Studies in Economics and Finance

Forecasting stock index volatility with GARCH models: international evidence

Prateek Sharma, Vipul _,
Article information:
To cite this document:
Prateek Sharma, Vipul _, (2015) "Forecasting stock index volatility with GARCH models: international
evidence", Studies in Economics and Finance, Vol. 32 Issue: 4, pp.445-463, https://doi.org/10.1108/
SEF-11-2014-0212
Permanent link to this document:
https://doi.org/10.1108/SEF-11-2014-0212
Downloaded on: 23 June 2018, At: 06:41 (PT)
References: this document contains references to 69 other documents.
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

To copy this document: [email protected]

The fulltext of this document has been downloaded 1211 times since 2015*
Users who downloaded this article also downloaded:
(2015),"Volatility behaviour of stock index futures in China: a bivariate GARCH approach",
Studies in Economics and Finance, Vol. 32 Iss 1 pp. 128-154 <a href="https://doi.org/10.1108/
SEF-10-2013-0158">https://doi.org/10.1108/SEF-10-2013-0158</a>
(2015),"Financial constraints, bank concentration and SMEs: evidence from Pakistan",
Studies in Economics and Finance, Vol. 32 Iss 4 pp. 503-524 <a href="https://doi.org/10.1108/
SEF-02-2014-0046">https://doi.org/10.1108/SEF-02-2014-0046</a>

Access to this document was granted through an Emerald subscription provided by emerald-
srm:178665 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald
for Authors service information about how to choose which publication to write for and submission
guidelines are available for all. Please visit www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as
well as providing an extensive range of online products and additional customer resources and
services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the
Committee on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for
digital archive preservation.

*Related content and download information correct at time of download.

 The current issue and full text archive of this journal is available on Emerald Insight at:
www.emeraldinsight.com/1086-7376.htm

Forecasting stock index Stock index

volatility with
volatility with GARCH models: GARCH
international evidence
Prateek Sharma and Vipul 445
Department of Finance and Accounting, Indian Institute of Management,
Received 12 November 2014
Lucknow, India Revised 2 February 2015
Accepted 5 February 2015

Abstract
Purpose – The purpose of this paper is to compare the daily conditional variance forecasts of seven
GARCH-family models. This paper investigates whether the advanced GARCH models outperform the
standard GARCH model in forecasting the variance of stock indices.
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

Design/methodology/approach – Using the daily price observations of 21 stock indices of the

world, this paper forecasts one-step-ahead conditional variance with each forecasting model, for the
period 1 January 2000 to 30 November 2013. The forecasts are then compared using multiple statistical
tests.
Findings – It is found that the standard GARCH model outperforms the more advanced GARCH
models, and provides the best one-step-ahead forecasts of the daily conditional variance. The results are
robust to the choice of performance evaluation criteria, different market conditions and the
data-snooping bias.
Originality/value – This study addresses the data-snooping problem by using an extensive
cross-sectional data set and the superior predictive ability test (Hansen, 2005). Moreover, it covers a
sample period of 13 years, which is relatively long for the volatility forecasting studies. It is one of the
earliest attempts to examine the impact of market conditions on the forecasting performance of GARCH
models. This study allows for a rich choice of parameterization in the GARCH models, and it uses a wide
range of performance evaluation criteria, including statistical loss functions and the Mince-Zarnowitz
regressions (Mincer and Zarnowitz 1969). Therefore, the results are more robust and widely applicable
as compared to the earlier studies.
Keywords GARCH, Volatility, Conditional variance, Forecast, Stock indices
Paper type Research paper

1. Introduction
Volatility forecasting has applications in portfolio selection, option pricing, risk
management and volatility-based trading strategies. The GARCH-family models are
extensively used for modeling and forecasting the volatility of financial assets (Poon
and Granger, 2003). Other popular approaches include the simpler time series models
like the exponentially weighted moving average (EWMA) model and the more
sophisticated stochastic volatility models (Barndorff-Nielsen, 1997; Barndorff-Nielsen
et al., 2002; Harvey and Shephard, 1996; Taylor, 1994, among others). Forecasting the
volatility across a variety of financial markets, Ederington and Guan (2005) found that
the GARCH (1,1) model generally outperforms the EWMA model. Similarly, while there Studies in Economics and Finance
Vol. 32 No. 4, 2015
is a strong argument for modeling volatility as a stochastic process, the out-of-sample pp. 445-463
© Emerald Group Publishing Limited
1086-7376
JEL classification – G10, G15, G17 DOI 10.1108/SEF-11-2014-0212
 SEF performance of stochastic volatility models is generally comparable to that of the
32,4 GARCH models (Fleming and Kirby, 2003; Lehar et al., 2002).
After the standard GARCH model (Bollerslev, 1986) was proposed by Bollerslev in
1986, a variety of more sophisticated GARCH parameterizations have been suggested
for modeling the conditional variance. These advanced models attempt to better capture
the empirically observed stylized facts of the conditional variance process. For instance,
446 the asymmetric effect of the negative return shocks is captured by the EGARCH model
(Nelson, 1991), the GJR model (Glosten et al., 1993), the TGARCH model (Zakoian, 1994)
and the NGARCH model (Higgins and Bera, 1992). More generalized parameterizations,
like the APARCH model (Ding et al., 1993) and the HGARCH model (Hentschel, 1995),
nest a variety of simpler GARCH models.
Nevertheless, the forecasting performance of the more sophisticated GARCH-family
models has not been impressive. Pagan and Schwert (1990) found that the GARCH
model is outperformed by the asymmetric EGARCH model, for the monthly volatility
forecasts of the US stocks. Bali and Demirtas (2008) forecast volatility for the S&P 500
index futures using the GARCH, EGARCH and TGARCH models. They found that
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

