American Journal of M athematics and Statistics 2013, 3(6): 315-331

DOI: 10.5923/j.ajms.20130306.04

Heteroscedastic Analysis of the Volatility of StockReturns

in Nairobi Securities Exchange
Mutai Cheruiyot Noah* , Mungatu Kyalo Joseph, Waititu Gichuhi Anthony

Jomo Kenyatta University of Agriculture and Technology, Nairobi, Kenya

Abstract Heteroscedasticity arises when the error term of a regression equation does not have a constant variance.
Financial markets are known to be very uncertain a phenomenon called volatility which is a key variab le used in many
financial applications such as investment, portfo lio construction, option pricing and hedging as well as market risk
management. This study models the heteroscedasticity of volatility of stock returns in Nairobi Securities Exchange, NSE of
Safarico m and Kenya Co mmercial Bank, KCB using daily return series fro m 9th June 2008, to 31st December, 2010, using
ARIMA-ARCH/ GA RCH models. The procedure for building the model involved model identificat ion, order determination,
estimation of parameters and diagnostic check.ShapiroWilk testrejected the null hypothesis of normality for both series at
5% level of significance while Ph ilip Perron (PP) and Aug mented Dickey Fuller (ADF) reveal that price series were not
stationary while returns series were stationary. All the return series exh ibit, leptokurtosis, volatility clustering and negative
skewness. The estimation results reveal that ARIMA (1, 0, 0) -GA RCH (1, 1) and ARIMA (0, 0, 2)-GA RCH (1, 1) best fits
Safarico m and KCB respectively. Investors who wish to avoid large, errat ic swings in portfolio returns may wish to structure
their investments to produce a leptokurtic distribution. Further, researches shou ld focus on the calculation of value-at-risk
(VaR) in the markets.

Keywords Heteroscedasticity, Volat ility, Returns, ARIMA-GA RCH-models

volatile and exh ib it volatility clustering. Due to the

1. Introduction exponential growth in those investing in stocks, modeling
and analyzing volatility of stock market returns has become
In recent years, modeling and analyzing stock return
an important research area in financial markets and has
volatility is one of the most important aspects of financial
received much attention from market practitioners, analysts
market develop ments, providing an important input for
and organizations with the aim of co ming up with robust
portfolio management, option pricing and market regulation
models that can predict future prices. This extensive research
[1]. An investors choice of a portfolio is intended to
reflects the importance of volatility in investment, security
maximize the expected return subject to a risk constraint, or
valuation, risk management and monetary policy making [1]
to minimize his risk subject to a return constraint. An
Both academicians and practitioners recognize that
efficient model for forecasting of an assets price volatility
volatility is not directly observable and that financial returns
provides a starting point for the assessment of investment
show certain characteristics that are specific to financial time
risk. To price an option, one needs to know the volatility of
series such as volatility clustering and leverage effect[3].
the underlying asset. This can only be achieved through
Financial economet ricians have developed many time-vary i
modeling the volatility. Vo latility also has a great effect on
ng volatility models among them, the Autoregressive
the macro-economy. High volatility beyond a certain
Conditional Heteroscedastic(ARCH) model proposed by
threshold will increase the risk of investor loses and raise
Engle[4] and its extension, the Generalized Autoregressive
concerns about the stability of the market and the wider
Conditional Heteroscedastic(GARCH) developed by
economy[2].
Bollerslev[3], and Taylor[5] wh ich have been applied widely.
In Kenya and other countries, investing in stocks has
This research seeks to investigate the dynamics of stock
attracted many indiv iduals. This can be ev idenced by the
return volatility in NSE. This is due to the growth in those
nu mb er o f peop le who sho wed int erest in bu y ing t he
investing in stocks in Kenya and it has become one way of
Safarico m IPOs during its inception in 2008. Returns from
building wealth. Investors normally anticipate for high
these stocks tend to fluctuat e over t ime. They are thus
returns but are also aware of the risk involved due to
* Corresponding author: fluctuation in prices.
ncheruiyot3@gmail.com (Mutai Cheruiyot Noah)
Published online at http://journal.sapub.org/ajms
Copyright 2013 Scientific & Academic Publishing. All Rights Reserved 2. Literature Review
316 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange
Financial time series modeling has been a subject of
considerable research both in theoretical and emp irical
statistics and econometrics. Various linear and non-linear
methods by which such forecasts can be achieved have been
developed in the literature and extensively applied in
practice to describe stock return volatility. Such techniques
range fro m linear to non-linear models. Poterba[6] take into
account the linear model and specify a stationary AR (1)
process for volatility of the S&P 500 index. Another study by
Frenchet al[7] uses a non-linear stationary ARIMA (0, 1, 3)
model to describe the volatility of the S&P 500 index.
The extensive use of such models is not surprising since
they provide good first order approximat ion to many
processes. Linear t ime series models however are not robust
to describe certain features of a volatility series. For instance
there are well-defined empirical ev idences that stock returns
have a tendency to exhibit clusters of outliers, imp ly ing th at
large variances tend to be followed by another large variance.
They are unable to exp lain a number of important features
common to much financial data, including leptokurtosis,
volatility clustering, long memo ry, volat ility smile and
leverage effects. That is, because the assumption of
homoscedasticity (or constant variance) is not appropriate
when using financial data, and in such instances it is
preferable to examine patterns that allow the variance to
depend upon its history. Thus such limitations of linear
models have motivated many researchers to consider
non-linear alternatives. The Autoregressive Conditional
Heteroscedastic (ARCH) model of Engle [3], the generalized
ARCH (GA RCH) model of Bo llerslev[3] and exponential
GA RCH (EGARCH) model of Nelson[8] are the common
non-linear models used in finance literature. These ARCH
class models have been found to be useful in capturing
certain non-linear features of financial t ime series such as
heavy tailed distributions and clusters of outliers.
A study by Akgiray[9] uses a GARCH(1,1) model to
investigate the time series properties of the stock returns and
reports GARCH to be the best of several models in
describing and forecasting stock market volat ility. Anil &
Higgins[10] investigated the volatility of the conventional
ordinary least squares to estimate optimal hedge rat io
estimates using future contracts. Similarly, Najand[11]
examines the relative ability of linear and non-linear models
to forecast daily S&P 500 futures index volatility. The study
finds that non-linear GA RCH models perform best. Benoit
[12] ut ilized the infinite variance distributions, when
considering the models for stock market price changes.
Fama[13] when modeling stock market prices attributed their
discrepancies to the possibility of the process having stable
innovations and thus fitted an adequate model on this basis.
Markov-Switching models have also been used to capture
the volatility dynamics of financial t ime series. This is
because they give rise to a plausible interpretation of
nonlinearities. Markov switching model of stock returns was
originally proposed by M, Start z, & Nelson[14]. Bhar[15],
among others employ markov switching models for the
modeling of stock returns.
There is a significant amount of research on volatility of
stock markets of developed countries. For instance, Gary[16]
applied the GA RCH model to the Shanghai Stock Exchange
while Bert ram[17] modeled Australian Stock Exchange
using ARCH models. Other studies on these stock markets
include Baudouhat[18] who utilized the GA RCH model in
analyzing the Nordic financial market integration. Walter
[19] applied the structural GA RCH model to portfolio risk
management for the South African equity market as well
Hongyu[2] who forecasted the volatility of the Chinese stock
market using the GARCH-type models. Elie[20] co mpared
the GA RCH model and the EGARCH under three
distribution assumptions: the Gaussian, the t-student and the
general error distributions. He showed that the distribution of
returns is far fro m being normally distributed with fat tails
and volatility clustering being persistent. Al-Jafari[21]
utilized a non-linear symmetric GA RCH(1,1) model and two
non-linear asymmetric models, TA RCH(1,1) and EGA RCH
(1,1) to Muscat Securities Market and the emp irical findings
provide no presence of day-of the week effect
The Sub-Saharan Africa has been under-researched as far
as volatility modeling is concerned. Studies carried out in the
African stock markets include, Frimpong Joseph Magnus
[22] who applied GA RCH models to the Ghana Stock
Exchange. Brooks[23] examined the effect of political
change in the South African Stock Market;Appiah-Kusi[24]
investigated the volatility and volat ility spillovers in the
emerging markets in Africa. More recently, Emen ike[25]
applied the EGA RCH model to the Kenyan and Nigerian
Stock Market returns. Fro m the availab le literature, the NSE
just like other Sub Saharan Africa Equ ity Markets has been
under-researched as far as market volatility is concerned and
therefore this study contributes to the small literature
available on the Nairob i stock market.
These developments in financial econo metrics suggest the
use of nonlinear time series structures to model the stock
market prices and the expected returns. The focus of
financial time series modeling has been on the ARCH model
and its various extensions. However, the ARCH has
limitat ions in that it treats negative and positive returns in the
same way. It is also very restrictive in parameters and often
over predicts the volatility because it responds slowly to
large shocks. GA RCH models have proved adequate in
modeling and forecasting volatility. GA RCH for instance
takes into account excess kurtosis i.e. fat tail behavior and
volatility clustering which are two important characteristics
if time series. It also provides accurate forecast of variances
and covariance of asset return through its ability to model
time varying conditional variances.
However, GA RCH is only part of a solution. Although
GA RCH models are usually applied in return series financial
decisions are rarely based solely on expected returns and
volatilities. GA RCH models are parametric specifications
that operate best under relatively stable market conditions.
Also GA RCH is explicit ly designed to model time -varying
conditional variances. GARCH models often fail to capture
highly irregular phenomenon. These include rebounds and
 American Journal of M athematics and Statistics 2013, 3(6): 315-331 317

