INDIAN INSTITUTE OF MANAGEMENT CALCUTTA

WORKING PAPER SERIES

WPS No. 588/ March 2006

Modeling daily volatility of the Indian stock market using intra-day data

by

Ashok Banerjee & Sahadeb Sarkar

ashok@iimcal.ac.in sahadeb@iimcal.ac.in

Professors, IIM Calcutta, Diamond Harbour Road, Joka P.O., Kolkata 700 104 India
Full Title: Modeling daily volatility of the Indian stock market using intra-day data

Authors:

Ashok Banerjee
Professor (Finance & Control)
Indian Institute of Management Calcutta
Diamond Harbour Road
Joka
Kolkata-700 104
INDIA
Email: ashok@iimcal.ac.in
Phone Number: 91-33-2467-8300
Fax Number: 91-33-2467-8307

Sahadeb Sarkar
Professor (Operations Management)
Indian Institute of Management Calcutta
Diamond Harbour Road
Joka
Kolkata-700 104
INDIA
Email: sahadeb@iimcal.ac.in

2
 ABSTRACT

The prediction of volatility in financial markets has been of immense interest among

financial econometricians. Using data collected at five-minute intervals, the present

paper attempts to model the volatility in the daily return of a very popular stock market

in India, called the National Stock Exchange. This paper shows that the Indian stock

market experiences volatility clustering and hence GARCH-type models predict the

market volatility better than simple volatility models, like historical average, moving

average etc. It is also observed that the asymmetric GARCH models provide better fit

than the symmetric GARCH model, confirming the presence of leverage effect. Finally,

our results show that the change in volume of trade in the market directly affects the

volatility of asset returns. Further, the presence of FII in the Indian stock market does

not appear to increase the overall market volatility. These findings have profound

implications for the market regulator.

KEY WORDS: Volatility; GARCH models; MDH; half-life of volatility shock;

leverage.

3
1. INTRODUCTION

The magnitude of fluctuations in the return of an asset is called its volatility. The

prediction of volatility in financial markets has been of immense interest among

financial econometricians. This interest is further rekindled by Bollerslev et al. (1994)

when they established that financial asset return volatilities are highly predictable. It is

true that unlike prices, volatilities are not directly observable in the market, and it can

only be estimated in the context of a model. However, Andersen et al. (2001) concluded

that by sampling intra-day returns sufficiently frequently, the realized volatility

(measured by simply summing intra-day squared returns) can be treated as the observed

volatility. This observation has profound implication for financial markets (Brooks

1998) in that (a) the realized volatility provides a better measure of total risk (value at

risk) of financial assets, and (b) it can lead to better pricing of various traded options.

It has been observed in early sixties of the last century (Mandelbrot, 1963) that stock

market volatility exhibits clustering, where periods of large returns are followed by

periods of small returns. Later popular models of volatility clustering were developed by

Engle (1982) and Bollerslev (1986). The autoregressive conditional heteroskedastic

(ARCH) models (Engle, 1982) and generalized ARCH (GARCH) models (Bollerslev,

1986) have been extensively used in capturing volatility clusters in financial time series

(Bollerslev et al., 1992). Using data on developed markets, several empirical studies

(Akgiray, 1989; West et al., 1993) have confirmed the superiority of GARCH-type

models in volatility predictions over models such as the naïve historical average,

moving average and exponentially weighted moving average (EWMA). GARCH

models can replicate the fat tails observed in many high frequency financial asset return

series, where large changes occur more often than a normal distribution would imply.

4
Financial markets also demonstrate that volatility is higher in a falling market than it is

in a rising market. This asymmetry or leverage effect was first documented by Black

(1976) and Christie (1982). Three popular GARCH formulations for describing this

asymmetry are Power GARCH model (Ding et al., 1993), Threshold GARCH model

(Glosten et al., 1993) and Exponential GARCH model (Nelson, 1991). We have used all

these three models in the present study.

Empirical results also show that augmenting GARCH models with information like

market volume or number of trades may lead to modest improvement in forecasting

volatility (Brooks, 1998; Jones et al, 1994). The association between stock return

volatility and trading volume was analyzed by many researchers (Karpoff, 1987). The

initial research on price-volume relation can be attributed to Osborne (1959) who

attempted to model stock price change as a diffusion process with the variance

dependent on the number of transactions. Later research on the empirical relationship

between daily price volatility and daily trading volume was based on Clark’s (1973)

mixture of distribution hypothesis (MDH). The essence of MDH is that if the stock

return follow a random walk and if the number of steps depends positively on the

number of information events, then stock return volatility over a given period should

increase with the number of information events (e.g., trading volume) in that period. In

a recent study on individual stocks in the Chinese stock market, Wang et al. (2005)

showed that inclusion of trading volume in the GARCH specification reduces the

persistence of the conditional variance dramatically, and the volume effect is positive

and statistically significant in all the cases for individual stocks. However, another study

on the Austrian stock market (Mestel et al., 2003) found that the knowledge of trading

volume did not improve short-run return forecasts. Most of the studies on the

5
relationship between return volatility and trading volume have used volume levels.

However, few studies (Ying, 1966; Mestel et al., 2003) have used a relative measure

(turnover ratio) for volume. We have used a relative measure of volume in the present

paper. The present paper also uses trading volume of foreign institutional investors as

another exogenous variable.

There have been a few attempts to model and forecast stock return volatilities in

emerging markets. For example, Gokcan (2000) finds that for emerging stock markets

the GARCH(1,1) model performs better in predicting volatility of time series data. In

another market specific study, Yu (2002) observes that the stochastic volatility model

provides better volatility measure than ARCH-type models. A few studies were

conducted (e.g., Varma, 1999 and 2002; Kiran Kumar and Mukhopadhyay, 2002; Raju

and Ghosh, 2004; Pandey, 2005; Karmakar, 2005) on modeling stock return volatility in

the world’s largest democracy, India. Varma (1999) showed, using daily data from

1990-1998 of an Indian stock index (Nifty), that GARCH (1, 1) with generalized error

distribution performs better than the EWMA model of volatility. In a later study,

Pandey (2005) showed that extreme value estimators perform better than the conditional

volatility models. In another recent study, Karmakar (2005) used conditional volatility

models to estimate volatility of fifty individual stocks and observed that the GARCH

(1,1) model provides reasonably good forecast. However, none of the studies, based on

Indian stock markets, attempted to fit a mean equation for the stock return series before

modeling volatility of stock returns. The present paper determines the best-fit mean

model for the index return, which is then used in GARCH model specifications.