EGARCH provides the most accurate forecasts of the future-realized volatility. Cao and
Tsay (1992) found that EGARCH provides the best long-horizon forecasts for the small
stocks, but is outperformed by the other time series models for the large stocks. Alberg
et al. (2008) found that EGARCH provides the best conditional variance forecast for the
stock indices of the Tel Aviv Stock Exchange (TASE). Forecasting the volatility for a
variety of asset classes, Ederington and Guan (2005) found no difference between the
GARCH and EGARCH models. Lee (1991) showed that the out-of-sample forecast
performance of the GARCH models depends on the loss evaluation criteria. Taylor
(2004) compared the forecasting performance of five different GARCH models, and
found that the GJR and IGARCH were the best performing models. Brailsford and Faff
(1996) found that the asymmetric GJR (1,1) model outperforms the others, using the root
mean squared error (RMSE), mean absolute error (MAE) and mean absolute percentage
error (MAPE) loss criteria. However, Franses and Van Dijk (1996) used the median of
squared error (MedSE) as the loss criterion, and concluded that the QGARCH and
GARCH models provide better out-of-sample forecasts than the GJR model. Forecasting
exchange rate volatility, Brooks and Burke (1998) found that GARCH(1,1) is preferable
on the MSE loss criterion, but not on the MAE criterion. Balaban (2004) documents that
EGARCH (GJR) is the best(worst) model for forecasting the exchange rate volatility.
However, the performance rankings were sensitive to the choice of loss criteria.
Because a heavily parameterized model is better able to capture the multiple
dimensions of volatility dynamics, it would generally have a better in-sample fit than a
relatively parsimonious model. However, a better in-sample fit may not necessarily
translate into a better out-of-sample forecasting performance. On the out-of-sample
forecasting ability, the simpler models often outperform the more complex models.
Comparing 330 ARCH-type forecasting models, Hansen and Lunde (2005) found no
evidence that the GARCH(1,1) model is outperformed by the more sophisticated models,
for the DM–$ exchange rate data. However, they found that the asymmetric GARCH
models perform better than the GARCH (1, 1) model for the IBM stock returns data. Bali
(2000) found that the asymmetric GARCH models give the best one-week-ahead
forecasts for the volatility of US T-bill yields. Numerous other studies have found that
on the out-of-sample forecasting performance, the parsimonious GARCH models
 perform better than the heavily parameterized models (Chiang and Huang, 2011; Stock index
Christoffersen and Jacobs, 2004; Hwang and Satchell, 2005; McMillan et al., 2000). volatility with
However, another set of studies found that the more sophisticated GARCH models
provide better volatility forecasts. Wei (2002) and Ulu (2005) found that the QGARCH
GARCH
model provides better out-of-sample forecasts than the GARCH model. Hansen and
Lunde (2006) compared several GARCH-family models, and found that the APARCH
model provides better forecasts than the more parsimonious GARCH models. 447
Forecasting the volatility of Madrid Stock Index (IBEX-35), Ñíguez (2008) found that the
fractionally integrated APARCH model provides the most accurate forecasts.
Antonakakis and Darby (2013) found that the FIGARCH model provides the best
volatility forecasts for the exchange rates of industrialized countries, whereas the
IGARCH model performs the best for the exchange rates of developing countries.
Finally, certain other studies provide mixed results, and suggest that the forecasting
horizon or market conditions may determine the choice of the best forecasting model.
Modeling and forecasting the exchange rate volatility, Akgül and Sayyan (2008) found
no clear winner among nine different GARCH models. McMillan and Garcia (2009)
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