other highly anticipated events that can lead to significant Where Yt is the price of the asset at day t. Yearly
structural change. Further, GA RCH models fail to capture arith metic returns are defined by:
the fat tails observed in asset return series. Some scholars
YT Y0
favor Markov-Switching models claiming that; Markov- R (3)
Switching models are more accurate and provide better Y0
forecasts than a variety of linear and non-linear GA RCH Where Y0 and YT are the prices of the asset at the first
models for instance[14]. In this paper we use ARIMA-
and the last trading day of the year, respectively. Then, R
ARCH/ GA RCH models of stock returns to model the
may be written as
heteroscedastic nature of volatility of stock returns in the
YT
Nairobi stock market over the period June 6, 2008 to
R 1
December 31, 2010. Y0
YT YT 1 Y1
3. Materials and Methods . ... 1
YT 1 YT 2 Y0
3.1. Materials T
Y
3.1.1. Data for the Study t 1 (4)
i 1 Yt 1
The data used in this study comprise Safarico ms and
KCBs daily returns series over the period June 6, 2008 to Definition: Let Yt and Yt 1 be todays and yesterdays
December 31, 2010. The closing prices were obtained prices of an asset or portfolio, then the geometric returns are
fro mNairobi Securit ies Exchange. Since the return of an defined as
asset is a complete and scale free su mmary of an investment Y
with attractive statistical features, we used return series X t log t (5)
rather than the price series[26]. Yt 1
Note: The yearly geometric returns are given by
3.2. Methods
Y
X t (6)
3.2.1. Volat ility Definition and Measurement
Y0
Vo latility refers to the fluctuation observed in some Fro m (6), we have that X may be written as
phenomenon over time. In modeling and forecasting
Y
literature it refers to the conditional variance o f the X t log t
underlying asset return. It is measured as the samp le standard Yt 1
deviation;
T Y
1 N log t
(ri )
N 1 i 1
2
(1)
t 1 Yt 1

T
Y
Where is the standard deviation, r i is the return on
log t
day i and is the average return over the N-day period t 1 Yt 1
T
3.2.2. Basic Statistics of Returns log X t (7)
t 1
3.2.2.1. Descriptive Statistics
i.e. the yearly geometric returns are equal to the sum of the
Analyzing financial prices d irectly is d ifficu lt because daily geo metric returns
consecutive prices are correlated, and the variances of prices
frequently increase with t ime. Consequently we use price 3.2.3. The Normality Test
changes to analyze prices. There are two main types of price This tests the likelihood that the given data set {x 1 x n }
changes that are used: arith metic and geometric returns.[27]
comes fro m a Gaussian distribution. A great number of tests
Definition: Let Yt and Yt 1 be todays and yesterdays have been devised for this problem. One of the tests used is
prices of an asset or a portfolio, the arith metic returns are the ShapiroWilk test. In statistics, the ShapiroWilk test
defined by tests the null hypothesis that a sample {x 1 x n } came
Yt Yt 1
rt
fro m a normally distributed population. It was published in
(2) 1965 by Samuel Shapiro and Martin Wilk. The test statistic
Yt 1 is:
318 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange

2 follows a 2 distribution. The null hypothesis is rejected if

n
ai x( i ) the test statistic is larger than critical value of 2.

