the partial derivative of is
=
T
t=2
2
t
2
t
,
=
1
2
T
t=2
_
1
2
t
r
2
t
4
t
_
_
_
_
_
1
r
2
t1
_
_
_
_
,
=
1
2
T
t=2
_
r
2
t
2
t
1
_
1
2
t
_
_
_
_
1
r
2
t1
_
_
_
_
. (3.15)
recalling that
2
t
=
0
+
1
r
2
t1
. Since

2
2
t
= 0, the Hessian is given by
=
T
t=2
4
t
2
t
2
t
,
=
1
2
T
t=2
_
r
2
t
(
2
t
)
3
+
_
r
2
t
2
t
1
_
1
4
t
_
_
_
_
_
1 r
2
t
r
2
t
r
4
t
_
_
_
_
.
The Fisher information matrix denoted by g is dened to be the negative of the
expected value of the Hessian, that is,
g = E
_

2
_
,
since
E
r
t
|H
t
_
r
2
t
2
t
1
_
1
4
t
_
_
_
_
1 r
2
t1
r
2
t1
r
4
t1
_
_
_
_
= 0,
20
and
E
r
t
|H
t
_
r
2
t
(
2
t
)
3
_
=
_
E
r
t
|H
t
r
2
t
(
2
t
)
3
_
.
It then follows that,
g =
1
2
T
t=2
_
1
4
t
_
_
_
_
_
1 r
2
t1
r
2
t1
r
4
t1
_
_
_
_
(3.16)
as in Engle, (1982). Non-linear optimization routines are iterative, thus if
i
denotes the parameter estimates after the i
th
iterations, then
i+1
has the form
i+1
=
i
+M
1
_
_
where is a step-length chosen to maximize the likelihood function in the direc-
tion

. For Newton Raphson based routines = 1 and M =

2
and for the

Fisher scoring method = 1 and M = g (Mills, 1994 and Engle, 1982).
3.2.3 Forecasting with the ARCH model
Forecasting is one of the main aims of developing a time series model. The
ARCH models also provide good estimates of the series before it is realized. We
now provide the theory of forecasting with the ARCH models in detail. Let
r
1
, r
2
, r
3
, . . . , r
t
, be an observed time series , then the step ahead forecast, for
= 1, 2, . . . , at the origin t, denoted as r
t
(), is taken to be the minimum mean
squared error predictor, that is, r
t
() minimizes
E(r
t+
f(r))
2
where f(r) is a function of the observations, then
r
r
() = E[r
t+
|r
1
, . . . , r
t
],
21
(Tsay, 2002)
However for the ARCH(1) model
r
t
() = E[r
t+
|r
1
, . . . , r
t
] = 0
The forecasts for the r
t
series provide no much helpful information. It is therefore
important to consider the squared returns r
2
t
given as
r
2
t
= E[r
2
t+
|r
2
1
, . . . , r
2
t
],
(Shephard, 1996)
Hence the step ahead forecast for r
2
t
is given by
r
2
t
(1) =
0
+
1
r
2
t
which is equivalent to
2
t
(1) = E[
2
t+1
|r
t
]
=
0
+
1
r
2
t
where
0
and
1
are the conditional maximum likelihood estimates of
0
and
1
.
Similarly a 2 step ahead forecast for r
2
t
is given by
r
2
t
(2) = E[r
2
t+2
|r
t
]
= E[
t+2
|r
t
]
=
0
+
1
E[r
2
t+1
|r
t
]
=
0
+
1
(
0
+
1
r
2
t
)
22
=
0
(1 +
1
) + (
1
2
r
2
t
)
=
2
t
(2)
In general the step ahead forecast for the r
2
t
is given by,
r
2
t
() = E[r
2
t+
|r
t
]
=
0
(1 +
1
+
1
2
+. . . +
1
1
) +
1
r
2
t
,
=
2
t
().
The obvious possible problem in using the ARCH formulation is that the approach
can lead to a high parametric model if the lag q is large. This necessitated the
introduction of the GARCH model. Section 3.3 below gives an account of the
theory of the GARCH modelling.
3.3 The GARCH Model
The Generalized ARCH (GARCH), as developed by Bollerslev (1986), is an ex-
tension of the ARCH model similar to the extension of an AR to ARMA process.
When modelling using ARCH, there might be a need for a large value of the
lag q , hence a large number of parameters. This may result in a model with
a large number of parameters, violating the principle of parsimony and this can
present diculties when using the model to adequately describe the data. An
ARMA model may have fewer parameters than AR model, similarly a GARCH
model may contain fewer parameters as compared to an ARCH model, and thus
a GARCH model may be preferred to an ARCH model. There are a variety of
extensions of the ARCH family of models that include the Exponential GARCH
(EGARCH), the Integrated GARCH (IGARCH) (Amos, 2009). These are not
23
discussed in this research study. For interested reader, thesis by Talke (2003)
is good source of such information. The GARCH(p, q) model employs the same
equation (3.1) for the log-returns r
t
but the equation for the volatility, includes
q new terms, that is
r
t
=
t
t
,
t
N(0, 1), (3.17)
2
t
=
0
+
1
r
2
t1
+. . . +
q
r
2
tq
+
1
2
t1
+. . . +
p
2
tp
. (3.18)
where now t >max(p, q) and the remaining components are as in the ARCH
model. The parameters of the model are
0
,
1
, . . . ,
q
,
1
, . . . ,
q
, for some pos-
itive integers p, q . We see that if p = 0 the above model is reduced to the
ARCH(q) . Thus the GARCH model generalizes the ARCH by introducing val-
ues of
2
t1
,
2
t2
, . . . in equation(3.17). Let {r
t
} be the mean corrected return,
t
be a Gaussian white noise with mean zero and unit variance. Let also H
t
be the
information set or history at time t given by H
t
=
_
r
1
, r
2
, . . . , r
t1
_
as in the
ARCH model. Then the process {r
t
} is GARCH(1, 1) if
r
t
=
t
t
,
t
N(0, 1), (3.19)
and
2
t
=
0
+
1
r
2
t1
+
1
2
t1
. (3.20)
The restrictions
0
0,
1
0, and
1
0 are imposed in order for the variance
2
t
to be positive. Clearly equations (3.18) and (3.19) show that large past mean
corrected squared returns r
2
t1
or past conditional variances
2
t1
give rise to large
values of
2
t
(Tsay, 2002). The conditional mean E(r
t
|H
t
) = 0 implies that
_
r
t
_
is a martingale dierence, thus E(r
t
) = 0 and {r
t
} is an uncorrelated series. At
this point it is important to point out that the information set is now strictly
given by
[r
1
;
2
1
, . . . , r
t1
;
2
t1
].
24
Furthermore ARCH(1) can also be dened as GARCH(0, 1) . Taking v
t
= r
2
t
2
t
,
we write
r
2
t
=
2
t
+v
t
,
=
0
+
1
r
2
t1
+
1
(r
2
t1
v
t1
) +v
t
,
=
0
+ (
1
+
1
)r
2
t1
+v
t
1
v
t1
. (3.21)
As in ARCH(1) model, E(v
t
|H
t
) = 0 and v
t
is another martingale dierence,
meaning that E(v
t
) = 0 and cov(v
t
, v
tk
) = 0 for k 1 and v
t
is series uncorre-
lated from equation (3.18). So
E(r
2
t
) =
2
t
=
0
+ (
1
+
1
)E(r
2
t1
),
2
t
=
0
+ (
1
+
1
)E(
2
t
2
t
),
=
0
+ (
1
+
1
)
2
t
E(
2
t
),
=
0
+ (
1
+
1
)
2
t
.1,
(1
1
B
1
B)
2
t
=
0
,
=