forecast intraday volatility using a variety of GARCH-models. They found that the
choice of the best forecasting model is dependent on the forecasting horizon. Kışınbay
(2010) found that for the forecasting horizons of 10-30 days, the performance of the
asymmetric GARCH models and the linear GARCH model is statistically not different
from each other. However, for shorter forecasting horizons, asymmetric models provide
better forecasts than the GARCH model. Recently, Chiang and Huang (2011) found that
the GARCH model performs better in the up-trending (bull) markets, whereas the
EGARCH model performs better in the down-trending (bear) markets.
To sum it up, there is no consensus on which GARCH model provides the best
forecasting performance. Different studies, with different sample periods, different asset
classes and different performance evaluation criteria favor different GARCH
specifications. However, the asymmetric GARCH models are generally preferred over
the symmetric GARCH model.
We examine the forecasting performance of seven GARCH-family models for 21
world stock indices, with specific attention to the choice of appropriate benchmark and
loss criteria and the prevention of data-snooping bias. This paper contributes to the
GARCH forecasting literature in a number of ways. The problem of data-snooping is
pervasive in the GARCH forecasting literature. When comparing the forecasting models
based on a single data set, a certain model may outperform the benchmark model due to
chance, instead of a superior forecasting ability. It is common to find that different data
sets favor different GARCH specifications, in terms of out-sample performance. As a
result, the existing literature is rife with conflicting empirical evidence (discussed in
detail in the next section). Our empirical results confirm that the forecasting
performance of GARCH models is sensitive to the choice of data set, and there is no
unique GARCH model that provides the best forecasts for all the stock indices.
Therefore, a wide cross-sectional data set is essential to evaluate the relative forecasting
ability of the GARCH models, an aspect that has received limited research attention. For
instance, Hansen and Lunde’s (2005) study, which is arguably the most comprehensive
study in the literature that compares the forecasting ability of 330 ARCH-type models, is
based on just two data sets – the daily DM–$ exchange rate and the daily IBM returns.
Our major contribution is that we avoid the data-snooping problem in two ways. First,
 SEF we compare the forecasting performance of GARCH models for 21 international stock
32,4 indices, over a period of 13 years. To the best of our knowledge, no prior study uses such
an extensive dataset. Second, we apply Hansen’s (2005) superior predictive ability test
(SPA) to compare the forecasting performance of GARCH models. The SPA provides a
statistical framework for comparing multiple forecasting models, which is robust to the
data-snooping bias. This ensures that our empirical results are not sensitive to the
448 choice of data set, and they reflect the genuine forecasting ability of the GARCH models.
Another novel contribution of this study is that it examines the effect of market
conditions on the forecasting ability of various GARCH models. Market conditions have
a significant influence on the return generation process and the investor behavior.
However, this aspect has not attracted sufficient academic inquiry in the volatility
forecasting literature. We compare the forecasting performance of GARCH models
under the up-trending (bull) and down-trending (bear) market conditions. This
investigation yields some valuable insights. For instance, the relative forecasting
performance of the GARCH models is sensitive to the prevailing market conditions, and
the difference in the performance is more pronounced in the up-trending market.
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

From the methodological perspective, we allow for a rich choice of parameterization

in our models (16 different lag specifications), to ensure sufficient flexibility in
simulating the complex volatility dynamics. To ease the computational burden of
rolling estimation of GARCH models, several studies assume the simplest lag
specification (p ⫽ q ⫽ 1) for all GARCH processes. However, this assumption can be too
restrictive[1]. We determine the optimal lag-length using the Akaike information
criterion (AIC). Additionally, we use numerous performance evaluation criteria to
compare the forecasting models. Lee (1991) notes that the out-of-sample performance of
volatility forecasting models is sensitive to the choice of forecast evaluation criteria. To
ensure that our findings are robust, we compare the forecasts using four statistical loss
functions, the Mince-Zarnowitz regressions (Mincer and Zarnowitz, 1969) and the SPA
test.
We find that the advanced GARCH models do not provide better out-of-sample
forecasts than the standard GARCH model. A bewildering number of GARCH models
have been proposed in the literature, which attempt to incorporate the stylized facts
(such as long memory and leverage effects) of the conditional variance process. While
these advanced GARCH models may be good for understanding various features of the
conditional variance process, our results indicate that they do not provide any additional
value in forecasting applications. Advanced GARCH specifications are more likely to
model the idiosyncrasies of the historical data, which have no relation with the future
realizations of the conditional variance process. This has an adverse effect on their
forecasting performance. From the standpoint of the practitioners, this study addresses
the model selection problem. We recommend that the standard GARCH model is good
enough, if not better than the more advanced GARCH models for the volatility
forecasting of stock indices. For future research, the key implication is that additional
investigation may be required to test the persistence of the parameters that incorporate
various stylized facts. For instance, if the estimates for the leverage effect parameter are
highly unstable over time, excluding it from the forecasting model might improve the
out-of-sample performance of the model.
The remaining article is organized as follows. Section describes the data set used in
this analysis. Section 3 describes the methodology used to compare the forecasting
 performance of the GARCH models. Section 4 discusses the empirical results, and Stock index
Section 5 concludes the article. volatility with
GARCH
2. Data
We use the daily price observations of 21 major stock indices of the world, for the period
January 1, 2000 to November 30, 2013. The sample of the study includes eight stock
indices from the Europe region, nine from the Asia-Pacific region and four from the 449
Americas region. Appendix 1 lists the selected stock indices. All data are sourced from
the Bloomberg financial database. We forecast the daily volatility for each of the stock
indices, using the GARCH-family models. For comparability, it is desirable that the size
of the estimation period (in-sample period) and the number of out-of-sample forecasts is
the same across all the stock indices. For ensuring this, we truncate the price series
samples of all the indices at 3,395 trading days, starting from January 1, 2000. In total,
3,395 observations are chosen, as these are the daily observations available for the
Jakarta Stock Exchange Composite Index (JCI), which has the least number of trading
days for the sample period. Table I provides the descriptive statistics for the daily index
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

returns.