S=
i 1 (8)
2 3.2.6. Testing for Stationarity and Autocorrelation
n
_

xi x
i 1
Test for stationarity is conducted with the Augmented
Dickey Fuller (ADF) and Ph ilip Perron (PP) test. The null
Where hypothesis is that the return series have unit roots or in other
i) x (i ) with parentheses enclosing the subscript index (i) words, the series is non-stationary. The null hypothesis is
rejected if the test statistic is larger in the absolute term than
is the i th order statistic, i.e., the i th -smallest number in the the critical value[28]
sample; Having confirmed that all return series are stationary, we
( x1 ... x n ) shall continue to examine the autocorrelat ion and the partial
ii) x is the sample mean;
autocorrelation in the series to identify their proper structures.
n
This is done through the Ljung-Bo x Q-statistic test by (Bo x
iii) the constants a i are given by & George[29] wh ich is defined as:
mT V 1 = +2 2
=1
2
~ (12)
(a1 ,..., an ) =
m V
1 (9)
1 Where is the samp le autocorrelat ion coefficient; T is
T
V 1m 2
the sample size and m is the maximu m lag length
Where m (m1 ,..., m n ) T are the expected values of The null hypothesis that all are zerois rejected if the
value of the computed Q is larger than the critical Q-statistic
the order statistics of independent and identically-d istributed
fro m the chi-square distribution at the g iven level of
random variables sampled fro m the standard normal
significance. According to Harvey & Jaeger[30], choosing
distribution, and V is the covariance matrix of those order
the number of lags for the test is a practical issueas a small
statistics.
number of lags might fail to detect the autocorrelations at
3.2.4. Volat ility Clustering high-order lags, whereas, a large nu mber of lags might result
in diluting the significant correlat ion at one lag by
This is determined by computing the ACF. Given that insignificant correlations at other lags.

X t is a stationary time series, with constant expectation
3.2.7. Volat ility Modeling Techniques
and time independent covariance. The ACF for the series is
defined as
3.2.7.1. ARCH Model
Cov( X t , X t k ) (k )
k (10) The ARCH model was introduced by Engle [4] in his study
Var( X t )Var( X t k ) (0) Autoregressive Conditional Heteroscedasticity with
estimates of the Variance of United Kingdom In flat ion, as
for k 0 and k k The value k denotes the lag.
the first formal model which seemed to capture the
Plots of ACF as a function of k shall be done, and phenomena of changing variance in t ime series data. It is
determine if the autocorrelation decreases as the lag gets most widely used discrete time model for analysis of
larger or of if there is any particular lag for wh ich the financial data. The fo rmulat ion of his model is given below:
autocorrelation is large q

3.2.5. Testing for ARCH Effects t i t i 2 Where t ~IID (0, 1) (13)

i 1
Before fitting the autoregress ive models to each of the
daily returns series, the presence of A RCH effects in the 2t = 0 1 t 1 2 2 2 t 2 ... q t 1 2
residuals is first tested. If there does not exist a significant
where t is the variance at time t, t is square
2 2
ARCH effect in the residuals then the ARCH model is
mis-specified. Testing the hypothesis of no significant residuals at rime t, and q is the nu mber of lags. The effect of a
ARCH effects is based on the Lagragian Mult iplier (LM) return shock i period ago (i q) on current volatility is
approach, where the test statistic is given by governed by the parameter . In an A RCH model, o ld news
arrived at the market mo re than q period ago has no effect at
LM =nR 2 (11)
all on current volatility. For ARCH (1, 1) the model is
2
Where n =sample size , R =the coefficient of
determination for the regression in the ARCH model using
t 2 0 1 2 t 1
the residuals. The null hypothesis is that there is no ARCH
effect up to order in the residuals. The test statistic is 3.2.7.2. GA RCH Model
calculated as the number of observations mult iplied by 2 Bollerslev[3] extended the basic ARCH model by
fro m the regression. The LM test statistic asymptotically introducing the GARCH model which has proven to be quite
 American Journal of M athematics and Statistics 2013, 3(6): 315-331 319

useful in empirical work. He suggested that the conditional
variance function be specified as follows: Y T =X T +
t is the mean equation. Where Y t is the stock return, X t
is the exogenous variables or belonging to the set of
informat ion Y t 1 , is a fixed parameter vector and
By considering the observed values x 1 , x 2 x n to be
fixed parameters of this function, whereas will be the
function's variable and allowed to vary freely. And this
function is called likelihood
L( /x 1 , x 2 ,, x n )= f{x 1 , x 2 ,, x n / }
n

conditional variance is,
q p = f (x / ) i (17)
h t = 0 i t i i ht i
2
(14)
i 1

i 1 i 1 In practice, it is often more convenient to work with the

logarith m of the likely-hood function and called the
Where 0 0, 1 , 2 ,..., q 0 and log-likelihood:
1 , 2 ,..., p 0 n

The GARCH (p, q) above defined as stationary when ( 1

ln L( / x 1 , x 2 ,, x n )=
i 1
ln f ( x / ) i (18)

+ 2 +. + 0 ) + (1+ 1+... + p ) 1 . In this study Assume observations x 1 , x x follow normal

2 n
we are going to use GA RCH (1, 1). The model for GA RCH
distribution with un-known parameters = { , 2 } then
(1, 1) is given by 2 t 0 1 2 t 1 2 t 1 where,
n ( xi )2
1
0 >0, 1 0 and 1 0. ln L( / x 1 , x 2 ,, x n ) = ln{ e 2 2
} (19)
2 i 1
2

3.2.8. Bu ild ing a Vo latility Model n

1 ( x )2
= ( ln 2 ln i 2 ) (20)
3.2.8.1. Model Identification i 1 2 2
2
Under the identification stage the following wasdone: n 1 n
i. Converting of daily closing price series to return series. =-
2
ln 2 n ln 2
2
( xi )
i 1
(21)
Let Y t denote the daily closing price of a stock at the end of
the day t, the daily stock return series is generated by In this case we have and 2 as the un-known
Yt parameters
2
r t =ln (15)
n 1 n
2
Yt 1 L( , ) = - ln 2 n ln
2
( xi ) (22)
2 2 i 1
ii. Stationarity of the return series was checked using unit
L 1 n
root test. Lagrange Multip lier (LM) and Ljung-Bo x
statistics is used to test for ARCH effects on thesquared
=
2