The present paper attempts to model the daily volatility, using high frequency intra-

day data, in the stock index return of a very popular stock market in India, called the

6
National Stock exchange (NSE). The financial markets in India have gone through

various stages of liberalization that has increased its degree of integration with the world

markets. Some instances of new policy reforms introduced in the Indian stock markets

include introduction of trading in index futures in June 2000, trading in index options in

June 2001, trading in options on individual securities in July 2001, introduction of VaR

(value at risk)-based margin, and introduction of the T+2 settlement system from April,

2003. After implementation of such reforms, the Indian securities market has now

become comparable with securities markets of developed and other emerging

economies. In fact, India has a turnover ratio, which is comparable with that of other

developed markets and also one of the highest in the emerging markets (NSE, 2005).

These developments in the Indian securities market have drawn attention of researchers

from across the globe to look at the price behaviour of the Indian securities market. The

daily gross activity (purchase and sales) of the Foreign Institutional Investors (FII) in

the Indian stock market has increased almost three-fold in three and half years, namely,

from Indian Rupees (Rs.) 6 billion in October 2000 to Rs. 17 billion by the end of

January 2004. The increasing interests of foreign investors in the Indian market call for

greater research on various properties of this market. The present paper examines the

evidence of stylized properties in the Indian stock market. The findings of this study

would greatly help fund managers have a better understanding of the Indian stock

market volatility.

This paper also attempts to forecast volatility using various competing models and

measure their predictive power using standard evaluation measures. Results show that

the Indian stock market experiences volatility clustering and GARCH-type models

predict the market volatility far better than the simple volatility models, like historical

7
average, moving average, EWMA etc. It is also observed that the asymmetric variants

of GARCH models give better fit than the symmetric GARCH model, confirming the

presence of the leverage effect. Finally, our results show that the change in volume of

trade in the market directly affects the volatility of asset returns. However, we did not

find similar convincing evidence in support of the FII investment affecting market

volatility.

The remainder of the paper is organized as follows. The next section discusses the

data structure and methodology used in the study. The third section briefly describes

various competing volatility models. The fourth section presents the results and

interpretations and the final section summarizes the findings and indicates scope for

further research.

2. DATA AND METHODOLOGY

The Indian capital market has witnessed significant regulatory changes since 1992 with

the creation of an independent capital market regulator, the Securities and Exchange

Board of India (SEBI). Subsequent changes (e.g., screen based trading, derivatives

trading, trading cycles etc.) have further developed the market and brought it in line

with international capital markets. Presently only two exchanges in India, the NSE and

the BSE (Stock Exchange, Mumbai) provide trading in the security derivatives. We

have used the most popular index of the National Stock Exchange in India, called S&P

CNX NIFTY (Nifty) to model the volatility in the Indian capital market. The Nifty, a

market capitalization weighted index, is an index of 50 scrips accounting for 23 sectors

of the Indian economy. Although the BSE, the oldest stock exchange in India (and also

in Asia), has been in existence for more than 100 years, the reason for choosing the

8
Nifty is its increasing popularity. The BSE was established in 1875 and the NSE started

its capital market (equities) segment in late 1994. However, the market capitalization of

this segment in March 2000, before the launch of stock index futures contracts, was

Rs.10204 billion as against the BSE’s Rs. 9128 billion. The average daily turnover

during May 2005 was around Rs. 40 billion in the NSE as against Rs. 20 billion in the

BSE.

Our sample contains a total of 60,631 data points consisting of the Nifty values at

five-minute intervals from 01 June 2000 through 30 January 2004. The choice of the

period is guided by the fact that a lot of policy reform initiatives in the Indian securities

market have started during this period. For example, trading on index futures was

allowed in India since June 2000. High-frequency data have now become a popular

experimental bench for analyzing financial markets (Dacorogna et al., 2001). High

frequency data are direct information from the market. Hence, instead of using the daily

closing value of the index, this paper uses the directly observable data. It cannot be

denied that very high frequency data have microstructure effect (e.g., how the data are

transmitted and recorded in the data base). In order to avoid serious microstructure

biases and at the same time reduce the measurement error due to data generation at low

frequency; we have used data at regularly spaced five-minute intervals (Andersen et al.,

2001). The present paper uses index values rather than stock prices and thus there are no

bid and ask prices. We have used the last quoted value of the index at five-minute

intervals. The daily index return is estimated using these five-minute interval values.

A large part of the data set, from June 01, 2000 through December 16, 2003, is used

to model volatility using various established volatility models. The remaining data set,

9
from December 17, 2003 through 30 January, 2004, is used to test the efficacy of

various models using one-day-ahead volatility forecasts.

Let rti , i = 1, L , mt , denote log of price relatives at an intra-day time-point i on day t,

where mt is the number of return observations obtained by using prices m times per

day. Then daily return on day t is calculated as rt = ∑i =t1 rti . Following Andersen et al.
m

(2001), we define the daily realized volatility ( Vt 2 ) as the sum of squares of returns

collected at 5-minute intervals:

mt
Vt2 = ∑ rti2 (1)
i =1

For example, on June 01, 2000, the first 5-minute price observation was at 9.55AM and

the trading on that day ended at 3.25PM, giving us m = 67 prices and hence mt = 66

return observations. So, the realized volatility for day 1 is computed as V12 = ∑i =1 r12i .
66

Let Vt denote the positive square root of Vt2 . The realized volatility represents a natural

approach to measure actual (ex post) realized return variation over a given period of

time (Andersen et al., 2005).

10
 Fig.1:Daily Realized Volatility

0.008
0.006
0.004 Real-vol
0.002
0
Date
20001122
20010518
20011109
20020509
20021031
20030505
20031025
We have computed 913 daily return ( rt ) and realized volatility ( Vt2 ) values using

the 5-minute interval intra-day data. The first 883 daily return points are used to model

Nifty volatilities and the remaining 30 return points are used for volatility prediction

using various models. The realized volatility series (Fig.1) shows that the volatility in

the Nifty has significantly reduced over a period of time. One of the reasons for the

decline in market volatility could be the launch of stock index futures in June 2000.