3. Methodology
3.1 Modeling conditional mean
Awartani and Corradi (2005) compared the forecasting performance of GARCH-type
models under six different specifications of conditional mean equation. They found that

Index Mean SD Skewness Excess-Kurtosis Maximum Minimum

AEX ⫺3.700 24.017 ⫺0.053 3.378 10.028 ⫺9.590

AORD 3.765 15.780 ⫺0.608 3.405 5.360 ⫺8.554
ATX 5.713 23.749 ⫺0.304 4.090 12.021 ⫺10.253
BEL20 ⫺1.003 20.934 0.046 2.847 9.334 ⫺8.319
SENSEX 9.823 25.725 ⫺0.182 3.392 15.990 ⫺11.809
IBOV 8.215 29.583 ⫺0.090 0.767 13.677 ⫺12.096
CAC ⫺2.253 24.318 0.027 1.654 10.595 ⫺9.472
UKX ⫺0.013 20.017 ⫺0.124 2.915 9.384 ⫺9.265
GD ⫺10.970 27.907 ⫺0.018 1.149 13.431 ⫺10.214
DAX 2.335 25.014 0.024 1.139 10.797 ⫺7.433
SPX 1.543 20.839 ⫺0.174 4.661 10.957 ⫺9.470
SPTSX 3.513 18.942 ⫺0.640 5.727 9.370 ⫺9.788
HSI 2.290 25.045 ⫺0.068 4.774 13.407 ⫺13.582
JCI 13.408 23.401 ⫺0.678 2.902 7.623 ⫺10.954
KOSPI 4.793 26.927 ⫺0.530 2.277 11.284 ⫺12.805
MERVAL 17.093 33.647 ⫺0.121 1.788 16.117 ⫺12.952
NKY ⫺1.415 24.966 ⫺0.419 3.246 13.235 ⫺12.111
SHCOMP 3.408 25.456 ⫺0.087 1.172 9.401 ⫺9.256
SMI 0.910 19.385 0.015 3.151 10.788 ⫺8.108
STI 1.473 19.416 ⫺0.423 3.127 7.531 ⫺9.216 Table I.
TWSE ⫺0.297 23.827 ⫺0.236 ⫺0.269 6.525 ⫺9.936 Descriptive statistics
for the daily return
Note: The mean returns and standard deviations are annualized series
 SEF the forecasting performance rankings remain consistent under different specifications
32,4 of the conditional mean equation. We use a general specification, the GARCH-in-mean
suggested by Engle et al. (1987), for modeling the conditional mean[2].

rt ⫽ ␮0 ⫹ ␮1␴t⫺1
2
⫹ ␧t, ␧t ⬃ N( 0, ␴t2 ) (1)

450 Where rt and ␴t2 represent the conditional mean and conditional variance for the day t, and ␧t
is the innovation for the day t. Note that these specifications nest the simpler models like the
constant mean model ( ␮1 ⫽ 0), and the zero mean model ( ␮0 ⫽ ␮1 ⫽ 0).

3.2 Modeling conditional variance

The literature on the GARCH-family models is extensive. However, for compactness, we
restrict our investigation to the seven more popular models, namely, GARCH
(Bollerslev, 1986), EGARCH (Nelson, 1991), GJR-GARCH (Glosten et al., 1993), TGARCH
(Zakoian, 1994), AVGARCH (Schwert, 1989; Taylor, 1986), APARCH (Ding et al., 1993)
and NGARCH (Higgins and Bera, 1992). The conditional variance specifications for
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

these models are as follows. The notations are consistent with Hansen and Lunde (2005):
p a
GARCH: ␴t2 ⫽ ␻ ⫹ ⌺ ␣i␧t⫺i
2
⫹ ⌺ ␤j␴t⫺j
2
(2)
i⫽1 j⫽1

p a
TGARCH: ␴t ⫽ ␻ ⫹ ⌺ ␣i关 ( 1 ⫺ ␥i )␧t⫺i
⫹
⫺ ( 1 ⫹ ␥i )␧t⫺i
⫺
兴 ⫹ j⫽1
⌺ ␤j␴t⫺j (3)
i⫽1

p a
AVGARCH: ␴t ⫽ ␻ ⫹ ⌺ ␣i ⱍ ␧t⫺i ⱍ ⫹ ⌺ ␤j␴t⫺j (4)
i⫽1 j⫽1

p a
NGARCH: ␴t␦ ⫽ ␻ ⫹ ⌺ ␣i ⱍ ␧t⫺1 ⱍ ␦ ⫹ ⌺ ␤j␴t⫺j
␦
(5)
i⫽1 j⫽1

p a
APARCH: ␴t␦ ⫽ ␻ ⫹ ⌺ ␣i关 ⱍ ␧t⫺1 ⱍ ⫺ ␥i␧t⫺1 兴␦ ⫹ ⌺ ␤j␴t⫺j
␦
(6)
i⫽1 j⫽1

p a
GJR: ␴t2 ⫽ ␻ ⫹ ⌺ 关 ␣i ⫹ ␥iI兵␧t⬎0其 兴␧t⫺i
2
⫹ ⌺ ␤j␴t⫺j
2
(7)
i⫽1 j⫽1

p a
EGARCH: log ( ␴t2 ) ⫽ ␻ ⫹ ⌺ ␣iet⫺i ⫹ ␥i( ⱍ et⫺i ⱍ ⫺ E ⱍ et⫺i ⱍ ) ⫹ ⌺ ␤jlog ( ␴t⫺j
2
) (8)
i⫽1 j⫽1

Here, ␴t2 is the variance for the day t. ␧t is the innovation and et ⫽ ␧t␴t⫺1 is the
standardized innovation for the day t. ␧t⫹ ⫽ ␧t if ␧t ⬎ 0, and ␧t⫹ ⫽ 0 if ␧t ⱕ 0. Likewise,
␧t⫺ ⫽ ␧t if ␧t ⱕ 0, and ␧t⫺ ⫽ 0 if ␧t ⬎ 0. I兵␧t⬎0其 ⫽ 1 if ␧t ⬎ 0, and I兵␧t⬎0其 ⫽ 0 if ␧t ⱕ 0. ␻, ␣i,
␤j, ␥i and ␦ are the parameters estimated using the method of maximum likelihood.