I
(x )
1
i (23)
residuals of the regressed AR (p) process, since GARCH (p,
L n 2
n
3 ( xi )
q) model imp lies ARCH (r = q + p) model. Under the null
(24)
hypothesis that there is no ARCH effects ( 1 = p ), i 1
2
the LM test statistic equal to TR has asymptotic chi Equating this to zero and solving for and gives
-squared distribution with p degree of freedo m.
iii. An ARIMA(p,d,q) model was fitted to the data to x (25)
remove serial dependence n 2
iv. A CF, PACF and AICc was used to determine the
2 (x )
i
(26)
order of the models i 1

n
3.2.8.2. Parameter Estimation
Parameter Es timation for GARCH (p, q) model
The estimation of the models parameters was
implemented by Maximu m Likelihood Method under the Let us now look at the application of Maximu m
normal distribution. Th is involves choosing values for the Likelihood Estimation, MLE to estimate the parameters of
parameters that maximizes the chance (or likelihood) of the GA RCH (p, q). To estimate parameters of GARCH (p, q)
given k, p and q we have
data occurring. Given a samp le {x 1 , x 2 x n } of n, IID k
observations, which comes fro m a distribution f(x) with
unknown parameter , then; the joint density function is
y t =C+
i
ay
1
i t i t (27)

f{x 1 , x 2 ,, x n / }=f(x 1 / )xf(x 2 / )xx f(x n / )(16) t v t ht (28)

320 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange

q p ii. The normal plots, ACF p lot and time series plot was
h t = 0 i t i 2 i ht i (29) done. The normal probability plot should be a straight line
i 1 i 1 while the time p lot should exh ibit random variation. For
Where v t is the wh ite noise term, t is normally ACFs all the correlation should be within the test bounds
which indicates stationarity in the data.
distributed with mean zero and variance h t iii. Ljung-Bo x test is employed to check for adequacy of
the fitted model. The Ljung-Bo x test was named after Greta

2t
p( t / t 1 ,..., 0 ) =
1
e 2 ht
(30)
M Ljung and George E. P. It is a type of statistical test
which test whether any of a group of autocorrelat ions of a
2 h t time series is different fro m zero. It performs a lack-of-fit
The log-likelihood function of the parameter vector = hypothesis test for model specification, wh ich is based on
the Q-statistic
( 0 , 1 ,..., q , 1 ,..., p )
T
n
p2 ( j)
becomes Q=n(n+2)
j n j
(39)
n n
1 1 2 1
L( )= l ( ) t q 2 ln 2 2 ln h 2h ) (31)
t q 1
t
1
t
t
where n= sample size, h= number o f auto-correlat ion lags
2
t
included in the statistic, and p ( j ) is the squares sample
l t ( ) t 1 ht
2
autocorrelation at lag j. Under the null hypothesis of no
( 2 ) (32)
2h t 2ht serial correlation, the Q-statistic is asymptotically
Chi-Square distributed. If the value of the test statistic is
2lt ( ) 2t 1 2 ht greater than the critical value fro m the Q-statistics, then the
( )
T 2h 2t 2ht T (33) null hypothesis can be rejected. Alternatively, if p-value is
smaller than the conventional s ignificance level, the null
1 2t ht ht hypothesis that there are no autocorrelation will be rejected.
( )
2h 2t h3t T 3.2.9. Volat ility Forecasting
Where The challenge in Econometrics is to specify how the
ht p
h
(1, 2t 1 ,..., 2t q , ht 1 ,..., ht p )T i t i (34) informat ion will be used to forecast the mean and variance of
i 1 the return, conditional on the past information. Various
methods have been considered for the mean return to
Thus the gradient is
forecast future returns. The most widely used specification is
1 n 2t 1 ht
L( ) ( 2 ) (35)
the GARCH (1, 1) model introduced by Bo llerslev[3] as a
2 t q 1 h t ht generalization of Eng le[4].
Consider the following GA RCH (1, 1) model:
And the Fisher Information matrix is
n

2
2t ht ht
1 ht
2
1
yt ut , u t N(0, t ),
2
(40)
J= E[( 2 ) ( 2 3 )
t
] (36)
t q 1 2h t 2ht
T
2h t h t T t 2 0 1u t21 t 1 2 (41)

1 n 1 ht ht What is needed to generate are forecasts of T+1 2 T,

=- E( 2
2 t q 1 h t T
) (37) T+2 2 T...T+s 2 T where T denotes all information
available up to and including observation T. Adding one to
each of the time subscripts of the above conditional variance
3.2.8.3. Diagnostic Checking equation, and then two, and then three would yield the
following equations
Goodness-of-fit needs to be performed after fitting the
T+12 = 0 + 1 +T2 (42)
appropriate model Tackle[31]. This is based on the
standardized residuals. The fo llo wing was performed: T+2 2 = 0 + 1 +T+1 2 (43)
i. The standardized residuals of the fitted model are T+3 2 = 0 + 1 +T+2 2 (44)
2
analyzed to ascertain their randomness. The standardized Let 1,T f be the one step ahead forecast for 2 made at
residuals
time T. Th is is easy to calculate since, at time T, the values of

t t (38)
1,T f
2
t all the terms on the right hand side are known. ,
are IID random variables following either a standard normal will be obtained by taking the conditional expectation of (40).

Given 1,T
f2
or student-t distribution. If the model fits well then neither
2,T f
2

2
, the 2-step ahead forecast for 2
t nor t should exh ibit serial correlat ion. made at t ime T is obtained by taking the conditional
 American Journal of M athematics and Statistics 2013, 3(6): 315-331 321

expectation of (41) amp litude vary with time a phenomenon called A RCH
f2
2,T f2
= 0 + 1 E ( u 2 T 1 T) + 1,T (45)
effects. Volat ility clustering is also evident.

KCB CLOSING PRICE

Where E( u 2 T 1 T) is the expectation, made at time T,
2
of u T 1 , wh ich is the squared disturbance term. We can
write

30
2
E ( u T 1 t ) = T+1 2 (46)

ClosingPrice
But T+1 2
is not known at time T, so it is replaced with the

25
1,T f
2
forecast for it, , so that the 2-step ahead forecast is
given by

20
2,T f 1,T f 1,T f
2 2 2
= 0 + 1 + (47)

2,T f
2 2
= 0 + (1 +) 1,T
f

15
(48)
0 100 200 300 400 500 600
By similar argu ments, the 3-step a-head forecast will be
given by Time
2
3,T f a)
= ET ( 0 + 1 + T+2 ) 2
(49)
2
= 0 + (1 + ) 2 ,T
f
(50) SAFARICOM CLOSING PRICE

= 0 + (1 + )[ 0 + ( 1 +) 1,T
f2
] (51) 8
2
= 0 + 0 (1 + ) + ( 1 +)
2
1,T f (52)
7
Closing_Price