0
1 (
1
+
1
)
.
provided |
1
+
1
| < 1
3.3.1 Estimation of the GARCH(1, 1)
Estimation of the parameters of the GARCH(1, 1) model is performed in the same
approach as in the ARCH(1) model. However, since the conditional variance of
the GARCH(1, 1) model depends also on the past conditional variance, an initial
value of the past conditional variance
2
1
is needed. Suppose as before that we
have a sample of log-returns r
1
, . . . , r
n
and we wish to nd estimates
0
,
1
and
25
1
that maximize the log-likelihood function, the conditional on
0
,
1
,
1
and
r
t1
, the distribution of r
t
must be normal, that is
r
t
|
0
,
1
, . . . ,
q
, r
t1
N(0,
0
+
1
r
2
t1
+
1
2
t1
) = N(0,
2
t
) (3.22)
This is follows from equations (3.18) and (3.19). The likelihood function is
L(
0
,
1
,
1
; r
n
) = p(r
1
, . . . , r
n
|
0
,
1
,
1
),
=
n
t=1
p(r
t
|
0
,
1
,
1
, r
t1
),
=
n
t=1
1
2
t
exp
_
r
2
t
2
2
t
_
,
= (2)
n
2
n
t=1
1
t
exp
_
1
2
n
t=1
r
2
t
2
t
_
. (3.23)
whereas from (3.19),
2
t
=
0
+
1
r
2
t1
+
1
2
t1
. The log-likelihood function of
0
,
1
,
1
is
(
0
,
1
,
1
; r
n
) = ogL(
0
,
1
,
1
; r
n
)
=
n
2
og(2)
1
2
n
t=1
og
2
t

1
2
n
t=1
_
r
2
t
2
t
_
(3.24)
The partial derivatives of (
0
,
1
,
1
; r
n
) with respect to
0
,
1
and
1
respec-
tively are,
(
0
,
1
,
1
; r
n
)
0
=
1
2
n
t=1
1
2
t
+
1
2
n
t=1
r
2
t
4
t
, (3.25)
26
(
0
,
1
,
1
; r
n
)
1
=
1
2
n
t=1
r
2
t1
2
t
+
1
2
n
t=1
r
2
t
r
2
t1
4
t
, (3.26)
(
0
,
1
,
1
; r
n
)
1
=
1
2
n
t=1
2
t1
2
t
+
1
2
n
t=1
r
2
t
2
t1
4
t
, (3.27)
Recalling that
2
t
=
0
+
1
r
2
t1
+
1
2
t1
and equating equations (3.25) , (3.26)
and (3.27) to zero, we obtain the system of three (3) equations with three (3)
unknowns.
n
t=1
_
1

0
+
1
r
2
t1
+

2
t1
_
=
n
t=1
_
r
2
t
(
0
+
1
r
2
t1
+

2
t1
)
2
_
, (3.28)
n
t=1
_
r
2
t1

0
+
1
r
2
t1
+

2
t1
_
=
n
t=1
_
r
2
t
r
2
t1
(
0
+
1
r
2
t1
+

2
t1
)
2
_
, (3.29)
n
t=1
_

0
+
1
r
2
t2
+

2
t2

0
+
1
r
2
t1
+

2
t1
_
=
n
t=1
_
r
2
t
(
0
+
1
r
2
t2
+

2
t2
)
(
0
+
1
r
2
t1
+

2
t1
)
2
_
. (3.30)
where
2
t1
and
2
t2
can be expressed in terms of the log-returns only, given that
some initial values for
2
1
and
2
0
are given.The maximum likelihood estimator of
2
t
is
2
t
=
1
n
n
i=1
r
2
i
. Alternatively, the above sums can start from t = 2 or t = 3
. The above system can be solved for
0
,
1
,

1
using numerical methods and it
can be veried that the Hessian matrix evaluated at
0
=
0
,
1
=
1
,
1
=

1
dened by
H =
_
_
_
_
_
_
_
_
2
0
2
1
2
1
_
_
_
_
_
_
_
_
0
=
0
,
1
=
1
,
1
=

1
27
is a negative denite matrix and so
0
,
1
,

1
are the maximum likelihood esti-
mates of
0
,
1
,
1
. However, since the conditional variance of the GARCH(1, 1)
model depends also on the past conditional variance, an initial value of the past
conditional variance,
2
1
is needed. The unconditional variance of r
t
can be taken
as an initial value for this variance, that is,
2
1
can be taken to be

0
1
1
1
.
Sometimes the sample variance of the return series can be taken to be the initial
value.
3.3.2 The GARCH(p, q) Model
The GARCH(p, q) is a generalization of GARCH(1, 1) with p as the autoregressive
lag p and lag q is the moving average lag. Formally a process
_
r
t
_
is GARCH(p, q)
if
r
t
=
t
t
,
2
t
=
0
+
q
i=1
i
r
2
ti
+
p
j=1
2
tj
,
=
0
+(B)r
2
t
+(B)
2
t
. (3.31)
where
t
is a Gaussian white noise, while (B) and (B) are polynomials in the
backshift operator given by
(B) =
1
B +. . . +
q
B
q
,
and
(B) =
1
B +. . . +
p
B
p
,
with the restrictions
0
> 0,
i
0 and
j
0 for i = 1, 2, . . . , q and j =
1, 2, . . . , p, being imposed in order to have the conditional variance remaining
positive. Equation (3.31) can be expressed as
(1 (B))
2
t
=
0
+(B)r
2
t
.
28
Note that GARCH(0, q) is the same as an ARCH(q) model and that for p =
q = 0 we have a GARCH(0, 0) model, which is a simple white noise. Similar
to the ARCH(q) model, the conditional mean of {r
t
} is zero, that is E(r
t
|H
t
)
implies the series {r
t
} is a martingale dierence and observing {r
t
} is uncorrelated
Gourieroux, et al., (1997). Assuming the GARCH(p, q) process is second order
stationary, that is
V ar(r
t
) = E(r
2
t
),
= E(
2
t
2
t
),
= E(
2
t
E(
2
t
|r
t1
)),
= E(
0
+(B)r
2
t
+(B)
2
t
),
=
0
+(B)E(r
2
t
) +(B)E(
2
t
),
=
0
+(B)
2
t
+(B)
2
t
,
=

0
1
q
i=1
p
j=1
j
. (3.32)
The autocovariance of a GARCH(p, q) model for k 1 where k is the lag,
E(r
t
r
tk
) = 0,
since r
t
is a martingale dierence Gourieroux, et al., (1997). Thus the GARCH(p, q)
model does not show autocorrelation in the return series {r
t
} . However the
squared returns show autocorrelation even though the returns are not correlated.
Considering writing r
2
t
in terms of v
t
= r
2
t

2
t
, yields
r
2
t
=
2
t
+v
t
,
=
0
+
q
i=1
i
r
2
ti
+
p
j=1
j
(r
2
tj
v
tj
) +v
t
=
0
+
q
i=1
i
r
2
ti
+
p
j=1
j
r
2
tj
j=1
j
v
tj
+v
t
. (3.33)
29
Now let m =max(p, q) then
r
2
t
=
0
+
q
i=1
(
i
+
i
)r
2
tj
j=1
j
v
tj
+v
t
, (3.34)
where
i
= 0 for i > q and
j
= 0 for i > p . Thus the equation of r
2
t
has
ARMA(m, p) representation. In order to nd the GARCH(p, q) process, we con-
sider solving for
0
in equation (3.34) and let the variance of r
t
be
2
which
yields
0
=
2
(1
q
i=1
j=1
j
), (3.35)
and substituting the value of
0
of equation (3.35) into equation (3.34) gives
r
2
t
=
2
(1
m
i,j=1
(
i
+
j
)) + (
m
i,j=1
(
i
+
j
))r
2
tj
j=1
j
v
tj
+v
t
,
=
2
+
m
i,j=1
(
i
+
j
)(r
2
tj
)
p
j=1
j
v
tj
+v
t
. (3.36)
Therefore,
r
2
tk
=
m
i,j=1
(
i
+
j
)(r
2
tj
)
p
j=1
j
v
tj
+v
t
. (3.37)
Multiplying both sides by (r
2
tk
), yields
(r
2
tk
)(r
2
t