3.3 Forecasting methodology

For each index, we generate 1,394 one-step-ahead forecasts[3], using fixed-size rolling
windows of the last 2,000 observations, re-estimating the parameters for each period.
The parameters are estimated using the method of maximum likelihood. The selection
of the estimation window is primarily subjective, as there is no specific theoretical
 prescription for the optimal size of the estimation window. Generally, using a large Stock index
estimation window is statistically desirable, as it increases the efficiency of the volatility with
parameter estimates. A practical consideration is to use a sufficiently long estimation
window to avoid the non-convergence problem in the likelihood-based estimation,
GARCH
particularly when estimating the more richly parameterized GARCH models. Using a
long estimation window may be detrimental under an equal-weighting scheme, like the
moving average model, as the forecasts would get influenced by the past information 451
that may no longer be relevant. However, this consideration would not apply to the
GARCH models that have exponentially decaying weights, which limit the influence of
the data that is too old.
Our choice of the length of estimation window (N ⫽ 2000 days) is based on these
considerations and on previous forecasting applications of the GARCH models. The
previous forecasting studies that make use of similar estimation windows include:
Ferulano (2009) – N ⫽ 2,000 days for 75 financial assets belonging to multiple asset
classes; da Veiga et al. (2012) – N ⫽ 2,000 days for the S&P 500 index; Hansen and Lunde
(2005) – N ⫽ 2,378 days for the IBM stock data; Alberg et al. (2008) – N ⫽ 1911 days for
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

the TA100 stock index; Karmakar and Shukla (2015) – N varies from 2,210 to 2,272 days
for six international stock indices.
Next, we introduce some notations to facilitate the discussion. N is the size of the
estimation window. M is the number of out-of-sample forecasts (M ⫽ 1,394 for our
analysis). ␴t2 is the true (latent) conditional variance for the day t. ␴ˆ t2 is the
one-step-ahead forecast of ␴t2, based on the information rt⫺1, rt⫺2,…,rt⫺N and ␺ˆ , where,
t ⫽ N ⫹ 1 …, N ⫹ M and ␺ˆ is the vector of estimated parameters of the forecasting
model. We allow for a rich choice of parameterization by allowing the lag-lengths p and
q to vary from one to four, as against p ⫽ q ⫽ 1, used by most of the earlier studies.
Therefore, we compare 16 lag specifications for each model. The best model is selected
based on the AIC.

3.4 Forecast performance evaluation

The evaluation of forecasting performance is a non-trivial task due to two reasons:
(1) The true conditional variance ␴t2 is not directly observable.
(2) There is no unique criterion for selecting the best forecasting model.

To address the first issue, the high-frequency-data-based realized variance is often

considered as a good approximation for the true conditional variance (Andersen et al.,
2001a, 2001b, 2003; Barndorff-Nielsen and Shephard, 2002; Barndorff-Nielsen, 2002).
Because it is not practical for us to obtain the high-frequency intradaily data for all the
stock indices, we use a range-based estimator as the proxy for the true conditional
variance, as the next best alternative. Moreover, processing high-frequency data for so
many indices is computationally complex owing to large number of observations,
asynchronous and irregularly spaced timestamps and the microstructure noise. The
range-based estimators are 5 to 14 times more efficient than the historical variance
(Garman and Klass, 1980; Parkinson, 1980; Rogers and Satchell, 1991; Yang and Zhang,
2000). Alizadeh et al. (2002) showed that the range-based volatility proxies are highly
efficient and robust to the microstructure effects. The range-based estimators capture
the extreme intraday fluctuations in the asset price, and, therefore, they contain more
information than the squared daily returns. The empirical evidence also favors the use
 SEF of range-based estimators to model the latent volatility. Shu and Zhang (2006) found that
32,4 the range-based estimators provide very precise estimates of the daily integrated
volatility. Jacob and Vipul (2008) found that the volatility forecasts based on the
range-based estimators are almost as efficient as those based on the realized variance
estimator. More recently, forecasting the volatility for the German DAX 30 index,
Todorova (2012) found that all the range-based estimators outperform the classical
452 daily-return-based estimator. Molnár (2012) modeled the daily volatility of the
component stocks of Dow Jones Industrial Average. He found that the performance of
the range-based Garman Klass estimator (Garman and Klass, 1980) is close to the
realized variance estimator (Andersen et al., 2001a, 2001b).
Out of the four major range-based volatility estimators [the Parkinson estimator
(Parkinson, 1980), the Garman and Klass estimator (Garman and Klass, 1980), the
Rogers and Satchell estimator (Rogers and Satchell, 1991) and the Yang and Zhang
estimator (Yang and Zhang, 2000)], we select the Yang and Zhang estimator. This is
owing to its robustness to both price drift and opening jumps. The Yang and Zhang
daily volatility estimator, ␴ˆ YZ
2
, is given by:
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

␴ˆ YZ
2
⫽ ␴ˆ o2 ⫹ k␴ˆ c2 ⫹ ( 1 ⫺ k )␴ˆ RS
2
(9)

where:

1 n
␴ˆ o2 ⫽ ⌺ ( o ⫺ oˉ )2
n ⫺ 1 i⫽1 i

1 n
␴ˆ c2 ⫽ ⌺ ( c ⫺ cˉ)2
n ⫺ 1 i⫽1 i

1 n
␴ˆ RS
2
⫽ ⌺ 关 h ( h ⫺ ci ) ⫺ li( li ⫺ ci ) 兴
n i⫽1 i i

0.34
k⫽
1.34 ⫹ ( n ⫹ 1 ) / ( n ⫺ 1 )

Here, Oi, Hi, Li and Ci represent the opening, highest, lowest and closing prices on the day
i, respectively. The log returns are given by oi ⫽ ln Oi ⫺ ln Ci⫺1 (close-to-open return),
hi ⫽ ln Hi ⫺ ln Oi (open-to-high return), li ⫽ ln Li ⫺ ln Oi (open-to-low return) and ci ⫽
ln Ci ⫺ ln Oi (open-to-close return). The mean returns are oˉ ⫽ 兺 i⫽1n
oi, cˉ ⫽ 兺 i⫽1
n
ci, and n
is the number of days used to estimate the Yang and Zhang daily variance. Yang and
Zhang (2000) proved that the peak efficiency is reached for n ⫽ 2. Accordingly we use
n ⫽ 2 (two-days’ data) for this study.
The second issue is identifying a suitable performance evaluation criterion for
selecting the best forecasting model. The general approach is to evaluate the competing
forecasts on multiple loss functions. Following Koopman et al. (2005), we use the
following four loss functions:

n
Mean squared error: MSE( ␴ˆ t2, ␴t2 ) ⫽ n ⫺1 ⌺ ( ␴ˆ t2 ⫺ ␴t2 )2 (10)
t⫽1
 n
Heteroskedasticity-adjusted MSE: HMSE( ␴ˆ t2, ␴t2 ) ⫽ n ⫺1 ⌺ ( ␴ˆ t2 ⫺ ␴t2 )2␴t⫺4 (11) Stock index
t⫽1
volatility with
n GARCH
Mean absolute error: MAE( ␴ˆ t2, ␴t2 ) ⫽ n ⫺1 ⌺ ⱍ ␴ˆ t2 ⫺ ␴t2 ⱍ (12)
t⫽1

n
Heteroskedasticity-adjusted MAE: HMAE( ␴ˆ t2, ␴t2 ) ⫽ n ⫺1 ⌺ ⱍ ␴ˆ t2 ⫺ ␴t2 ⱍ ␴t⫺2 (13) 453
t⫽1

Here, ␴ˆ t2 is the variance forecast for the day t; ␴t2 is the true (latent) variance for the day
t. The Yang and Zhang (2000) estimator is used as a proxy for the true (latent)
conditional variance ␴t2. Patton (2011) showed that, in the presence of a noisy proxy,
some loss functions may prefer an imperfect forecast to the true conditional variance.
Among the selected loss functions, only MSE is robust to the noise in the conditional
variance proxy (Patton, 2011). Nonetheless, for completeness and comparability with
the existing studies, we evaluate the forecasts across all the loss criteria. Moreover, as
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

the range-based proxies provide fairly accurate estimates of the true conditional
variance (Alizadeh et al., 2002; Shu and Zhang, 2006), the noise is expected to be minimal.
We also use the Mince-Zarnowitz regressions (Mincer and Zarnowitz, 1969), which is
a widely used forecast evaluation methodology. In this procedure, the true variance
series ( ␴t2 ) is regressed on the forecasted variance series ( ␴ˆ t2 ). The R2 of this regression
is used to evaluate the accuracy of forecasts. Finally, we compare the forecasting models
using the Hansen’s (2005) superior predictive ability (SPA) test. It is based on the
contention that a single sample comparison of forecasting models provides little
information about how the models would fare against each other, for a different data
sample. In a particular sample, a forecasting model may outperform the others just by
chance (or worse, due to data-snooping), instead of superior forecasting ability. The SPA
test provides a statistical framework for comparing multiple forecasting models, which
is robust to the data-snooping bias. We implement the SPA test using the stationary
block bootstrap procedure of Politis and Romano (1994), with 10,000 block bootstrap
resamples. The lengths of the blocks are independent and randomly drawn from a
geometric distribution with the mean 1/q. Following Koopman et al. (2005), we have
taken q ⫽ 0.5. A detailed discussion on the implementation of the SPA test is provided
by Hansen (2005). For the remainder of this article, the GARCH model is referred to as
the benchmark model, and the other six models, i.e. the TGARCH, AVGARCH,
APARCH, GJR, NGARCH and EGARCH models, are referred to as the competing
models.

4. Empirical results and discussion

Table II compares the forecasting accuracy of the GARCH models on various loss
criteria and provides the model ranking statistics. Panel A compares the forecasts of the
seven GARCH models using the MSE criterion. For each of the stock indices, we rank the
forecasting models from one (the best model) to seven (the worst model). The GARCH
model outperforms all the other models with a mean rank of 2.952. TGARCH is the next
best model with a mean rank of 3.667, and APARCH is the worst model, with a mean
rank of 5.095. Interestingly, the APARCH model is the most generic GARCH
specification of our study, and it nests the GARCH, TGARCH, AVGARCH, GJR and
NGARCH models. The Ranked Best metric reports the number of indices (out of 21) for
 SEF
32,4 Index GARCH TGARCH AVGARCH NGARCH APARCH GJR EGARCH