Any s-step a-head forecast (s 2) would be produced by

6

s 1
hsf,T 0 (1 ) i 1 (1 ) s 1 h1f,T (53)
5

i 1
4
3

4. Results and Discussion

4.1. Data Explorati on 0 100 200 300 400 500 600

The data employed in this study comprise Safarico ms and Time

KCBs closing price over the period June 6, 2008 to b)
December 12, 2010 wh ich constitutes a sample of 653 Figure 1. Time series plot of KCB and Safaricom closing price
observations. The closing prices were obtained fro mNairobi
Securities Exchange. Let Y t denote the daily closing price SAFARICOM RETURNS

of a stock at the end of the day t, the daily stock return series
was be generated by
0.05

Yt
r t = ln (54)
Yt 1
0.00
Safrt

Fro m Figure 1 above the closing prices are very irregular

with varied degree of fluctuations. The time plots clearly
-0.05

show that the mean and variance are not constant, showing
non-stationarity of the data. It also shows a drop in prices
-0.10

fro m a high value in 2008 to a low value in 2010. Series such

as thesecannot be used for further statistical in ferences
because of their implications Gujarat i[32], thus theneed to
0 100 200 300 400 500 600
transform them to returns. The plots of daily returns of
Safarico m and KCB are p resented in Figure 2 belo w. The Time
plots for returns are stationary and exhibit no trendand the a)
322 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange

4.1.1. Descriptive Statistics for the Prices and Returns

KCB RETURNS
0.10 Table 1 below shows summary statistics for the two
companies return series. The results indicate high volatility
and the risky nature of the market since the standard
0.05

deviation of the market returns is h igh in co mparison with

the mean. A lso the standard deviations are very close for
Kcbrt

both Safarico m and KCB with Safarico m being slig htly

0.00

volatile. Both price series have positive skewness imp lying

that the distribution has a long right tail. On the other hand,
-0.05

the return series for Safarico m have negative skewness

imply ing that the distribution has a long left tail and positive
-0.10

for KCB imp lying that the distribution has long right tail.
The values for kurtosis are high (above three) for both return
0 100 200 300 400 500 600
series imply ing they are leptokurtic. The Shapiro-Wilk test
Time rejects normality at the 5% level for all series. So, the
b) samples have all financial characteristics: volatility
Figure 2. Plots of Safaricoms and KCBs returns r t clustering and leptokurtosis.

Table 1. Descriptive statistics for prices and returns

KCB Closing price Safaricom Closing price

A. Prices
Mean 22.243 4.6474
Median 21.750 4.6500
Minimum 15.500 2.7000
Maximum 33.000 8.1500
Standard deviation 3.5203 1.1823
C.V. 0.15826 0.25441
Skewness 1.2007 0.35244
Ex. kurtosis 1.3145 -0.36618
Shapiro-Wilk 0.887923 0.959455
Observations 653 653
B. Returns KCB returns Safaricom returns

-0.00062188
Mean
0.00000 -0.00079954
Median
-0.093090 0.00000
Minimum
0.093932 -0.11935
Maximum
0.021731 0.084557
Standard deviation
34.945 0.022302
C.V.
0.29534 27.893
Skewness
3.6479 -0.086079
Ex. kurtosis
0.905841 3.9471
Shapiro-Wilk
652 0.916475
Observations
652

4.1.2. Test for Normality and Unit Root

Shapiro-Wilk test is used to test for normality in the series wh ich are shown in the table below
Table 2. Shapiro-Wilk test for Normality for the two series

Shapiro-Wilk test
W p-value
12
Safaricom 0.9595 1.927x10
16
KCB 0.8879 2.2x10

A stationary check for both closing prices and returns using Augmented Dickey Fuller (A DF) Philip Perron (PP) test
shows that under the null hypothesis, unit root is not detected in both returns
 American Journal of M athematics and Statistics 2013, 3(6): 315-331 323

Table 3. ADF and PP test for prices and returns for Safaricom and KCB

Safaricom Prices Returns

ADF Test PP test ADF Test PP Test
ADF Value -3.5 PP Value -3.27 ADF Value -9.13 PP Value -22.7
P-value 0.04 P-value 0.075 P-value 0.01 P-value 0.01
ADF Value -2.71 PP Value -2.86 ADF Value -9.38 PP Value -22.38
KCB
P-value 0.27 P-value 0.21 P-value 0.01 P-value 0.01

ACF OF SAFARICOM RETURNS ACF OF SAFARICOM SQUARED RETURNS

1.0

1.0
0.8

0.8
0.4 0.6

0.6
ACF

ACF

0.4
0.2

0.2
0.0

0.0
0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag

ACF OF KCB RETURNS ACF OF KCB SQUARED RETURNS

1.0
1.0

0.8
0.8
0.6

0.6
ACF
ACF

0.4

0.4
0.2

0.2
0.0

0.0

0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
Figure 3. ACF of Asset Returns and Squared Asset Returns for Safaricom and KCB

4.2. ARIMA (p, d, q) Modeling

4.2.1. Model Identification

Use ACF and PA CF identify the A RIMA model for the mean equation
The upper left graphs show ACF of Log Safarico m closing price, showing the A CF slowly decreases. It is probably that the
model needs differencing. The lower left is PA CF of Log Safarico m closing price, in dicat ing significant value at lag 1 and
then PACF cuts off. Therefore, the model for Log Safarico m closing price might be ARIMA (1, 0, 0). The upper right shows
ACF of d ifferences of log Safarico m with no significant lags. The lo wer right is PACF of d ifferences of log Safarico m,
reflecting no significant lags. The model for differenced log Safarico m series is thus a white noise, and the original model
resembles random walk model ARIMA (0, 1, 0)
324 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange

ACF OF SAFARICOM LOG CLOSING PRICE ACF OF SAFARICOM LOG DIFFERENCE

1.0

1.0
0.8

0.8
0.6

0.6
ACF

ACF
0.4

0.4
0.2

0.2
0.0

0.0
0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
a) b)

PACF OF SAFARICOM LOG CLOSING PRICE PACF DIFFERENCE LOG SAFARICOM

1.0

0.10
0.8

0.05
Partial ACF
Partial ACF

0.6
0.4

0.00
0.2

-0.05
0.0

0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
c) d)
Figure 4. ACF and PACF Safaricom closing and log differenced closing price

ACF OF KCB LOG CLOSING PRICE ACF OF KCB LOG DIFFERENCE

1.0

1.0
0.8

0.8
0.6
0.6
ACF

ACF

0.4
0.4

0.2
0.2

0.0
0.0

0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
a) b)
 American Journal of M athematics and Statistics 2013, 3(6): 315-331 325