2
) =
m
i,j=1
(
i
+
j
)(r
2
ti
)(r
2
tk
p
j=1
j
v
tj
((r
2
tk
) +v
t
(r
2
tk
),
(3.38)
and taking expectations
E[(r
2
tk
)(r
2
t

2
)] = E[
m
i,j=1
(
i
+
j
)(r
2
ti
)(r
2
tk
)]
E[
p
j=1
j
v
tj
((r
2
tk
)] +E[v
t
(r
2
tk
)].
(3.39)
30
But
E[v
t
(r
2
tk
)] = E[(r
2
tk
)E(v
t
|r
t
)] = 0,
since v
t
is a martingale dierence and also
E[
j
v
tj
((r
2
tk
)] = E[(r
2
tk
)E(v
tj
|r
tk
)] = 0. (3.40)
for k < j. . Thus the autocovariance of the squared returns for the GARCH(p, q)
model is given by
cov(r
2
t
, r
2
tk
) = E[
m
i,j=1
(
i
+
j
)(r
2
ti
)(r
2
tk
)],
=
m
i,j=1
(
i
+
j
)cov(r
2
t
, r
2
tk+i
), (3.41)
Dividing both sides by var(r
2
t
) gives the autocorrelation function at lagk as
k
=
m
i=1
(
i
+
i
)
ki
. (3.42)
for k (p +1). This result is analogous to the Yule-Walker equations for an AR
process. Thus the autocorrelation function (ACF) and the partial ACF (PACF)
of the squared returns in a GARCH process has the same pattern as those of an
ARMA(m, p) process. As in ARMA modelling the ACF and PACF are useful in
identifying the orders p and q of the GARCH(p, q) process. The ACF are also
important for checking model accuracy, in which case, the ACFs of residuals
should be indicative of a white noise if the model is adequate. Thus the rst
autocorrelations depend on the parameters
0
,
1
, . . . ,
q
;
1
,
2
, . . . ,
q
, but given
p
, . . . ,
p+1m
, , the expression in (3.39) determines uniquely the autocorrelations
at higher lags, (Bollerslev, 1986 ). Thus letting
mm
denotes the m
th
partial
autocorrelation for r
2
t
, then
k
=
m
i=1
mm
ki
, k = 1, . . . , m.
31
By equation (3.39)
mm
cuts o after lag q for an ARCH(q) process such that
mm
= 0 for k q and
mm
= 0 for k > q . This is identical to the PACF for an
AR(q) process and decays exponentially (Bollerslev, 1986). After identifying the
orders p and q , we can now estimate the parameters (
0
,
1
, . . . ,
q
,
1
, . . . ,
q
),
of the GARCH(p, q) model as explained in section 3.3.3 .
3.3.3 Estimation of GARCH(p, q) Model
The maximum likelihood estimation is also applicable in estimating the parame-
ters of the model. As in the GARCH(1, 1) model estimation, initial values of both
the squared returns and past conditional variances are needed in estimating the
parameters of the model. As suggested by Bollerslev, (1986) and Tsay, (2002),
the unconditional variance given in equation (3.31) or the past sample variance
of the returns for the past variances may be used as initial values. Therefore as-
suming r
1
, . . . , r
q
and
2
1
, . . . ,
2
p
are known, the conditional maximum likelihood
estimates can be obtained by maximizing the conditional log-likelihood given by
= ogf(r
q+1
, . . . , r
t
,
2
p+1
, . . . ,
2
t
|, r
1
, . . . , r
q
,
2
1
, . . . ,
2
p
),
=
1
2
T
t=m+1
og(2
2
t
)
1
2
T
t=m+1
_
r
2
t
2
t
_
.
with, = (
0
,
1
, . . . ,
p
;
1
,
2
, . . . ,
q
) and m =max(p, q).
3.3.4 Model Checking
Goodness of t of the ARCH-GARCH model is based on residuals and is more
specically on the standardized residuals (Takle, 2003). The residuals are as-
sumed to be independently and identically distributed following either a normal
32
or standardized tdistribution (Tsay, 2002) and (Gourieroux, 2001). Plots of the
residuals such as the histogram, the normal probability plot and the time plot
of the residuals can be used. If the model ts the data well the histogram of
the residuals should be approximately symmetric. The normal probability plot
should be a straight line while the time plot should exhibit random variation.
The ACF and the PACF of the standardized residuals are used for checking the
adequacy of the conditional variance model. The Engles test and the Ljung Box
Qtest are used to check the validity of the ARCH eects in the data. Having
established that our model ts the data well, we can now use the tted model to
compute forecasts.
3.3.5 Forecasting with GARCH(p, q) models
As we have outlined, the GARCH and the ARMA models are similar, such that
forecasting using the GARCH model is the same as using the ARMA model.
Thus the conditional variance of {r
t
} is obtained simply by taking the conditional
expectation of the squared mean corrected returns. Assuming a forecasting origin
of T, then the step ahead volatility forecast is given by
r
2
t
(1) = E[r
2
t+1
|r
t
],
=
0
+
m
i=1
(
i
+
i
)E(r
2
t+i
|r
t
)
p
i=1
i
(v
t+i
|r
t
).
where r
2
t
, . . . , r
2
t+m
,
2
t
, . . . ,
2
t+1p
are assumed known at time t and the true
parameter values
i
and
i
for i = 1, . . . , m are replaced by their estimates. Fur-
thermore, the step ahead forecast of the conditional variance in a GARCH(p, q)
model is given by
2
t
() = E[r
2
t+1
|r
t
],
33
=
0
+
m
i=1
(
i
+
i
)E(r
2
t+i
|r
t
)
p
i=1
i
(v
t+i
|r
t
).
where E[r
2
t+1
|r
t
] for i < can be given recursively as for i , E(v
t+i
|r
t
) = 0
for i < , E(v
t+i
|r
t
) = v
t+i
for i We now consider the techniques that are
used for selecting the best tting model in line of two or more competing models
based on the likelihood ratios.
3.4 Model selection criteria
In statistical modelling, one of the main challenges is to select a suitable model
from a candidate family to characterize the underlying data. Model selection
criteria provide useful tools in this regard. Selection criteria assess whether a
tted model oers an optimal balance between goodness-of-t and parsimony.
Ideally, a criteria will identify candidate models that are either too simplistic
to accommodate the data or unnecessarily complex. The most common model
selection criteria are the AIC and the BIC.
3.4.1 The Akaike Information Criterion
The Akaike information criterion (AIC) was the rst model selection criterion
to gain widespread acceptance. AIC was introduced in 1973 by Hirotogu Akaike
as an extension to the maximum likelihood principle. Conventionally, maximum
likelihood is applied to estimate the parameters of a model once the structure
of the model has been specied. Akaikes seminal idea was to combine estima-
tion and structural determination into a single procedure. The minimum AIC
34
procedure is employed as follows. Given a family of candidate models of various
structures, each model is t to the data through maximum likelihood. An AIC is
computed based on each model t. The tted candidate model corresponding to
the minimum value of AIC is then selected. AIC serves as an estimator of Kull-
backs directed divergence between the generating, or true, model (that is, the
model that presumably gave rise to the data) and a tted candidate model. The
directed divergence assesses the disparity or separation between two statistical
models. Thus, when entertaining a family of tted candidate models, by select-
ing the tted model corresponding to the minimum value of AIC, one is hoping to
identify the tted model that is closest to the generating model (Salkind, 2007).
The AIC is dened by
AIC = 2(oglikelihood) + 2(N).
where N denotes the number of observations
3.4.2 The Bayesian information criterion
The Bayesian information criterion (BIC) is related to the Bayes factor and
is useful for model comparison in its own right. The BIC of a model is dened as
BIC = 2n(likelihood) + (k +knN).
where k denotes the number of parameters and N denotes the number of ob-
servations or equivalently, the sample size. BIC penalizes more complex models
(those with many parameters) relative to simpler models. This denition permits
multiple models to be compared at once; the model with the highest posterior
probability is the one that minimizes BIC.
A desirable model is one that minimizes the AIC or the BIC. The other criteria are
the R
2
associated with the model which is the proportion of variability in a data
35
set that is accounted for by the statistical model (Salkind, 2007 ). However, as
Harvey (1991) indicated that the coecient of determination (R
2
) has a limitation
in that, a model which can pick out the trend reasonably well, will have R
2
close
to unit. In general a model selected by two dierent criteria mentioned above may
dier and thus it should be emphasized that the selection of an ARCH-GARCH
model depends on the selection criteria used (Talke, 2003 ).
3.5 Forecasting performance
As pointed out in section 3.2.3, forecasting is one of the most important objectives
of time series modelling, thus one of the criteria for selecting the best model can
be based or centred on the best forecasting model. Forecast in this context
means the estimates of the conditional variances obtained from models. There
are several measures for assessing the predictive accuracy of an ARCH-GARCH
model. Among these is the mean square error (MSE). The MSE is dened as
the average of the squared dierence between the actual variance and volatility
forecast
2
t
. In the absence of the observed true variance the squared time series
observations r
2
t
is used. The MSE is thus given by
MSE =
1
T
T
t=1
_
r
2
t