Panel A: Comparison of the forecasting models based on the MSE criterion

Mean MSE 2.513 2.670 2.735 2.803 3.010 2.902 2.865
Median MSE 2.275 2.427 2.337 2.577 2.967 2.637 2.642
454 Mean rank 2.952 3.667 3.81 3.81 5.095 4.19 4.476
Ranked best 6 2 2 3 0 2 6
Binomial p-value (0.068) (0.823) (0.823) (0.594) (1.000) (0.823) (0.068)
Ranked worst 0 0 1 3 4 4 9**
Binomial p-value (1.000) (1.000) (0.961) (0.594) (0.352) (0.352) (0.001)
Wilcoxon test Statistic 102 98 40.0** 35.0** 52.0* 93
p-value (0.329) (0.281) (0.005) (0.002) (0.014) (0.226)
Panel B: Comparison of the forecasting models based on the HMSE criterion
Mean HMSE 0.779 0.833 0.860 0.818 0.860 0.849 0.812
Median HMSE 0.547 0.542 0.536 0.571 0.533 0.613 0.529
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

Mean rank 3.333 3.476 4.476 4.143 4.429 4.381 3.762

Ranked best 5 5 1 5 1 0 4
Binomial p-value (0.171) (0.171) (0.961) (0.171) (0.961) (1.000) (0.352)
Ranked worst 1 1 7* 8** 3 1 0
Binomial p-value (0.961) (0.961) (0.022) (0.006) (0.594) (0.961) (1.000)
Wilcoxon test Statistic 102 86 37.0** 72.0 66.0 104
p-value (0.329) (0.160) (0.003) (0.069) (0.075) (0.354)
Panel C: Comparison of the forecasting models based on the MAE criterion
Mean MAE 0.939 0.959 0.961 0.966 0.983 0.975 0.955
Median MAE 0.880 0.841 0.858 0.858 0.900 0.874 0.850
Mean rank 2.762 3.524 4.143 4.143 4.905 4.286 4.238
Ranked best 9** 1 1 3 0 1 6
Binomial p-value (0.001) (0.961) (0.961) (0.594) (1.000) (0.961) (0.068)
Ranked worst 0 0 4 5 3 3 6
Binomial p-value (1.000) (1.000) (0.352) (0.171) (0.594) (0.594) (0.068)
Wilcoxon test Statistic 105 103 35.0** 46.0** 52.0* 110.5
p-value (0.367) (0.341) (0.005) (0.007) (0.013) (0.438)
Panel D: Comparison of the forecasting models based on the HMAE criterion
Mean HMAE 0.640 0.640 0.647 0.652 0.647 0.643 0.635
Median HMAE 0.576 0.551 0.538 0.584 0.555 0.553 0.540
Mean rank 3.714 3.81 4.238 4.048 4 4.095 4.095
Ranked best 4 2 3 7* 1 1 3
Binomial p-value (0.352) (0.823) (0.594) (0.022) (0.961) (0.961) (0.594)
Ranked worst 1 1 6 8** 2 1 2
Binomial p-value (0.961) (0.961) (0.068) (0.006) (0.823) (0.961) (0.823)
Wilcoxon test Statistic 100 94.5 42.5** 97 110.5 107
p-value (0.433) (0.354) (0.010) (0.390) (0.438) (0.393)

Table II. Notes: For each index, we rank the models from one (best) to seven (worst). Ranked best (worst) is the
Comparing the number of times a forecasting model is ranked as the best (worst). Wilcoxon signed-ranks test is used to
forecast accuracy of test the null hypothesis that the median loss function of the GARCH model is equal to that of a
GARCH-models competing model. Statistical significance at the 1 and 5% levels is indicated by ** and * , respectively
based on loss criteria
 which a particular model is ranked the best. To check whether this number is Stock index
significantly more than that expected by chance (H0: each model is equally likely to be volatility with
the best), we find the cumulative binomial probability of observing it under the null
hypothesis. The resulting p-values that represent the statistical significance of the
GARCH
Ranked Best metric are reported in the next row (row five of Panel A). Similarly,
the Ranked Worst metric counts the number of indices for which a particular model is
ranked the worst, and its associated binomial p-value is given in the subsequent row 455
(row seven of Panel A). The GARCH and EGARCH models are ranked as the best model
for the maximum number of stock indices (six for each model). The associated binomial
test suggests that this number is statistically significant at 10 per cent level of
significance. However, the EGARCH model also ranks as the worst model for nine stock
indices, whereas the GARCH model does not rank as the worst model for any stock
index. Finally, we select the GARCH model as our benchmark and compare its MSE with
those of the competing models. We use the one-tailed Wilcoxon signed-rank test under
the null hypothesis that the median of MSE of the benchmark model is equal to that of a
competing model (one competing model at a time). We find that the median MSE for the
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