However, some sharp spikes can be seen during 2001. The maximum daily volatility

was around 0.00576, which happened in March 2001. India witnessed multi-billion

rupees stock market scam in March 2001 which led to a freeze on the flagship scheme

of India’s largest mutual fund (Unit Trust of India) in June 2001. The market witnessed

a 15% decline in value in just one month! This aberration could prompt any researcher

(Alexander, 2001) to remove outliers and then model volatility. However, it cannot be

denied that stock market scams are results of systemic failure and hence these may

recur. Therefore, any robust volatility model should be able to capture this phenomenon

and hence we have not removed these extreme observations from our sample.

11
 In order to test whether the volume of trade influences the volatility of the Indian

stock market, we have included traded volume in the variance equation of competing

GARCH models. The volume X 1t , at time t, is defined as the change in the daily traded

volume:

X 1t = (vt − vt −1 ) / vt −1 ,

where v = traded volume measured in Rs. billion. Similarly, we have used FIIs’

activities in the Indian stock market to capture their influence, if any, on the market

volatility. The FIIs have, in the recent past, shown tremendous interest in the Indian

stock markets. These investors regularly participate in the Indian capital market by

buying and selling large volume of stocks. The FIIs are currently net buyers in the

Indian stock market on most of the days. We have collected the daily purchase and sale

volume of FIIs in the stock market from the official website of the capital market

regulator in India (SEBI). In order to capture the activity of the FIIs in India, we have

considered the daily gross trade (sum of purchase and sale volume). We did not have

data on the FII trade for all 883 days that we have considered for GARCH models

without exogenous variable(s). We have used a total of 798 data points on FII trade data

from October, 2000 through mid-December, 2003. In other words, while the other

GARCH models (Table 4) use 883 data points, the models with FII trade (as an

exogenous variable in the variance equation) use 798 data points. In the conditional

variance equation of competing GARCH models, the FIIs’ daily gross trade

volume X 2 t , at time t, is defined as follows:

X 2 t = log(κ t )

where κ t = sum of purchase and sales volume. To compare predictive power of the

competing volatility models, we have predicted daily conditional volatility for thirty

12
days forward. Under each model, all the associated parameters are estimated each time

the day-ahead forecast is made. The first day-ahead forecast is made using parameter

estimates from Table 4. The sample is then rolled forward by removing the first

observation of the sample and adding one to the end to get the next day-ahead forecast.

This process is repeated to calculate the subsequent day-ahead forecasts. Such daily

volatility forecasts are compared with the realized volatility for the designated thirty

days.

Let σˆ t2 denote volatility forecast. One can assess the accuracy of the daily volatility

forecasts under a model by considering the simple linear regressions (Andersen et al.,

2005) of yt on σˆ t2 :

y t = a + bσˆ t2 + ε t

where y t = Vt 2 or rt2 , and then computing the coefficient of determination R 2 . The

model with the highest R 2 value may be treated as the best model for predicting yt. On

the other hand, if the R 2 value turns out to be generally higher for one choice of yt, then

that choice ( Vt 2 or rt2 ) may be considered to be a better measure of observed volatility.

In addition to the regression-based framework, we have used the standard measures

of predictive power of a model − the root mean square error (RMSE), the mean absolute

error (MAE), and the Theil-U statistic. They are defined as follows:

K
1
K ∑ (σˆ k2 − Vk2 ) 2
∑k =1 (σˆ k2 − Vk2 ) 2 , ∑k =1| σˆ k2 − Vk2 |, k =1
K K
RMSE = 1
K
MAE = K1 Theil − U =
K
1
K ∑ (Vk2−1 − Vk2 ) 2
k =1

3. VOLATILITY MODELS

13
We have fitted various competing volatility models to the daily Nifty series, and

predicted the one-day-ahead volatility of the series for the last thirty days and compared

relative performance of the models in estimating the daily realized volatility.

Let

µ t = E (rt | Ft −1 ) , σ t2 = V (rt | Ft −1 ) = E ( ( rt − µt ) 2 | Ft −1 ) (2)

where Ft −1 is the information made available up to time (t−1). The return series { rt }

may be either serially uncorrelated or may have minor lower order serial correlations,

but it may yet be dependent. Volatility models are used to capture such dependence in a

return series. Generally, the conditional mean µ t of such return series { rt } can be

modelled using a simple time series model such as a stationary ARMA(p,q) model, i.e.,

µ t = φ0 + ∑i =1 φi rt −i − ∑ j =1 θ j at − j ,
p q
rt = µ t + at , (3)

where the shock (or mean-corrected return) at represents a white noise series with

mean zero and variance σ a2 , and p, q are non-negative integers. From equations (2)

and (3) we have

σ t2 = V (rt | Ft −1 ) = V ( a t | Ft −1 ) (4)

The equation for µ t is called the mean equation for the rt , and that for σ t2 its

volatility (or conditional variance) equation. A GARCH model expresses the volatility

evolution through a simple parametric function. To compute the predicted value σˆ T2 +1 of

σ T2 +1 at t= (T+1), parameters are first estimated by fitting an appropriate mean and

volatility equations jointly to the data upto time point t=T and then an appropriate

prediction formula used. The most commonly employed method of estimation is the

maximum likelihood estimation under certain regularity conditions and a detailed

14
discussion on this is given in Andersen et al. (2005). Exogenous explanatory variables

such as the change in daily traded volume (X1) and the FII investment (X2) may be

included in the volatility equation to improve prediction accuracy. We now discuss the

models briefly.

Random walk: The random walk model is the simplest of the models considered

and it is given by: σ t2 = σ t2−1 + ε t , where ε t iis a white noise series.

Historical Average: The historical average model (Yu, 2002) is given by:

t −1
∑i=1σ i2 .
1
σ t2 =
t −1

EWMA: The exponentially weighted moving average (EWMA) model is given by:

σ t2 = λ (σ t2−1 ) + (1 − λ )(Vt2−1 ) , where 0 ≤ λ ≤ 1 is estimated by maximizing an appropriate

likelihood function (Hull, 2003).