2
t
_
2
, (3.43)
where
2
t
, for t = 1, . . . , T is the estimated conditional variance obtained from
tting ARCH-GARCH model. The MSE is critized, as in Tsay (2002), in that,
although the r
2
t
is a consistent estimator of
2
t
, it is nevertheless noisy, therefore
unstable. Among the alternative measures are the mean absolute error (MAE)
by Lopez (1999) dened by
MAE =
1
T
T
t=1
r
2
t

2
t
2
, (3.44)
36
and the MSE of the log of the squared error dened by
MSLE =
1
T
T
t=1
_
n(r
2
t
) n(
2
t
)
_
2
. (3.45)
The advantage of the MSE of the log of the squared error is that it penalizes
inaccurate variance forecasts more heavily when the squared innovations, r
2
t
are
low. Another criterion which is popular is Theils Ustatistic. The Theils U
test which is used to test accuracy of the future forecasts/predictions is dened
as (Diebold and Lopez, 1995)
U =
_
T1
t=1
(FPE
t+1
APE
t+1
)
2
T1
t=1
(APE
t+1
)
2
_1
2
. (3.46)
where
FPE
t+1
=
(

X
t+1
X
t
)
X
t
is forecasted relative change, and
APE
t+1
=
_
X
t+1
X
t
X
t
_
is actual relative change.
If the forecasts are good then U should be close to zero. A Ustatistic of one
implies that the model under consideration and the benchmark model are equally
(in)accurate, while a value of less than one implies that the model is superior to
the benchmark, and vice versa for U > 1.
CHAPTER FOUR
DATA ANALYSIS AND METHODOLOGY
4.1 Introduction
Having explored the general theory of ARCH-GARCH models in the preceding
chapter, this chapter is dedicated to tting the GARCH family of models to
the Tanzania ination rate data which we obtained courtesy of the Tanzania
National Bureau of Statistics. A description of the data is given in section 4.2
and application of GARCH in real life data is given in section 4.3 through 4.7. In
section 4.2 we investigate the general statistical features of the given time series:
Ination rate data based on Tanzania.
4.2 Data description and basic statistics
The data employed in this study comprise of 168 monthly observations of the
Tanzania ination rates spanning from January 1, 1997 to December 31, 2010. In
Tanzania the Consumer Price Index (CPI) is calculated by the National Bureau
of Statistics under the auspices of the Ministry of Finance. The national CPI
covers prices collected from twenty towns in Tanzania mainland. All prices are
the prevailng market prices and are gathered for two hundred and seven items.
Table 4.1 gives the descriptive statistics for monthly consumer prices and returns,
where Std. Dev. stands for Standard Deviation. The ination series which is
the basic data in this work, has a constant general mean of 8.1, the standard
37
38
deviation of 3.7, positive skewness of 0.8138 implying that the distribution has a
long right tail and kurtosis of 2.8019. The minimum value in this data is 0.2 and
the maximum value is 17.2 , giving a data range of 17.0. On the other hand, the
return series has a negative skewness of 0.3811 implying that the distribution
has a long left tail. Skewness is a measure of symmetry. For univariate data sets
X
1
, X
2
, . . . , X
N
, the formula for Skewness is: Skewness =
N
i=1
(X
i

X)
3
(N1)s
3
, where

X
is the mean, s is the standard deviation, and N is the number of data points.
Table 4.1: Descriptive Statistics for the Ination rate series.
Price Return
Mean 8.0500 1.9773
Median 6.9000 1.9314
Std. Dev. 3.6475 0.5023
Kurtosis 2.8019 6.9162
Skewness 0.8138 -0.3811
Maximum 17.2 2.6391
Minimum 0.2 -1.6094
Jarque-Bera 55.8432 53.8218
ADF(Level) -0.576881
ADF(1st di) -11.642264
The value for kurtosis is high (close to three), so, the distribution are peaked
relative to normal. Kurtosis is a measure of whether the data are peaked or
at relative to a normal distribution. That is, data sets with high kurtosis tend
to have a distinct peak near the mean, decline rather rapidly, and have heavy
39
tails. Data sets with low kurtosis tend to have a at top near the mean rather
a sharp peak. For univariate data X
1
, X
2
, . . . , X
N
, the formula for Kurtosis is:
Kurtosis =
N
i=1
(X
i

X)
4
(N1)s
4
where

X is the mean, s is the standard deviation and
N is the number of data points. The Kurtosis for standard normal distribution
is three. For this reason, ExcessKurtosis =
N
i=1
(X
i