PACF OF KCB LOG CLOSING PRICE PACF OF KCB LOG DIFFERENCE

1.0

0.10
0.8

0.05
Partial ACF
0.6
Partial ACF

0.4

0.00
0.2

-0.05
0.0

0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
c) d)
Figure 5. ACF and PACF KCB closing and log differenced closing price

Figure 5 shows the ACF and PA CF o f closing price and Based on the AICc, values presented in Table 4, the
log difference of closing price. The graphs show the same ARIMA (1,0,0) model is identified to be the one that best
trend as for Safarico m discussed above. Further, AICc fits the daily returns for Safarico m and ARIMA(0,0,2) for
provides another way to check and identify the model. This KCB fro m June 2008 to Dec 2010.
can be calculated by the formu la:
4.2.2. Parameter Estimation
AICc = N* log (SS/N) + 2(p + q + 1)*N/ (N p q 2), if
no constant term in model The parameters of the fitted ARIMA models are shown in
AICc = N* log (SS/N) + 2(p + q + 2)*N/ (N p q 3), if the table below
constant term in model Table 5. Estimated parameters for ARIMA (1, 0, 0) and ARIMA (0, 0, 2)
N: the number o f items after differencing (N = n d)
SS: sum o f squares of differences Model 1 Intercept 1 2
p& q : the order of autoregressive and moving average
model, respectively. According to this method, the model 0.1188 ------
with lowest AICc will be selected. Fitting the various orders Safaricom ARIMA(1,0,0)
-0.0008 ------
of ARIMA in R g ives the values in Table 3 belo w. ------- 0.1140
KCB ARIMA(0,0,2)
-0.0008 0.0672
Table 4. AICc values for the candidate ARIMA (p, d, q) models
Thus the complete models become for the fitted ARIMA
Model Safaricom KCB
(1, 0, 0) and ARIMA (0, 0, 2) for Sfarico m and KCB
AICc AICc respectively becomes;
ARIMA(1,0,0) -3113.21 -3113.206 Safaricom: r t = r t = -0.0008+ 0.1188 1 + t
ARIMA(0,0,1)
ARIMA(1,0,1)
-3047.012
-3049.893
-3112.235
-3112.479
KCB: r t = + 0.1140r t 1 + 0.067r t 2 + t
ARIMA(1,1,0) -2952.173 -3048.920
ARIMA(0,1,1) -3117.038 -3112.480 4.2.3. Diagnostic Checking
ARIMA(0,1,2) -3103.62 3103.62
ARIMA(0,0,2) -3059.893 -3157.22 We plot he ACF and the PACF of residuals to check for
model adequacy.
326 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange

ACF OF SAFARICOM RESIDUALS ACF OF KCB RESIDUALS

1.0
1.0

0.8
0.8

0.6
0.6

ACF
ACF

0.4
0.4

0.2
0.2

0.0
0.0

0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
a) b)

PACF OF SAFARICOM RESIDUALS PACF OF KCB RESIDUALS

0.05
0.05

Partial ACF
Partial ACF

0.00
0.00

-0.05
-0.05

0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
c) d)

Figure 6. ACF and PACF of Safaricom and KCB residuals

The ACF and PA CF shows no significant lag hence the models are appropriate

4.3. ARCH/ GARCH Modeling

Although ACF & PACF of residuals have no significant lags, the time series plot of residuals shows some cluster of
volatility (not reported here). ARIMA is a method to linearly model the data and the forecast width remains constant because
the model does not reflect recent changes or incorporate new informat ion. Ho wever, we fit an ARIMA (p, d, q ) model to
remove serial dependence in the series. Inspection of residual plot d isplays and squared residual plot shows cluster of
volatility. The ACF & PA CF of squared residuals confirms this and thus if the residuals (noise term) are not independent and
can be predicted. Hence, ARCH/ GARCH should be used to model the volat ility of the series to reflect more recent changes
and fluctuations in the series . Followings are the plots of squared residuals.
 American Journal of M athematics and Statistics 2013, 3(6): 315-331 327

SAFARICOM SQUARED RESIDUALS SAFARICOM ACF SQUARED RESIDUALS

1.0
res.arima100 * res.arima100

0.012

0.8
0.6
0.008

ACF

0.4
0.004

0.2
0.0
0.000

0 100 200 300 400 500 600 0 5 10 15 20 25

Time Lag
a) b)

SAFARICOM PACF SQUARED RESIDUALS KCB SQUARED RESIDUALS

0.000 0.002 0.004 0.006 0.008

0.25

res.arima002 * res.arima002
0.15
Partial ACF

0.05
-0.05

0 5 10 15 20 25 0 100 200 300 400 500 600

Lag Time
c) d)

KCB ACF SQUARED RESIDUALS KCB PACF SQUARED RESIDUALS

0.25
1.0
0.8

0.15
Partial ACF
0.6
ACF

0.4

0.05
0.2

-0.05
0.0

0 5 10 15 20 25 0 5 10 15 20 25

Lag Lag
e) f)
Figure 7. ACF and PACF plots of residuals and squared residuals of Safaricom and KCB

residuals is tested. If there does not exist a significant ARCH

4.3.1. Testing for ARCH effects in Returns of
Fitted ARIMA (1,0,0) and ARIMA (0,0, 2)
Before fitting the autoregressive models to each of the
daily returns series, the presence of A RCH effects in the
t in the effect in the residuals then the ARCH model is mis -specified.
Testing the hypothesis of no significant ARCH effects is
based on the Lagragian Multiplier (LM ) approach as stated
earlier on the methodology.
Fro m Table 6, the p-values for both series are less than
328 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange

0.05 hence we reject the null hypothesis of no significant This study has included GARCH (1, 0) GARCH (0, 2) and
arch effect in the daily returns of Safarico m and KCB and GA RCH (2, 2) in order to check if they are appropriate for
conclude there are significant arch effects for the June 6, modeling time vary ing variance. We select the model with
2008 to December 31, 2010. the lowest AIC and BIC`
Fro m Table 7 above the model given in bold is taken to be
Table 6. Lagragian Multiplier test for Arch effects
the most appropriate according to the criteria above. The
Re turns Chi-square df p-value GA RCH models for different values of p and q were fitted to
the data, diagnosed and from the diagnosis and goodness of
KCB 117.15 <0.001
fit statistics, the GA RCH (1, 1) was found to be the best
Safaricom 74.5019 4 <0.001 choice. This is consistent with most empirical studies
involving the application of GARCH models in financial
4.3.2. Model Identification time series data. We thus fit a GARCH (1; 1) to the residuals
of ARIMA (1, 0, 0) and ARIMA (0, 0, 2) of Safarico m and
Since this study deals with daily returns, it is restricted to
KCB respectively.
pure ARCH (p) models. For GA RCH (p, q) models, those
with p, q 2 are typically selected by AIC and BIC. Low Table 7. AIC, BIC values of the candidate GARCH model
order GA RCH (p, q) models are generally preferred to a high
Company
order ARCH(p) for reasons of parsimony and better Model AIC BIC
numerical stability of estimat ion Safaricom GARCH(0,1) 1.534990 1.534990
GARCH(0,2) 1.534990 1.534990
4.3.3. Order Determinat ion
GARCH(1,1) 1.510699 1.521877
Determining the ARCH order p and the GARCH order q GARCH(1,2) 1.553465 1.601506
for a particular series is an important practical p roblem. The GARCH(2,1) 1.534990 1.583031
AIC, BIC and Log likelihood ration tests are used in
GARCH(2,2) 1.533339 1.588244
selecting the appropriate order of the GARCH fro m
competing models. Table 7 below g ives the suggested order
KCB GARCH(0,1) 3.752802 3.790843
with their respective fit statistics. The aim is to have a GARCH(0,2) 3.763026 3.791068
parsimonious model that captures as much variation in the
GARCH(1,1) 3.716712 3.757891
data as possible. Usually the simple GA RCH model captures
GARCH(1,2) 3.742802 3.790843
most of the variability in most stabilized series. Small lags
GARCH(2,1) 3.743026 3.791068
for p and q are common in applications. Typically GA RCH
(1, 1); GA RCH (2, 1) or GA RCH (1, 2) models are adequate GARCH(2,2) 3.739183 3.794087
for modeling volatilities even over long sample periods[33].

4.3.4. Estimat ion

Fro m R output
Estimate Std. Error t value Pr(>|t|)
a0 4.101e-05 8.491e-06 4.829 1.37e-06 ***
a1 1.866e-01 2.617e -02 7.131 9.98e-13 ***
b1 7.209e-01 3.363e-02 21.438 < 2e-16 ***

For Safarico m the fitted GARCH (1, 1) model is

r t = 5.76 + t t

t 2 =0.00004+0.186 Z 2
t 1 +0.7209
2
t 1

Fro m the fo llo wing output for KCB

Estimate Std. Error t value Pr(>|t|)
a0 2.895e-05 5.833e -06 4.963 6.95e-07 ***
a1 1.912e-01 2.592e -02 7.376 1.63e-13 ***
b1 7.597e-01 2.561e-02 29.662 < 2e-16 ***

The fitted GA RCH (1, 1) model is

r t = 20.18 + t t
t 2 =0.000028+0.19 Z 2
t 1 +0.7597 2 t 1
 American Journal of M athematics and Statistics 2013, 3(6): 315-331 329

To assess the accuracy of the estimates, the standard errors are used the smaller the better. Model fit statistics used to assess
how well the model fit the data are the AIC and BIC. Fro m the standard errors the estimates are precise. Based on 95%
confidence level, the coefficients of the fitted GARCH (1, 1) model are significantly different fro m zero.

4.3.5. Diagnostic Checking

Here the adequacy of the selected models is done. This is done by using standardized residuals which are assumed to
follow either normal or standardized t distribution. It must satisfy the requirement of a white noise. The p lots include
normal plots, ACF p lot time series plot and h istogram. If the model fits the data well the histogram of the residuals should
be symmetric. The normal probability plot should be a straight line wh ile the time p lot should exhib it rando m variation. For
ACF p lots all the correlation should be within the boundary line mean ing the data is stationary.

ACF PLOTS OF SAFARICOM RESIUALS ACF PLOTS OF KCB RESIDUALS

1.0

1.0
0.8

0.8
0.6

0.6
ACF

ACF
0.4

0.4
0.2

0.2
0.0

0.0

0 5 10 15 20 25
0 5 10 15 20 25
Lag
Lag

Figure 8. ACF plots of residuals for Safaricom and KCB

It is clear that for Safarico m all the correlations are within the test bounds imply ing the fitted model is adequate. As for
KCB the first two auto-correlat ions are outside the test bounds which might imply the model is not adequate. However, this
might be by chance and we proceed to use Q-Q p lot and

SAFARICOM ARIMA(1,0,0)-GARCH(1,1) KCB ARIMA(0,0,2)-GARCH(1,1)

4
4

Sample Quantiles
Sample Quantiles

2
2

0
0

-2
-2

-4
-4

-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3

Theoretical Quantiles Theoretical Quantiles

a) b)
Figure 9. Q-Q plots and Normal probability plot of Safaricom and KCB residuals
330 M utai Cheruiyot Noah et al.: Heteroscedastic Analysis of the Volatility
of StockReturns in Nairobi Securities Exchange

Fro m the Q-Q plots and normal probability plot the General Error Distribution to capture the stylized facts of
residuals seem to be roughly normally d istributed although return series. Further, in emerging markets, diversification
some points remain off the line. and return benefits provided have attracted significant
investors interest which have led to significant portfolio
4.4. Vol atility Forecasting equity inflo ws into these financial systems, and as a result,
The final objective of this study was to forecast the motivated the study of various aspects of stock return
volatility. Table 8 and 9 belo w shows the forecasts behavior in these markets. For that reason, an imperat ive and
Although forecast performance was not one of the contemporary filament of emp irical researches should focus
objectives of the study, comparing GARCH(1,1) and mixed on the calculation of value-at-risk (VaR) in the markets.
models using Mean Squared Erro r it was found that mixed
models outperform GA RCH(1,1)(results not presented here)
Table 8. Forecast results for Safaricom and KCB
REFERENCES
Point Forecast Low High
653 -0.0011994094 -0.04450269 0.04210387 [1] Granger, O., & Poon, S. H. (2003). Forecasting Volatility in
654 -0.0014993927 -0.04508323 0.04208444 Financial M arkets: A review. Journal of Economic Literature,
655 -0.0008070325 -0.04448788 0.04287381 478-539.
656 -0.0008070325 -0.04448788 0.04287381
[2] Hongyu, P. (2006). Forecasting Financial Volatility:
657 -0.0008070325 0.04287381 -0.04448788 Evidence from Chinese Stock M arket. Working paper in
658 -0.0008070325 -0.04448788 0.04287381 economics and finance No. 06/02. University of Durham.
United.
659 -0.0008070325 -0.04448788 0.04287381
660 -0.0008070325 -0.04448788 0.04287381 [3] Bollerslev, T. (1986). Generalized Conditional
661 -0.0008070325 -0.04448788 0.04287381 Heterosedasticity. Journal of Econometrics(31), 307-327.
662 -0.0008070325 -0.04448788 0.04287381
[4] Engle, R. (1982). Autoreationgressive Conditional
Table 9. Forecast results for KCB Heteroscedasticity with Estimates of the Variance of UK Inf.
Econometrica, 987-1008.
Point Forecast Low High
653 -0.000855497 -0.0427695 0.04105851 [5] Taylor, S. (1986). M odelling Financial Time Series.
654 -0.0007207025 -0.04296716 0.04152576 Chichester: Wiley.
655 -0.0006276681 -0.04321562 0.04196029
656 -0.0006276681 -0.04321562 0.04196029 [6] Poterba. (1986). The Persistence of Volatility and Stock
657 -0.0006276681 -0.04321562 0.04196029 M arket Returns. American Economic Review, 1142-1151.
658 -0.0006276681 -0.04321562 0.04196029
659 -0.0006276681 -0.04321562 0.04196029 [7] French, K. R., Schwert, G. W., & Stambaugh, R. F. (1987).
660 -0.0006276681 -0.04321562 0.04196029 Expected Returns and Volatility. Journal of Financial
661 -0.0006276681 -0.04321562 0.04196029 Economics, 3-29.
662 -0.0006276681 -0.04321562 0.04196029
[8] Nelson, D. (1991). Conditional Heteroscedasticity in Asset
Returns: A new approach. Ecnometrica, 59(2), 347-370.