Next we briefly discuss the popular GARCH-type models. The GARCH model

focuses on the time-varying variance of the conditional distributions of returns. The

GARCH and related models capture the stylized fact that financial market volatility

appears in clusters, where tranquil periods of small returns are interspersed with volatile

periods of large returns. The feature of volatility clustering or volatility persistence went

unrecognized in traditional volatility models such as the historical average, which

assumed market volatility to be constant.

GARCH: The volatility model for the rt or at is said to follow a GARCH(m, s)

model (Bollerslev, 1986; Bollerslev et al., 1994) if

∑i=1α i at2−i + ∑ j =1 β jσ t2− j ,

m s
at = σ t ε t , σ t2 = α 0 + (5)

where α 0 > 0, α i ≥ 0, β j ≥ 0,

15
 ∑i=i
max( m, s )
(α i + β i ) < 1, with α i = 0 for i>m and β j = 0 for j >s,

and {ε t } is a sequence of iid random variables with mean 0 and variance 1, which is

often assumed to have a standard normal or standardized Student-t distribution.

An exogenous explanatory variable Xk may be included in the GARCH model. For

example, the GARCH(1,1) model can be augmented as

at = σ t ε t , σ t2 = α 0 + α1at2−1 + β1σ t2−1 + η1 X kt (6)

Moving average based on half-life of the volatility shock: The high or low

persistence in volatility is generally captured in the GARCH coefficient(s) of a

stationary GARCH model. For a stationary GARCH model the volatility mean-reverts

to its long-run level, at the rate given by the sum of ARCH and GARCH coefficients,

which is generally close to one for a financial time series. The average number of time

periods for the volatility to revert to its long run mean level is measured by the half life

of the volatility shock and it is used to forecast the Nifty series volatility on a moving

average basis.

A covariance stationary time series { y t } has an infinite order moving average

representation of the form (Fuller, 1996)

∞ ∞
yt = µ + ∑ψ i at −i , ψ 0 = 1, ∑ψ i2 < ∞ .
i =0 i =0

The plot of the ψ i against i is called the impulse response function (IRF). The decay

rate of IRF is sometimes reported as a half-life, denoted by Lhalf , which is the lag at

which the IRF reaches 12 . Calculation of half-life of volatility shock for a stationary

16
GARCH(1,1) process σ t2 = α 0 + α1at2−1 + β1σ t2−1 is done as follows. For this model the

mean-reverting form is given by

(at2 − σ 2 ) = (α1 + β1 )(at2−1 − σ 2 ) + ut − β1ut −1 ,

where σ 2 = α 0 /(1 − α1 − β1 ) is the unconditional long-run level of volatility, and

u t = (at2 − σ t2 ) is the volatility shock. The half-life of the volatility is given by the

formula (Zivot and Wang, 2002): Lhalf = ln( 12 ) / ln(α1 + β1 ) .

EGARCH : A stylized fact of financial volatility is that bad news (negative shocks)

tend to have a larger impact on volatility than good news (positive shocks). However, a

GARCH model fails to respond differently to positive and negative shocks. Nelson

(1991) developed the exponential GARCH (E-GARCH) model, which corrects this

weakness by appropriately weighting innovation ε t .

g (ε t ) = θε t + γ [| ε t | − E (| ε t |)] , (7)

where θ and γ are parameters. An EGARCH(m, s) model is of the form

1 + β1 B + L + β s B s
at = σ t ε t , ln(σ t2 ) = α 0 + ( ) g(ε t -1 ) , (8)
1 − α1 B − L − α s B m

where α0 is a constant, B is the back-shift or lag operator such that B ( g (ε t )) = g (ε t −1 ) ,

and (1 + β1 B + L + β s B s ) , (1 − α1 B − L − α s B m ) are polynomials in B. The EGARCH

model uses logged conditional variance to relax the constraint of model coefficients

being nonnegative, and through g (ε t ) can respond asymmetrically to positive and

negative lagged values of at. An alternative representation for EGARCH(m, s) model is

given by

∑i=1α i (| ε t −i | +γ iε t −i ) + ∑ j =1 β j ln σ t2− j ,
m s
at = σ t ε t , and ln σ t2 = α 0 +

17
where γi’s are the leverage parameters.

TGARCH: Another GARCH variant which is capable of responding

asymmetrically to positive and negative innovations is the threshold GARCH

(TGARCH) model (Zakoian, 1991; Glosten et al., 1993). The TGARCH(m,s) model has

the form

σ t2 = α 0 + ∑ ∑i=1γ i (S t −i at2−i ) + ∑ j =1 β jσ t2− j ,

m m s
at = σ t ε t , α a2 + (9)
i =1 i t −i

where

⎧1 if a t −i < 0
S t −i = ⎨
⎩0 if a t −i ≥ 0

PGARCH: A third GARCH variant capable of responding asymmetrically to

positive and negative innovations is the power GARCH (PGARCH) model (Ding et al,

1993). The PGARCH(m,s) model has the following form

σ td = α 0 + ∑ ∑ j =1 β jσ td− j
m s
at = σ t ε t , α (| at −i | + γ i a t −i ) d + (10)
i =1 i

where the parameter d is positive and γ i ’s denote the coefficients of leverage effects.

4. RESULTS AND ANALYSIS

Table 1 shows the descriptive statistics of the daily realized volatility and daily return

for the entire sample. The sample kurtosis of 70.08 and skewness of 6.92 indicate that

the distribution of daily volatility is not a normal distribution. This is further supported

by the normality test statistics in Table 2. The constructed daily return series has an

insignificant mean (Table 1) of around 6.30% per annum, using a standard 250-trading

days in a year. The negative skewness of the return series usually indicates that there is

at least one very large negative return in the data, which is what we observe in our case

18
with the minimum daily return being −6.15%. The existence of excess kurtosis also

indicates that the daily return series is not normal.