X)
4
(N1)s
4
3. The Jarque-Bera
test rejects the null hypothesis at 5% level that the distribution is normal. So,
the sample has all nancial characteristics, that is, volatility clustering and lep-
tokurtosis. Furthermore, in terms of stationary the results from the Augmented
Dickey Fuller(ADF) test in Table 4.1 indicates that the series has unit root and
therefore time series models can be used to examine the behavioour of volatility
over time.
4.3 Transformations
As stated earlier most applied time series are non stationary, their variance is
changing with time. Consider, for example, a plot of the CPI series for the
Figure 4.1: Time plot of monthly ination in Tanzania
40
period January 1997 through December 2010 in Figure 4.1 , it is evident that
variances are changing with time. The ination series we have in this study is
non-stationary. Therefore we convert the prices to returns by logarithmic trans-
formations. The logarithmic returns is based upon the following mathematical
denition, that is,
r
t
= n
P
t
P
t1
where
r
t
is the return for any time, t,
P
t
is the Consumer Price Index value at time , t,
P
t1
is the Consumer Price Index value at time , t 1.
Figure 4.2: First Dierence of Log of CPI
Figure 4.2 illustrates the plot of return series converted from original data. The
returns appear to be quite stable overtime and the transformation has produced
a stationary process. This behaviour of ination series returns is in line with
most nancial theories and models which usually assume the prices to returns,
to be stationary process. This is also noticed in Figure 4.9(b) where the normal
41
probability plot shows a nearly straight line suggesting that the residuals follow
an approximately normal distribution.
4.4 Testing for ARCH eects and Serial corre-
lation in the return series
This section uses a formal statistical test to establish the presence of ARCH eects
in the data. The test for heteroscedasticity has been done using Ljung-Box-Pierce
Qtest and Engles ARCH test. Table 4.2 shows the Ljung-Box-Pierce Qtest
in the form of binary number that is 0 and 1. H = 0 implies that no signicant
correlation exists. H = 1 means that signicant correlation exists. From Table
4.2 it can be veried that there is no signicant correlation in the raw returns
when tested for up to 10, 15 and 20 lags of the ACF at the 5% level of signicance.
However, there is signicant serial correlation in the squared returns in Table 4.3
and Table 4.4, all the p values show that the Q test and the ARCH test at
lag 10, lag 15 and lag 20 are signicant showing presence of ARCH eects.
Table 4.2: Ljung-Box-Pierce Q-test for autocorrelation: (at 95% condence).
Lag H p value Stat CriticalV alue
10 0.0000 0.0721 17.1019 18.3070
15 0.0000 0.2359 18.529 24.9958
20 0.0000 0.3888 21.1416 31.4104
The dependence in the data x
1
, . . . , x
n
are ascertained by computing correlations
for data values at varying time lags. This is done by plotting the sample auto-
42
Table 4.3: Ljung-Box Q-test for squared returns: (at 95% condence).
10 1.0000 0.0000 57.3782 18.3070
15 1.0000 0.0000 76.7057 24.9958
20 1.0000 0.0000 82.7525 31.4104
Table 4.4: Engle ARCH test for heteroscedasticity: (at 95% condence).
10 1.0000 0.0000 68.6467 18.3070
15 1.0000 0.0000 67.9727 24.9958
20 1.0000 0.0000 66.2913 31.4104
correlation function (ACF):
ACF(h) = (h) =
(h)
(0)
,
against the time lags h = 0, 1, . . . , n1 and where (h) is the sample autocovari-
ance function (ACVF) given by:
ACV F(h) = (h) =
1
n
nh
t=1
(x
t+h
x)(x
t
x),
and x is the sample mean. If the time series is an outcome of a completely
random phenomenon, the autocorrelation should be near zero for all time-lag
separations. Otherwise, one or more of the autocorrelations will be signicantly
non-zero.
Another useful method to examine serial dependencies is to examine the partial
43
Figure 4.3: ACF with Bounds for Raw Return Series
Figure 4.4: PACF with Bounds for Raw Return Series
autocorrelation function (PACF) ( an extension of autocorrelation),where the
dependence on the intermediate elements (those within the lag) is removed. The
partial autocorrelation is similar to autocorrelation, except that when calculating
it, the autocorrelation with all the elements within the lag are eliminated.
In Figure 4.3 and Figure 4.4 the ACF and PACF provide no indication of the
correlation characteristics of the returns. The ACF of squared returns in Figure
4.5 show signicant correlation and die out slowly. This results indicates that
the variance of returns series is conditional on its past history and may change
44
Figure 4.5: ACF of the squared Returns
over time. Each of these tests extracts the sample mean from the actual ination
series. This is consistent with the denition of the conditional mean equation of
the default model, in which the innovations process is
t
= r
t
C and C is the
mean of r
t
.
4.5 Model estimation and evaluation
4.5.1 Model selection and analysis
The strategy used in selecting the appropriate model from competing models
is based on the Akaike information criterion (AIC), the Bayesian information
criterion (BIC), standard error (SE) and on the signicance tests. MATLAB
software is used to perform trial and error to determine the best tting model. Ta-
ble 4.5 gives the suggested models with their respective t statistics. The idea is
to have a parsimonious model that captures as much variation in the data as pos-
sible. Usually the simple GARCH model captures most of the variability in most
stabilized series. Small lags for p and q are common in applications. Typically
45
GARCH(1, 1), GARCH(2, 1) or GARCH(1, 2) models are adequate for modelling
volatilities even over long sample periods (Bollerslev, Chou and Kroner, 1992).
However in the Table 4.5 below we have included GARCH(1, 0), GARCH(0, 2) and
GARCH(2, 2) in order to check if they are appropriate for modelling time varying
variances of our data. In the Table 4.5 below the smaller the AIC and BIC the
better. Larger AIC
s, BIC
s and standard error makes the model unfavourable.