5. Conclusions and Suggestions [9] Akgiray, V. (1989). Conditional Heteroscedasticity in time

The objectives of the study of this research work have
been largely achieved. The return series of Safarico m and
KCB series have been modeled. They reveal so me stylized
facts such as negative skewness, leptokurtosis, and
volatility clustering, nonlinear data generating process,
serial dependence and leverage effects which are common
observations in other stock markets. This is agreement with
previous researches. The null hypothesis of significant
correlations is rejected at 5% level of significance for the
two series. The results of LM test finds presence of arch
effects and that the standardized residuals are normally
distributed. It has been shown that the price is not normally
distributed,which suggests that there is evidence of fat tails
or is leptokurtic which is a common feature of financial
market returns. Investors who wish to avoid large, errat ic
swings in portfolio returns may wish to structure their
investments to produce a leptokurtic distribution.
Other researchers can use heavy tailed distributions e.g
series of stock and returns: Evidence and Forecast. Journal of
Business, 55-80.
[10] Anil, B., & Higgins, M . (1993). ARCH M odels: Properties,
Estimation and Testing. Journal of Economic Surveys,
305-362.
[11] Najand, M . (1991). Forecasting Stock Index Future Volatility:
Linear Verses Non-linear M odels. Retrieved 1 10, 2013, from
http://dx.doi.org/10-1111/1540-6288.00006
[12] Benoit, M . (1963). The variation of certain speculative
prices. Journal of Business, 36, 394-419.
[13] Fama, E. F. (1965). The Behavior of Stock-M arket Prices.
Journal of Business, 34-105.
[14] M , T. C., Startz, R., & Nelson, C. R. (1989). A M arkov model
of Heteroscedasticity ,Risk and Learning in the Stock M arket.
Journal of Financial Economics, 25, 3-22.
[15] Bhar, S. H. (2004). Empirical Characteristics of the
Traspitory Component of Stoc Return: Analysis in a M arkov
Switching Heteroscedastic Framework. Australia, Japan.
 American Journal of M athematics and Statistics 2013, 3(6): 315-331
[16] Gary, T. a. (2004). Intraday Data and Volatility M odels:
Evidence from Chinese Stocks. University of Western
Sydney.
[17] Bertram, W. (2004). An empirical investigation of Australian
Stock Exchange Data. School of M athematics and Statistics,
University of Sydney.
[18] Baudouhat, A. (2004). Nordic Financial M arket
Integration:An Analysis with GARCH. Unpublished M sc
Thesis,Gteborg University.
[19] Walter, A. (2005). A structural GARCH model: An
application to portfolio risk management Unpublished PhD
Thesis, University of Pretoria.
[20] Elie, B. (2012). An Attemt to Capture Leptokurtic of Returns
and to M odel Volatility of Returns: The case of Beirut Stock
Exchange. International Research Jornal of Finance and
Economics(90), 111-122.
[21] Al-Jafari, K. M . (2012). An Empirical Investigation of the
Day-of-the-Week Effect on Stock Returns and Volatility:
Evidence from M uscat Securities M arket. International
Journal of Economics and Finance(4), 141-149.
[22] Frimpong, J. a.-A. (2006). M odelling and Forecasting
Volatility of Returns on the Ghana Stock Exchange Using
Garch M odels. American Journal of Applied Sciences, 3(10),
2042-2048.
[23] Brooks, R. D. (1997). An Examination of the Effects of M ajor
Political Change on Stock M arket Volatility: The South
African Experience. Journal of International Financial
M arkets , Institutions and M oney(7), 255 - 275.
331
[24] Appiah-Kusi, J. a. (1998). Volatility and Volatility
Spill-Overs in Emerging M arkets CERF Discussion Paper
Series No. 98-04.
[25] Emenike, K. O. (2010). M odelling Stock Returns Volatility In
Nigeria Using GARCH M odels. Enugu State: M unich
Personal RePEc Archive.
[26] Campell, J., Lo, A., & M acKinlay, C. (1997). The
Econometric of Financial M arket. New Jersey: Princeton
University Press.
[27] Jorion, P. (1997). Value at Risk: The New Benchmark for
controlling M arket Risk. Chicago: Irwin Professional Pub.
[28] Gajurati. (2010). Basic Econometrics (4 ed.). Pearson-
Addison Wiley.
[29] Box, G. M ., & George, L. E. (1987). On a measure of a Lack
of Fit in Time Series M odels. Biometrika(65), 297-303.
[30] Harvey, A. C., & Jaeger, A. (1993). Detrending, Stylized
Facts and the Business Cycle. Journal of Applied
Econometrics, 231-247.
[31] Takle. (2003). M odelling Volatility in Time Series data, M sc
thesis,. University of Kwa-Zulu Natal.
[32] Gujarati, D. N. (2004). Basic Econometrics. 4th ed. Tata
M cGraw-HillPublishing Company Limited, New Delhi,
India.
[33] Bollerslev, T., Chou, R. Y. and Kroner, K. F. (1992) ARCH
modelling in finance, Journal of Econometrics, Vol 52, pp.
5-59