Table 1: Descriptive Statistics and test for nonstationarity of the daily return and
unconditional volatility series

Series Mean Median Maximum Minimum Skewness Excess ADF

Kurtosis
Return 0.0252% 0.1045% 5.2994% -6.1585% -0.04670 1.6596 -12.01*
Volatility 0.0222% 0.0126% 0.5757% 0.0014% 6.9179 70.0842 -9.623*
*Significant at 1% level.

Table 2: Normality and Autocorrelation Test for Unconditional volatility series

Test Model Test-statistic p-value
Normality Jarque-Bera 192038.4 0.0000
Shapiro-Wilks 0.4471 0.0000
Autocorrelation Ljung-Box 302.6835 0.0000

To test whether the daily return and realized volatility series is stationary, the

augmented Dickey-Fuller (ADF) statistic is calculated on the entire sample and the

results (Table 1) fail to accept the unit root null hypothesis at 1% level. The

autocorrelation test (Ljung-Box) for the realized volatility series shows that (Table 2)

the first-order autocorrelation of volatility is quite high, although we observed the

higher-order autocorrelations to be generally diminishing (the numerical results not

presented here, for brevity). However, autocorrelations up to order four are statistically

significantly different from zero. This demonstrates the evidence of volatility clustering

and hence any autoregressive heterskedastic volatility model should be a better fit.

In order to apply the GARCH-type models, one needs to first identify the best model

for the mean equation and then fit a model for variance equation. We have applied

various models on daily returns to identify the best-fit mean model. We have used first

883 daily returns to model the return series. A random walk mean model without a drift

shows that the return series is non-normal (Table 3, Panel A). We have also used a

19
random walk model with a drift. However, the intercept coefficient is not different from

zero. In order to test whether the return series is autocorrelated, we have used a higher-

order autoregressive model, namely, AR(25):

rt = φ 0 + φ1 rt − 1 + φ 2 rt − 2 + L + φ 25 rt − 25 + at

We observed that the Nifty series is largely uncorrelated. However, we find, from

the AIC (Akaike Information Criterion) values that a moving average model, namely

[MA (1)], is a better fit. It may be noted that neither the autoregressive nor the moving

average models contain an intercept term.

Table 3: Statistics for Mean Model

Panel A: Normality and Autocorrelation Tests for in-sample Mean Return Series
Test Model Test-statistic p-value
Normality Jarque-Bera 155.0179 0.0000
Shapiro-Wilks 0.9684 0.0000
Autocorrelation Ljung-Box 27.0563 0.3531

Panel B: Test Results of Various Mean Models

Model Coefficient t-statistic p-value AIC
Random walk with 0.0002 0.4561 0.6484
drift
AR(25) -4843.07
AR(1): AR term 0.0743 2.2135 0.0271 -5009.63
MA(1): MA-term -0.0755 -2.2499 0.0245 -5013.73
ARMA(1,1) -5009.21
AR-term -0.2527 -0.6122 0.5404
MA-term -0.3251 -0.8059 0.4203

After modeling the mean of the Nifty return series, the Nifty volatility series is

modeled using various competing GARCH models as described above. The parameter

estimates of various GARCH models are given in Table 4. The sizes of arch and garch

parameters determine the short-run dynamics of the resulting volatility time series.

Large garch coefficient normally indicates persistence of volatility and large arch

coefficient implies that volatility is less persistent and more ‘spiky’. The sum of arch

20
and garch coefficients in GARCH (1,1) model for the Nifty is less than one indicating

that the variance process is stationary. In fact, the garch coefficient, though significant,

is relatively low and arch coefficient high. This indicates that volatilities do not ‘persist’

for long.

Table 4: Parameter Estimates of Various GARCH Models

Panel A: Parameter Estimates
Model Coefficient Value t-value p-value AIC BIC
GARCH (1,1) -5121.92 -5102.78
MA 0.1005 2.595 0.0048
Intercept 0.0000 4.589 0.0000
ARCH 0.2338 6.306 0.0000
GARCH 0.6053 10.199 0.0000
E-GARCH -5136.21 -5112.30
(1,1)
MA 0.1146 3.053 0.0012
Intercept -2.2207 -5.594 0.0000
ARCH 0.3852 7.838 0.0000
GARCH 0.7780 17.905 0.0000
Gamma( γ ) -0.3717 -5.293 0.0000
T-GARCH -5135.49 -5111.58
(1,1)
MA 0.1079 2.749 0.0030
Intercept 0.0004 4.906 0.0000
ARCH 0.1122 4.063 0.0000
GARCH 0.5740 8.975 0.0000
Gamma( γ ) 0.2528 3.935 0.0000
P-GARCH -5137.25 -5113.33
(1,1)
MA 0.1052 2.92 0.0018
Intercept 0.0030 5.029 0.0000
ARCH 0.2065 7.13 0.0000
GARCH 0.6172 10.51 0.0000
Gamma( γ ) -0.4449 -5.906 0.0000
GARCH (1,1)- -5133.13 -5109.21
with volume
MA 0.0954 2.388 0.0086
Intercept 0.0000 4.946 0.0000
ARCH 0.2524 6.918 0.0000
GARCH 0.5635 10.54 0.0000
volume 0.0001 5.02 0.0000
E- -5137.85 -5109.15
GARCH(1,1)- MA 0.1090 2.873 0.0021
with volume Intercept -2.4359 -6.148 0.0000
ARCH 0.4012 7.831 0.0000
GARCH 0.7562 17.449 0.0000
Lev -0.3820 -5.441 0.0000
volume 0.2345 2.339 0.0098