The model given in bold is judged to be the most appropriate according to the
criteria above.
Table 4.5: Comparison of suggested GARCH models.
Model AIC BIC SE Log Likelihood
GARCH(0, 1) 489.0200 498.3560 0.1026 241.5100
GARCH(1, 1) 474.8236 487.2715 0.0544 233.4118
GARCH(0, 2) 478.5589 491.0068 0.0860 235.2794
GARCH(1, 2) 475.5820 491.1419 0.0863 232.7910
GARCH(2, 1) 476.8236 492.3835 0.0735 233.4118
GARCH(2, 2) 477.1047 495.7766 0.0654 232.5524
The Table 4.5 shows the competing models to the data with their respective
AIC, BIC and SE. From our derived models, using the method of maximum
likelihood, the estimated parameters of the models with their corresponding stan-
dard error and other statistical tests are shown in the Table 4.6, 4.7 and 4.8. We
started in Table 4.6 by tting a GARCH(1, 1) to the data.
46
Table 4.6: Parameter estimates for GARCH(1,1).
Parameter C K GARCH(1) ARCH(1)
Estimates 0.0272 0.0753 0.5427 0.4573
Standard Error 0.0283 0.0254 0.0681 0.0958
t value 0.9627 2.9656 6.7136 5.6619
Pr(> |t|) 0.8315 < 0.0001 < 0.0001 < 0.0001
r
t
= 0.0272 +
t
, (4.1)
2
t
= 0.0753 + 0.45730r
2
t1
+ 0.54266
2
t1
. (4.2)
The standard errors are used to assess the accuracy of the estimates, the smaller
the better. The model t statistics used to assess how well the model t the
data are the AIC and BIC. The corresponding values are: AIC = 474.8 and
BIC = 487.3 with the log likelihood value of 233.4. The standard errors are
quiet small suggesting precise estimates. Based on 95% condence level, the co-
ecients of the GARCH(1, 1) model are signicantly dierent from zero and the
estimated values satisfy the stability condition, that is,
1
+
1
< 1, with
1
>
1
.
The second model to be considered is the GARCH(1, 2). This model is chosen
because it seemed to t the data well, as it has small log likelihood function,
small standard error and both their AIC and BIC are very close to that of
GARCH(1, 1). However the model was found to be violating
i
+
j
< 1, i, j Z
+
as
1
= 0.34705,
2
= 0.01764,
1
= 0.47655, that is
1
+
1
+
2
= 1, where
i
+
j
< 1 is a necessary condition for the model to be stationary. The corre-
sponding t statistics are: AIC = 475.6 and BIC = 491.1 with log likelihood
47
Parameter C K GARCH(1) ARCH(1) ARCH(2)
Estimates 0.0198 0.0892 0.4766 0.3471 0.1764
Standard Error 0.0422 0.0290 0.0915 0.1219 0.1468
t value 0.4706 3.0819 3.7920 3.9109 1.2013
function of 232.8. Thus the GARCH(1, 2) has bigger AIC and BIC values than
the stochastic volatility GARCH(1, 1) model, indicating that, GARCH(1, 1) is su-
perior in extracting statistical information from the data hence accommodating
conditional variances. The parameter estimates given in Table 4.7 giving rise to
the model,
r
t
= 0.0198 +
t
, (4.3)
2
t
= 0.0892 + 0.3471r
2
t1
+ 0.04766
2
t1
+ 0.1764
2
t2
. (4.4)
A corresponding stochastic GARCH(2, 1) model was also seen not to t the data
well, as the
2
coecient of the GARCH(2, 1) was not signicant, also , the sta-
tionarity condition was violated.
Another candidate model tted to data and tested was the GARCH(2, 2) . The
analysis showed a non signicant
1
coecient, this imply that the ARCH(1) pa-
rameter adds little explanatory power to the model. The t statistics are: AIC =
477.1 and BIC = 495.8 which are also greater than the GARCH(1, 1) model, sug-
gesting that the GARCH(2, 2) is less appropriate than the GARCH(1, 1) model.
Note also that the mean standard error under GARCH(2, 2) is bigger as compared
to the stochastic volatility GARCH(1, 1) model.
48
Parameter Estimates Standard Error t value
C 0.0205 0.0418 0.4864
K 0.1055 0.0527 3.2315
GARCH(1) 0.4543 1.5e-17 0.0015
GARCH(2) 0.3661 0.0572 3.1373
ARCH(1) 2.204e-20 0.1281 0.3.5456
ARCH(2) 0.1796 0.1124 3.2573
At this point in time series, it can be established that, amongst all the identied
models, the GARCH(1, 1) is proven to be the best t model. Section 4.5.2 below
provides a comprehensive comparison of the identied models.
4.5.2 A Comparison of the tted GARCH models
Table 4.5 shows the competing GARCH models together with their corresponding
estimates, standard errors, signicant statistical tests, AIC and BIC for each
parameter. The GARCH(1, 1) has the smallest AIC = 474.8 and BIC = 487.3
and all parameters are signicant. The standard error of the GARCH(1, 1) is
the smallest of the all models, suggesting that the model parameters are precise.
Therefore this model is selected to be the best t of all the models tted. The
other models have greater AIC, BIC and standard errors, therefore they were
provided only for comparison purposes. Having constructed our model, the next
task is to assess how well the model ts the data. This is done through residual
49
analysis or model checking as explained in section 4.6
4.6 Diagnostic checking of the GARCH(1, 1) model
As was explained in section 3.3.4 , model checking is done through analyzing
the residuals from the tted model. In time series modelling, the selection of
the best model t to the data is directly related to whether residual analysis is
performed well. One of the assumptions of GARCH model is that, for a good
model, the residuals must follow a white noise process, that is if the model ts the
data well, the residuals are expected to be random, independent and identically
distributed following the normal distribution. Figure 4.6 inspects the relationship
between the innovations (residuals) derived from the tted model, the correspond-
ing conditional standard deviations and the observed returns. We can see that
both innovations (top plot) and returns (bottom plot) in the Figure 4.6 exhibit
volatility clustering.
Figure 4.6: Plot of Return, Estimated Volatility and Innovations (residuals)
However if we plot the, standardized innovations (the innovations divided by their
50
conditional standard deviation), they appear generally stable with little clustering
as in Figure 4.7.The time plot of the residuals given in Figure 4.7 is used to check
whether the residuals are random. The normality check is also done by analyzing
the histogram of residuals and normal probability plot. Figure 4.9(a) gives the
histogram of the residuals from the GARCH(1, 1) model. The histogram shows
almost a symmetric bell-shaped distribution which is indicative of the residuals
following a normal distribution. The slight negative skewness is expected since
the residuals may also come from students t distribution. The negative skewness
tendency is also supported by negative large residuals in Figure 4.7. The normality
check of residuals can also be done by the normal probability plot. As we saw in
section 3.3.4 , if the residuals come from the normal distribution the plot should
be a straight line. Figure 4.9(b) gives the probability plot of the residuals. The
residuals are indeed normally distributed the graph Fig 4.9(b) shows a nearly
straight line were almost all the points has fallen along the dashed line.
Figure 4.7: Time plot of residuals from GARCH(1,1)
Moreover if the model is successful at modelling the serial correlation structure
in the conditional mean and conditional variance, then there should be no auto-
correlation left in the standardized residuals and squared standardized residuals.
51
Figure 4.8: ACF of the squared standardized residuals
Figure 4.8 gives the plot of the ACF of the squared standardized innovations;
the plot shows no correlation left. Table 4.9 and 4.10 compared the results of the
Qtest and the ARCH test with the results of the same tests in the pre-estimation
analysis in Table 4.2 and 4.3 respectively. In the pre-estimation analysis, both the
Qtest and the ARCH test indicated rejection (H = 0 with p value = 0) of their
respective null hypothesis showing signicant evidence in support of GARCH ef-
fects. In the post estimation analysis using standardized innovations based on
the estimated model, these same tests indicate acceptance (H = 0 with highly
signicant pvalues) of their respective null hypothesis and conrm the explana-