21
Table 4 : Panel A (Contd.)
Model Coefficient Value t-value p-value AIC BIC
T- -5149.39 -5120.69
GARCH(1,1)- MA 0.1063 2.622 0.0044
with volume Intercept 0.0000 5.257 0.0000
ARCH 0.1138 4.457 0.0000
GARCH 0.5405 9.279 0.0000
Gamma( γ ) 0.2753 4.339 0.0000
volume 0.0001 4.548 0.0000
P- -5147.05 -5118.35
GARCH(1,1)- MA 0.0932 2.479 0.0067
with volume Intercept 0.0033 6.022 0.0000
ARCH 0.2340 7.546 0.0000
GARCH 0.5565 10.226 0.0000
Gamma( γ ) -0.4550 -6.485 0.0000
volume 0.0044 4.041 0.0000
GARCH(1,1)- -4667.29 -4643.88
with FII MA 0.1030 2.556 0.0054
Intercept -0.0001 -1.581 0.0571
ARCH 0.1873 5.467 0.0000
GARCH 0.6637 11.275 0.0000
FII 0.0000 2.100 0.018
E- -4674.26 -4646.17
GARCH(1,1)- MA 0.1165 3.035 0.0012
with FII Intercept -2.8782 -4.491 0.0000
ARCH 0.3263 6.765 0.0000
GARCH 0.7827 16.155 0.0000
Leverage -0.4215 -4.267 0.0000
FII 0.1238 2.428 0.0077
T- -4677.62 -4649.52
GARCH(1,1)-
with FII
MA 0.1098 2.749 0.0030
Intercept -0.0001 -2.345 0.0096
ARCH 0.0771 3.117 0.0009
GARCH 0.6576 10.570 0.0000
Gamma( γ ) 0.2051 3.402 0.0003
FII 0.0000 2.910 0.0018
P- -4677.34 -4649.24
GARCH(1,1)-
with FII
MA 0.1188 3.1600 0.0008
Intercept -0.0026 -1.6820 0.0464
ARCH 0.1721 6.0700 0.0000
GARCH 0.6543 10.4640 0.0000
Gamma( γ ) -0.5156 -4.8960 0.0000
FII 0.0009 3.0050 0.0014

22
Table 4: Panel B: Normality and Autocorrelation Tests
Model JB- p-value SW- p-value LB- p-value
statistic statistic statistic*
GARCH(1,1) 38.48 0.0000 0.9849 0.2704 13.38 0.3420
E- 35.27 0.0000 0.9864 0.5093 17.93 0.1177
GARCH(1,1)
T- 35.56 0.0000 0.9848 0.2643 17.4 0.1351
GARCH(1,1)
P- 34.74 0.0000 0.9866 0.5498 19.46 0.07801
GARCH(1,1)
GARCH(1,1)- 40.94 0.0000 0.982 0.0325 14.69 0.2586
with volume
E- 37.32 0.0000 0.9859 0.4198 19.55 0.07613
GARCH(1,1)-
with volume
T- 36.86 0.0000 0.9829 0.07067 19.21 0.0836
GARCH(1,1)-
with volume
P- 38.67 0.0000 0.9845 0.2178 23.25 0.0256
GARCH(1,1)-
with volume
GARCH(1,1)- 28.72 0.0000 0.9874 0.7006 11.61 0.4772
with FII
E- 29.33 0.0000 0.9875 0.711 16.24 0.1805
GARCH(1,1)-
with FII
T- 30.71 0.0000 0.9868 0.6064 15.12 0.2351
GARCH(1,1)-
with FII
P- 28.64 0.0000 0.9879 0.7652 16.91 0.1529
GARCH(1,1)-
with FII
Note: JB stands for Jarque-Bera; SW for Shapiro-Wilk and LB stands for Ljung-Box.
* LB statistic is for squared standardized residuals.

In symmetric GARCH models, the signs of the residuals (estimated shocks) have no

impact as we consider only squared residuals ( at2−i ) in the GARCH equation. However,

in equity markets it is commonly observed (Black, 1976) that volatility is not

symmetric. The negative skewness in the return series (Table 1) indicates the existence

of leverage effect and hence any asymmetric GARCH model would be capable of

capturing such effect. The negative (and significant) coefficient of leverage in E-

GARCH confirms the leverage effect. The other versions of asymmetric GARCH model

23
(TGARCH and PGARCH) further support the leverage effect in the Nifty volatility

series. The AIC or BIC (Bayesian Information Criterion) value is minimum for the

PGARCH model without any exogenous explanatory variable and hence the asymmetric

PGARCH models the Nifty volatility best. However, when GARCH models are

augmented with exogenous variables in the variance equation, TGARCH (1,1) with

volume as the exogenous variable has given a better fit than PGARCH(1,1). It may be

noted that all versions of the GARCH model use only first-order autoregressive variance

and squared error terms. We have tried with higher-order GARCH models, but the

results did not improve (based on AIC or BIC values). The normality (Shapiro-Wilk)

and autocorrelation (Ljung-Box) tests show that the conditional variances (Table 4,

Panel B) are normal and uncorrelated. In other words, the effect of hetersoskedasticity

in the residuals has been controlled.

The coefficient of volume in various GARCH models is significantly different from

zero. The positive coefficient of volume indicates that higher the change in daily traded

volume, larger is the variance. In other words, the change in daily traded volume

impacts the market volatility. With the growth in the volume of trade in the Indian

bourses, market volatility is going to only increase with time. This finding has a

message for the market regulator – market surveillance and risk management practices

should be further strengthened to take care of greater market uncertainty associated with

an increase in the volume of trade.

We have also tried to test whether the flow of the FII investment in the Indian

capital market has increased the market volatility. It may be noted that when the FII

trade is used as an exogenous variable in GARCH-type models, the volume is not used.

The results show that although the level of FII trade has an impact on the market

24
volatility, the AIC or BIC values indicate that the volume is more significant exogenous

variable than the FII trade. These findings would dispel the fear, in the minds of the

capital market regulator, that FII participation in the Indian capital market would make

the market more uncertain. Instead, FIIs with better skill, vast experience and

investment acumen would steady the market volatility. The FIIs were net buyers in the

Indian stock markets even during most market crashes in recent years. One may,

however, argue that the total volume of trade subsumes the FII trade and hence FII trade

does impact market volatility. But a comparison of competing GARCH models with the

volume and FII trade appearing as an exogenous variable in respective models reveals

that the information about FII investments fails to improve volatility predictions. The

AIC or BIC values for FII-trade dependent GARCH models deteriorate significantly,

indicating that the level of the FII trade could weakly explain the market volatility.