tory power of GARCH(1, 1). The tests showed that no any ARCH eects left (no
heteroscedasticity). Also from t test result Table 4.9 , since the p values of
0.9964, 0.9997, 1.0000 are greater than 5% alpha level, we fail to reject the null
hypothesis that, there is no autocorrelation left in the residuals. Therefore we
proceed to use the model to forecast future values of the ination series.
52
Table 4.9: Ljung-Box-Pierce Q-test on standardized innovations.
10 0.0000 0.99412 2.2420 18.3070
15 0.0000 0.9992 3.3378 24.9958
20 0.0000 1.0000 4.0144 31.4104
Table 4.10: Engle ARCH test on standardized innovations.
10 0.0000 0.9964 1.9937 18.3070
15 0.0000 0.9997 2.8363 24.9958
20 0.0000 1.0000 3.5752 31.4104
4.7 Forecasting with the GARCH(1, 1) model
One of the main objectives of this study and time series analysis in general is
to use the constructed model to compute forecasts. In time series, forecasting
is a mathematical way of estimating future values using present and historical
values of the series (Aidoo, 2010). Forecasting as described by Box and Jenkins
(1976) provide basis for economic and business planning, inventory and produc-
tion control and optimization of industrial process. From previous studies, most
research work has found that the selected model is not necessary the model that
provides best forecasting. In this sense further forecasting accuracy test such as
MAE, RMSE and MSE must be performed.
53
4.7.1 Forecast Evaluation and Accuracy Criteria
Empirically taking, we have examined that Table 4.12 reports the various mea-
sures of forecasting errors, namely the mean absolute error (MAE), the root
mean squared error (RMSE), and Thieles U for two models seemed to be ade-
quately t the data. The rst two forecast error statistics depend on the scale of
the dependent variable. These are used as relative measure to compare forecasts
for the same series across dierent models, the smaller the error the better the
forecasting ability of that model accordingly. The remaining two statistics are
scale invariant. The Theil inequality coecient always lies between zero and one,
where zero indicates a perfect t. From the results, it can be seen that, most
Table 4.11: Forecast Accuracy Test on the most likely suggested GARCH models.
Model MSE MAE RMSE Thiels U test
GARCH(1, 1) 0.5848 0.7483 0.8651 0.8528
GARCH(1, 2) 0.6034 0.9812 0.7768 0.9821
of the accuracy test favour GARCH(1, 1) model. Also the Thieles statistics is
less than one (0.8528 as from Table 4.11 ) which indicates that, the forecasts
are fairly accurate. To measure the forecasting ability we have estimated within
sample forecasts. The purpose of forecasting within the sample is to test for the
predictability power of the model. If the magnitude of the dierence between the
forecasted and actual values is small then the model has good forecasting power.
In our case GARCH(1, 1) has shown good results as evident from Table 4.12 .
One can observe from the Figure 4.9 that the forecasted series are closer to the
accuracy series. Therefore it can be concluded that the prediction power of the
model is better and suitable for twelve periods forecasting.
54
Table 4.12: One year in-sample forecasts of Ination obtained from the
GARCH(1,1).
Month Forecast(%) Observed Value(%) Forecast Error
January 11.32 10.9 0.42
February 10.43 9.6 0.83
March 9.66 9.0 0.66
April 10.08 9.4 0.68
May 7.90 7.2 0.70
June 8.74 7.9 0.84
July 7.29 6.3 0.99
August 7.48 6.6 0.88
September 5.03 4.5 0.53
October 4.81 3.9 0.91
November 5.14 4.3 0.84
December 6.30 5.6 0.70
The Figure 4.10 displays the original ination rate and the predicted values pro-
duced by the GARCH(1, 1) model. The gure also displays how the forecasted
values behave. It can be conrmed that the model somehow t the data well.
55
Figure 4.9: (a)Histogram of residuals(top) and (b)Normal probability(bottom)
plot from GARCH(1,1)
56
Figure 4.10: Time plot of Ination rate and one year forecasts from GARCH(1,1)
CHAPTER FIVE
DISCUSSION, CONCLUSION AND FUTURE WORK
5.1 Discussion
The GARCH(1, 1) model has been tted. From this model we get
0
= 0.0753,
1
=
0.45730 and
1
= 0.54266 with small standard error 0.05441 . Ljung Box-Pierce
Qtest is applied to test the uncorrelatedness of the data, Engles test is applied
to test for the presence of heteroscedasticity. A forecasting plot has been drawn,
and both AIC = 474.8236 and BIC = 487.2715 has been obtained. According to
the GARCH(1, 1) model tted
1
reects the inuence of random deviations in
the previous period on
t
, where as
1
measures the part of the realized variance
in the previous period that is carried over into the current period. The sizes
of
1
and
1
determine the short-run dynamics of the resulting volatility time
series. Large GARCH coecient,
1
mean that volatility reacts intensely to the
ination rate movements. Large GARCH coecient
1
, indicates that, shocks
to the conditional variance take long time to die out. Our results also show that
there exist some persistent in volatility as measured by the sum of
1
and
1
in
the GARCH(1, 1) model.
5.2 Conclusion
The study has presented us with an opportunity to have an extensive under-
standing of the theory of time series analysis in the area of non linear models
57
58
and its application to real life situation. The stages in the model building (that
is the identication, estimation and checking) strategy has been explored and
utilized. Based on minimum AIC and BIC values, the best t GARCH mod-
els tend to be GARCH(1, 1) and GARCH(1, 2) . The GARCH(1, 1) model has
smaller AIC and BIC which is an indicative that it explains the data better than
GARCH(1, 2) model. After estimation of the parameters of selected models, a
series of diagnostic and forecast accuracy test were performed. Having satised
all the model assumptions, GARCH(1, 1) model was judged to be the best model
for forecasting.
5.3 Suggestions for future work
Although the application in this research study is based on ination data, future
research can also consider;
the multivariate GARCH, where other macro-economic variables such as
exchange rate and money supply will be involved in the model
other areas of application for instance, environmental and pollution data,
health researches in the context of longitudinal data, agriculture and geo-
statistics just to mention a few.
REFERENCES
Aidoo, E (2010), Modelling and Forecasting Ination Rates in Ghana. An Appli-
cation of SARIMA models, A dissertation submitted to the School of Technology
and Business Studies, Hogskolan Dalarna in partial fullment of the requirement
for the award of Masters of Science Degree in Applied Statistics.
Ahiati, V.S (2007), Discrete Time Series Analysis with ARMA Models, Revised
Edition, Holden-Day, Oakland.
Akaike, H (1974), A new look at the statistical model identication. IEEE trans-
actions on automatic control volume 19 number 6 page 719-723.
Amos, C (2010), Time Series Modelling with Application to South African In-
ation Data, A thesis submitted to University of Kwazulu Natal for the degree
of Masters of Science in Statistics.
Assis, K., Amran, A., Remali, Y (2006), Forecasting Cocoa Bean Prices Using
Univariate Time Series Models; Journal of Arts Science and Commerce ISSN
2229 - 4686.
Bollerslev, T (1986), Generalized autoregressive conditional heteroscedasticity.
Journal of Econometrics 31, 307-327.
Bollerslev, T., R.Y. Chou and K.F. Kroner (1992), ARCH Modeling in Finance:
A Selective Review of the Theory and Empirical Evidence, Journal of Economet-
rics, 52, 5-59.
59
60
Brockwell, P. J. and Davis, R. A (2002), Introduction to Time Series and Fore-
casting. 2nd ed. Springer Texts in Statistitics, Springer, New York.
Brooks, C (2008), Introductory Econometrics for Finance, 2nd ed. Cambridge
University Press, New York.
Campbell, J.Y, Lo, A.W. and Mackinlay, A.C (1997), Modelling daily value at
risk using realized volatility and ARCH type models, Journal of empirical nance
volume 11, no. 3 pages 379-398.