Once it is identified that the volatility in the Nifty can be best modeled with an

asymmetric GARCH model, we have tried to test the robustness of competing GARCH

models in predicting market volatility. We have made one-day-ahead forecast of

volatility for thirty days (from December 17, 2003 through January 30, 2004). We have

also attempted to compare conditional volatility predictions based on GARCH models

with other competing models of volatility, like random walk, historical average, moving

average and exponentially weighted moving average (EWMA). Yu (2002) used a five-

year and a ten-year moving average in forecasting monthly volatility. We have followed

a different approach in using moving average method for conditional volatility forecast

in this paper. It is necessary to capture the volatility persistence to decide about the

number of past observations to be included in the moving average. We have used the

concept of half-life of variance to identify the number of past observations to be used

25
for forecasting volatility through the moving average method. The GARCH(1,1) model

in Table 4 (Panel A) reports that the conditional volatility is stationary. Using the arch

coefficient (0.2338) and the garch coefficient (0.6053) of the symmetric GARCH(1,1)

model, the half-life of variance is estimated as 4 days. This implies that volatility would

not persist for a long period. It may be conjectured that frequent regulatory interventions

in the Indian capital markets have not allowed market volatility to persist for long. Thus,

moving average model of volatility prediction uses realized volatility of the immediately

past four days.

Results are presented in Table 5. Results show that (Table 5, Panel A) conditional

volatility models better predict actual volatility. In fact, EGARCH (1,1) with volume as

exogenous variable has the best predictive power. It can also be observed that R 2 has

significantly improved for the same volatility forecast when realized volatility Vt 2 is

used in place of its poor cousin rt2 . This result reiterates that even if a researcher is

interested in predicting volatility over daily horizon, it is always better to use Vt 2 as the

proxy for latent daily volatility (Andersen et al., 2005).

Table 5: Evaluation Measures for volatility prediction

Panel A: Correlation ( R 2 ) of general volatility forecast evaluation regression
Model rt2 Vt 2
Random-walk 0.5006 0.3651
Historical average 0.1413 0.2314
Moving average 0.2170 0.3054
EWMA 0.3341 0.3515
GARCH(1,1) 0.1464 0.3092
E-GARCH(1,1) 0.3979 0.6074
T-GARCH(1,1) 0.3685 0.5667
P-GARCH(1,1) 0.3973 0.5804
GARCH-volume 0.1524 0.3163
E-GARCH-volume 0.3820 0.6189
T-GARCH-volume 0.3797 0.5859
P-GARCH-volume 0.4289 0.6026
GARCH-FII 0.1448 0.3241
E-GARCH-FII 0.3692 0.5469

26
T-GARCH-FII 0.3552 0.5354
P-GARCH-FII 0.4107 0.5466

Panel B: Standard Evaluation Measures

Model RMSE (× 107) MAE (× 102) Theil-U
Random-walk 0.5318 0.5268 1.00
Historical average 0.8378 0.5841 1.57
Moving average 0.5137 0.4507 0.96
EWMA 0.4619 0.4161 0.86
GARCH(1,1) 0.5738 0.4448 1.07
E-GARCH(1,1) 0.3310 0.3757 0.62
T-GARCH(1,1) 0.3292 0.3509 0.61
P-GARCH(1,1) 0.3252 0.3758 0.61
GARCH-volume 0.5686 0.4322 1.06
E-GARCH-volume 0.3377 0.3788 0.63
T-GARCH-volume 0.3204 0.3446 0.60
P-GARCH-volume 0.3144 0.3788 0.59
GARCH-FII 0.5991 0.4763 1.12
E-GARCH-FII 0.7988 0.5969 1.50
T-GARCH-FII 0.5460 0.5284 1.02
P-GARCH-FII 0.6213 0.5682 1.16
Other standard measures (Table 5, Panel B) show the superiority of asymmetric

GARCH models for predicting volatility as compared to historical average, moving

average and EWMA models. The prediction error measures RMSE and Theil-U show

that the forecast error is minimum when the PGARCH (1,1) model containing the

volume as an exogenous variable in the variance equation model is used. However, the

MAE measure of forecast error reports that the TGARCH (1,1) model with volume as

an exogenous variable in variance equation is a better predictor of one-day-ahead

volatility. The Theil-U statistic is a poor evaluator of performance in our case as it treats

the random walk model as a benchmark to compare the forecast performance of other

volatility models. The poor forecasting ability of the random walk model implies that

the market volatility is not a random walk and hence can be conveniently modeled. The

GARCH models with the FII trade as an exogenous variable came out as poor predictors

of volatility. Results in Table 5 confirm two distinct features of the Nifty: (a) the

27
stylized fact of leverage effect in volatility clustering, and (b) the empirical evidence

that greater the volume of trade in the market, larger is the market volatility.

5. CONCLUSION

There have been attempts to model and forecast stock return volatilities in emerging

markets. The present paper attempted to model the volatility in the index returns of the

NSE, using high frequency intra-day data covering a period from June 2000 through

January 2004. This paper has four main findings: (a) existence of volatility clustering in

the Indian stock market; (b) evidence of leverage effect on volatility; (c) the change in

volume of trade positively affecting market volatility; and (d) participation of FIIs in the

Indian stock market not resulting in significant increase in market volatility.

Our results confirm the findings of past researchers (e.g., Akgiray, 1989; West et al,

1993; Brooks, 1998; Gokcan, 2000 etc.) in that GARCH-type models have better

forecasting ability of volatility than naïve models. It was also observed (Figure 1) that

the realized volatility in the NSE has declined over the period.

These findings convey some messages for the market regulator. First, a decline in

the market volatility after June 2000 conjectures that the introduction of derivatives

trading in the Indian stock market has led to reduction in market uncertainties. Second,

market volatilities form clusters and volatilities are not very persistent in India, contrary

to experiences in many countries. Third, the current method of computing market

volatility for estimating VaR-based margins using EWMA needs to be reviewed. Our

results show that asymmetric GARCH models can measure volatility more efficiently

than the EWMA model. Fourth, growth in the trading volume, caused by all players and

not specifically FII-led volume growth, appears to cause market volatility to increase.

28
Fifth, the FIIs’ participation in the Indian stock market has, over a period of time, not

led to a significant increase in the market volatility.

Our study has used intra-day data over a period of about three and half years. A

longer period of study could lead to different results. The paper has not attempted to

travel beyond GARCH-type models to predict the market volatility. The forecasting

performance of various volatility models is gauged using standard evaluation measures.