Chateld, C (2000), Time Series Forecasting: University of Bath, Chapman and
Hall text in statistical science, London.
Chong, C.W., Loo, C.S and Muhammad, A.M (2002), Modelling the Volatility of
Currency Exchange Rates Using GARCH Model, Malaysia Press, Putra. David,
F.H (2001), Modelling UK Ination, 1875-1991. Journal of Applied Economet-
rics, 16(3): 255-275.
Diabold, F.X., Lopez, J.A (1995), Comparing Predictive Accuracy. Journal of
Business and Economic Statistics 13 253-265, 2nd Edition. South-Western Pub-
lishing.
Engle, R (1982), The Use of ARCH/GARCH Models in Applied Econometrics;
Journal of Economic Perspectives-Volume 15, Number 4. pp. 157 - 168.
Franses, P.H., and Dijk, D (1996), Forecasting Stock Market Volatility Using
(Non-linear) GARCH Models, Journals of Forecasting, Volume 15, 229-235(1996)
61
Gourieroux, C (2001) Financial Econometrics, Problems, models and methods.
Princeton series in nance, New Jersey, USA.
Gourieroux, C, Monfort, A and Philips, P.C (1997), Time series and dynamic
models, Cambridge university press, United Kingdom.
Hall, R., (1982), Ination, Causes and Eects, Chicago University Press, Chicago.
Hamilton, J.D., (1994), Time Series Analysis. Princeton University Press, Prince-
ton New Jersey.
Harvey, A.C (1981), Time series models. Philip Allan publishers limited, Oxford.
Harvey, A.C (1991), Forecasting, structural time series models and Kalman lter.
Cambridge university press.
International Monetary Fund., (2010), Tanzania Economic Review, Journal of
Economy, Volume 3, No.12, pp. 19-23.
Igogo, T (2010), Real Exchange Rate Volatility and International Trade ows
in Tanzania, A Dissertation submitted at the University of Dar es Salaam in
Partial Fulllment of the requirement for the degree of M.A (Economics).
Garcia, R.C., Javier, C., Marko, A., Garcia, B.S., (2003), A GARCH Forecasting
Model to Predict Day-Ahead Electricity Prices; A paper submitted to the IEE
power system Journal in Berlin.
62
Jehovanes, A (2007), Monetary and Ination Dynamics: A Lag between Change
in Money Supply and the Corresponding Ination Responding in Tanzania. Mon-
etary and Financial Aairs, Department Bank of Tanzania, Dar Es Salaam
Ljung, G.M and Box, G.E.P (1978), On a measure of lack of t in time se-
ries models, Biometrika Volume 65 page 297-303.
Lopez, J.A (1999), Evaluating the predictive accuracy of volatility models, Eco-
nomic research, Department of federal reserve bank of San Francisco.
Meyler, A, G. Kenny and T. Quinn (1998), Forecasting Irish Ination using
ARIMA models, Central Bank of Ireland Technical Paper3/RT/98, December.
Perniaag, F.E., Nor, M., Ekonomi, P.P (2004), Modelling and Forecasting on
the Malaysian Ination Rates: An Application of GARCH Models, Bangi, UKM
press.
Peng, R.D and Dominici, F (2008), Statistical methods for Environmental Epi-
demiology with R. Springer Series in Statistics, New York.
Rutayisire, L. W (1986), Measurement of government budget decit and scal
stance in less developed economy: the case of Tanzania, 1966-84, The University
of Dar Es Salaam, Eastern Africa Economic Review 2(2).
Salkind, N.J (2007), Encyclopedia of Measurement and Statistics , Volume 1,
SAGE Publications, Inc.
Suardi, S., Olekalns, Henry, O.T (2006), Testing For Rate-Dependence and Asym-
63
metry in Ination Uncertainty: Evidence from the G7 Economies, University of
Queensland, Australian Research Council.
Shephard, N (1996), Statistical aspects of ARCH and stochastic volatility time
series in Econometrics, nance and other elds, Chapman and hall, London.
Tanzania National Bureau of Statistics (TNBS): Statistics and Services Unit,http:
//www.nbs.go.tz/ retrieved on Monday, 20th September, 2010.
Talke, I. S (2003), Modelling volatility in time series data, MSc thesis, University
of Kwa-Zulu Natal.
Tsay, R.S (2002), Analysis of Financial Time Series. ISBN 0-471-41544-8.
Webster, D (2000), Websters New Universal Unabridged Dictionary. Barnes
and Noble Books, New York.
APPENDIX
Matlab Codes for the CPI data
% reading data for CPI
load CPIdata;
returns=di(log(CPIdata));
plot([0:166],CPIdata,LineWidth,2);
set(gca,XTick,[1 14 28 42 56 70 84 98 112 126 140 154 168 ]);
set(gca,XTickLabel,1997 1998 1999 2000 2001 .. 2010);
xlabel(Time);
ylabel(Rate(%));
title(Tanzania Montly Ination Rate 1997:1-2010:12);
legend(CPI data);
gure;
plot(returns, LineWidth,2);
ylabel(Return);
title(Tanzanias Monthly Dierence of Log of CPI-Ination );
axis tight;
gure;
plot(returns, LineWidth,2);
ylabel(Return);
title(Tanzanias Monthly Dierence of Log of CPI-Ination );
axis tight;
gure;
% Pre-estimation Analysis
autocorr(returns);
title(ACF with Bounds for Raw Return Series);
64
65
gure;
parcorr(returns);
title(PACF with Bounds for Raw Return Series);
gure;
autocorr(returns.square);
title(ACF of the Squared Returns);
gure;
% Ljung-Box-Pierce Q-test for ACF
(H,pValue,Stat,CriticalValue) =
lbqtest(returns-mean(returns),(101520),0.05);
(H pValue Stat CriticalValue)
% Ljung-Box-Pierce Q-test for squared ACF
lbqtest(returns-mean(returns).square,(10 15 20),0.05); (H pValue Stat Critical-
Value)
Comment:Heteroscedasticity Engle ARCH test
archtest(returns-mean(returns),(10 15 20),0.05); (H pValue Stat CriticalValue)
% Estimating Parameters GARCH(1,1)
(coe,errors,LLF,innovations,sigmas,summary) = garcht(returns);
garchdisp(coe,errors);
spec11 = garchset(P,1,Q,1,Display,o);
(coe21,errors21,LLF21) = garcht(spec21,returns);
garchdisp(coe21,errors21);
% Estimating garch(0,1)
66
% specifying number of parameters for garch(2,1) and garch(1,1) models
n21 = garchcount(coe21);
n11 = garchcount(coe11);
% the AIC and BIC for garch(1,1) model
format long
(AIC,BIC) = aicbic(LLF21,n21,166);
(AIC BIC)
format long
(AIC,BIC) = aicbic(LLF11,n11,166);
(AIC BIC)
format long
(AIC,BIC) = aicbic(LLF01,3,166);
(AIC BIC)
67
format long
(AIC BIC)
format long
(AIC BIC)
format long
(AIC BIC)
% Post estimation Analysis
garchplot(innovations,sigmas,returns);
gure;
% standardised innovations
stdresiduals=innovations./sigmas;
mean(stdresiduals);
plot(stdresidiuals, LineWidth,2);
ylabel(Innovation);
title(Standardised Innovations);
gure;
histt(stdresiduals);
gure;
normplot(stdresiduals);
gure;
% ACF of standardized innovations
autocorr((innovations./sigmas).square);
68
title(ACF of the Squared Standardized Innovations);
gure;
% Ljung-Box-Pierce Q-test and Engle ARCH test on standardized innovations
(H, pValue,Stat,CriticalValue) =
lbqtest((innovations./sigmas).square,(10 15 20),0.05);
(H, pValue, Stat, CriticalValue) =
archtest(innovations./sigmas,(10 15 20),0.05);
% loading data: non stationary time series with 168 data transform into a sta-
tionary TS with 166 data specifying the model
spec=garchset(VarianceModel,GARCH,P,1,Q,2);
spec=garchset(spec,Display,o,Distribution,T);
%taking the rst 154 values to t the model and
forecasting the next 12 values of data
for i=1:12;
(coe,errors,LLF,eFit,sFit)=garcht(spec,data(i:153+i));
(sigmaForecast,meanForecast,sigmaTotal,meanRMSE)=
garchpred(coe,data(i:153+i),1);
garchdisp(coe,errors)
(sigmaForecast,meanForecast,sigmaTotal,meanRMSE)=
garchpred(coe,data(i:153+i),1);
sForecast(i,1)=sigmaForecast;
mForecast(i,1)=meanForecast;
sTotal(i,1)=sigmaTotal;
mRMSE(i,1)=meanRMSE;
i=i+1;
(sigmaForecast,meanForecast,sigmaTotal,meanRMSE)
69
end
% time plot and predicted plot of ination rate.
plot((0:167),CPIdata,m,(0:167),CPIdata,k, LineWidth,2);
title(Actual and Forecasted Values from GARCH(1,1));
xlabel(Time);
ylabel(Rate(%));
title(Actual and Forecasted values from GARCH(1,1));
legend(Forecasted values,Actual values);

Ngailo Edward

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ngailo Edward

Uploaded by

Copyright:

Available Formats

MODELLING AND FORECASTING USING TIME

SERIES GARCH MODELS: AN APPLICATION OF

. It is more practical to set condition on r

= 0, the Hessian is given by

. For Newton Raphson based routines = 1 and M =

and for the

s and standard error makes the model unfavourable.

You might also like