It was observed (Brooks and Persand, 2003) that relative accuracies of the various

volatility models are highly sensitive to the measure used to evaluate them. While the

parametric methods model the persistent, smoother aspects of volatility, the

nonparametric methods model the highly nonlinear response to large return shocks

(Pagan and Schwert, 1990). We intend to use nonparametric methods in our future

research for modeling volatility.

References:

Akgiray, V. 1989. Conditional heteroskedasticity in time series of stock returns:

evidence and forecasts, Journal of Business 62(1) 55-80.

Andersen, T.G., Bollerslev, T., Christoffersen, P.F., and Diebold, F.X. 2005. Volatility

forecasting, Working Paper 11188, National Bureau of European Research.

Andersen, T.G., Bollerslev, T, Diebold, F.X., and Labys, P. 2001. The distribution of

realized exchange rate volatility, Journal of the American Statistical Association 96

(453) 42-55.

Black, F. 1976. Studies in stock price volatility changes, Proceedings of the 1976

Business Meeting of the Business and Economic Statistics Section, American

Statistical Association, 177-181.

29
Bollerslev, T., Chou, R.Y., and Kroner, K.F. 1992. ARCH modeling in finance: a

selective review of the theory and empirical evidence, Journal of Econometrics, 52,

5-59.

Bollerslev, T., Engle, R.F., and Nelson, D.B. 1994. ARCH Models, in Handbook of

Econometrics, R.F. Engle and D.L. McFadden (eds.) 4, Elsevier Science B.V.

Bollerslev, T. 1986. Generalized autoregressive conditional heteroskedasticity, Journal

of Econometrics 31, 307-327.

Brooks, C., and Persand, G. 2003. Volatility Forecasting for Risk Management, Journal

of Forecasting 22, 1-22.

Brooks, C. 1998. Predicting stock index volatility: can market volume help? Journal of

Forecasting 17, 59-80.

Christie, A.A. 1982. The Stochastic Behavior of Common Stock Variances: Value,

Leverage and Interest Rate Effects, Journal of Financial Economics, 10, 407-432.

Clark, P.K. 1973. A subordinated stochastic process model with finite variance for

speculative prices, Econometrica, 41, 135-155.

Dacorogna, M. M., Gencay, R., Muller, U., Olsen, R.B., and Pictet, O.V. 2001. An

Introduction to High-Frequency Finance Academic Press.

Ding, Z., Granger, C.W., and Engle, R.F. 1993. A long memory property of stock

market returns and a new model, Journal of Empirical Finance, 1, 83-108.

Engle, R.F. 1982. Autoregressive conditional heteroskedasticity with estimates of the

variance of U.K. inflation, Econometrica, 50, 987-1008.

Fuller, W.A. 1996. Introduction to Statistical Time Series, 2nd edn. John Wiley & Sons

Ltd.

30
Glosten, L.R., Jagannathan, R., and Runkle, D.E. 1993. On the relation between the

expected value and the volatility of the nominal excess return on stocks, Journal of

Finance, 48(5), 1779-1801.

Gokcan, S. 2000. Forecasting Volatility of Emerging Stock Markets : Linear various

Non-linear GARCH models, Journal of Forecasting 19, 499-504.

Hull, J.C. 2003. Options, Futures, and other Derivatives, Prentice Hall India.

Jones, C. Kaul, G. and Lipson M. 1994. “Transactions, Volume and Volatility” Review

of Financial Studies 7, 631-651.

Karmakar, M. 2005. Modeling conditional volatility of the Indian stock markets,

Vikalpa, 30(3), 21-37.

Karpoff, J.M. 1987. The relation between price changes and trading volume: A survey,

Journal of Financial and Quantitative Analysis, 22(1), 109-126.

Kiran Kumar, K., and Mukhopadhyay, C. 2002. Equity Market Interlinkages

Transmission of Volatility – A case of US and India. NSE Working Paper Series,

September.

Mandelbrot, B. 1963. The variation of certain speculative prices, Journal of Business

36(3), 394-419.

Mestel, R. , Gurgul, H., and Majdosz, P. 2003. The empirical relationship between stock

returns, return volatility and trading volume on the Austrian stock market,

www.charttricks.com/Resources/ Articles/volume_volatility_Mestel.pdf (accessed

on 21 February 2006).

NSE 2005. Indian Securities Market, A Review (ISMA), National Stock Exchange.

Nelson, D.B. 1991. Conditional heteroskedasticity in asset returns: A new approach

Econometrica 53, 347-370.

31
Osborne, M.F.M. 1959. Brownian motion in the stock market, Operations Research, 7,

145-173.

Pagan, A.R. and Schwert, W.G. 1990. Alternative models for conditional stock

volatility, Journal of Econometrics, 45, 267-290.

Pandey, A. 2005. Volatility models and their performance in Indian capital markets,

Vikalpa, 30(2), 27-46.

Raju, M.T. and Ghosh, A. 2004. Stock Market Volatility – An International

Comparison, working paper Series No.8, Securities and Exchange Board of India,

April.

Varma, J.R. 1999. Value at risk models in the Indian stock market, Working Paper 99-

07-05, Indian Institute of Management Ahmedabad.

---------- 2002. Mispricing of volatility in the Indian index options market, Working

Paper 2002-04-01, Indian Institute of Management Ahmedabad.

Wang, P., Wang, p., and Liu,A. 2005. Stock return volatility and trading volume:

evidence from the Chinese stock market, Journal of Chinese Economic and Business

Studies, 3, 39-54.

West, K.D., Edison H.J., Cho, D. 1993. A utility-based comparison of some models of

exchange rate volatility, Journal of International Economics, 35, 25-45.

Ying, C.C. 1966. Stock market prices and volumes of sales, Econometrica, 34, 676-686.

Yu, J. 2002. Forecasting Volatility in the New Zealand Stock Market, Applied Financial

Economics, 12, 193-202.

Zakoian, J. 1991. Threshold Heteroskedasticity Model, Unpublished Manuscript

INSEE.

Zivot, E., Wang, J. 2003. Modeling Financial Time Series with S-Plus, Springer.

32