GARCH model is significantly lower than that of each of the NGARCH, APARCH and
GJR models at 5 per cent level of significance.
Panel B presents the comparisons on the HMSE criterion. On mean rank, the GARCH
model again outperforms the competing models with the lowest mean rank of 3.333. The
TGARCH model follows as a close second with a mean rank of 3.476, and the
AVGARCH model performs the worst with a mean rank of 4.476. For most of the indices,
the GARCH, TGARCH and NGARCH models are ranked as the best performers, and the
AVGARCH and NGARCH models are ranked as the worst performers. The Wilcoxon
signed-rank tests suggest that the median HMSE for the GARCH model is significantly
lower than that of each of the NGARCH, APARCH and GJR models at 10 per cent level
of significance.
Panel C compares the forecasting models on the MAE criterion. Here, the GARCH
(APARCH) model is the best (worst) model with a mean rank of 2.762 (4.905). The
TGARCH model is the next best with a mean rank of 3.524. The GARCH model is ranked
as the best model for nine stock indices and is never ranked as the worst model. The
Wilcoxon signed-ranks tests suggest that the median MAE for the GARCH model is
significantly lower than that of each of the NGARCH, APARCH and GJR models at the
5 per cent level of significance. Finally, Panel D compares the forecasting models on the
HMAE criterion. All the models perform quite close to each other on the HMAE
criterion. The GARCH model still has the lowest mean rank, although the difference
between the competing models is marginal. The performance of the NGARCH model
fluctuates considerably across the various stock indices. It is ranked as the best model
for seven indices, and the worst model for eight indices. The GARCH model is ranked as
the best model for four indices and the worst model for one index. The Wilcoxon
signed-rank tests suggest that the median HMAE for the GARCH model is significantly
lower than that for the NGARCH model. However, we fail to reject the null for the other
competing models.
To test the robustness of our results to different market conditions, we divide the
total sample period of 1,394 days into five contiguous sub-periods: four sub-periods of
279 days each and one sub-period of 278 days. For each stock index, we sort the
sub-periods based on their realized (ex-post) returns. We classify the sub-periods with
 SEF the highest and the lowest return, as the up-trending and the down-trending periods, and
32,4 we discard the remaining three sub-periods. The performance evaluation exercise is
repeated for the up-trending and down-trending market conditions separately, and
Table III presents the results. For brevity, we provide only the mean ranks for the
different GARCH-models (detailed results are available on request). In line with the
earlier results, GARCH outperforms the competing models on all the loss criteria, for
456 both the up-trending and down-trending market conditions. The difference in the
performance rankings of various GARCH-models is more pronounced for the
up-trending market. In the up-trending market, the GARCH and NGARCH models
perform the best, whereas the AVGARCH and EGARCH models perform the worst. In
the down-trending market, the GARCH and APARCH models are the best, whereas the
AVGARCH and NGARCH models are the worst.
Table IV reports the forecast accuracy of GARCH-models based on the R2 of
Mince-Zarnowitz regressions. The high values of R2 for a forecasting model indicate
that the pattern of one-step-ahead forecasts obtained from the model is close to
the pattern of the true variance. The R2 (goodness-of-fit) analysis confirms our earlier
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

results. The GARCH model has the lowest mean rank (1.810), and it is the best model for
14 out of 21 stock indices. We carry out the Wilcoxon signed-ranks tests under the null
that the median R2 for the GARCH model is equal to that of a competing model (one
competing model at a time). The null is rejected for all the competing models.
Next, we carry out the SPA tests with GARCH as the benchmark model. The null
hypothesis of our SPA test is that, for the selected loss criterion, none of the competing
models outperforms the benchmark model. Therefore, a low p-value (say 0.05) would
indicate that the benchmark model is outperformed by at least one of the competing
models. Table V reports the results of SPA tests. The benchmark model is preferred for
12, 13, 10 and 12 stock indices under the MSE, MAE, HMSE and HMAE loss criteria,

Loss criterion GARCH TGARCH AVGARCH NGARCH APARCH GJR EGARCH

Panel A: Mean ranks for the up-trending market condition

MSE 3.286 4.190 4.333 3.381 4.286 3.810 4.714
HMSE 3.190 4.095 5.190 3.524 3.571 3.762 4.667
MAE 2.810 4.429 4.667 3.381 4.190 3.667 4.857
HMAE 3.238 4.524 4.857 3.476 3.476 3.524 4.905
Average 3.131 4.310 4.762 3.441 3.881 3.691 4.786
Panel B: Mean ranks for the down-trending market condition
MSE 3.429 3.857 4.524 3.810 4.048 4.190 4.095
HMSE 3.048 4.571 4.143 4.905 3.619 4.000 3.714
MAE 3.619 3.857 4.190 3.714 4.190 4.143 4.286
HMAE 3.333 4.476 4.000 4.524 3.429 3.905 4.333
Average 3.357 4.190 4.214 4.238 3.822 4.060 4.107
Table III.
Performance Notes: For each of the 21 sample indices, we rank the models from one (best) to seven (worst), based on
rankings of GARCH- the various loss criteria (indicated in the rows). The scores reported in this table are the averages of these
models under 21 ranks (the mean rank for each model). Panels A and B report the mean ranks for the up-trending and
different market the down-trending market conditions. The last row of each panel provides the average of the mean
conditions ranks as a summary statistic
 respectively (at 5 per cent level of confidence). On an average, the GARCH model is Stock index
preferred over all the six competing models, for over half of the sample stock indices, volatility with
regardless of the choice of loss criterion. To check whether the number of times the
benchmark model is preferred, is significantly more than that expected by chance (H0:
GARCH
each model is equally likely to be the best), we find the cumulative binomial probability
of observing it under the null hypothesis. For all the loss functions, the null hypothesis
is rejected even at 0.1 per cent level of significance. Therefore, there is compelling 457
evidence that the benchmark model provides significantly better forecasts than the
competing models, regardless of the choice of the loss criteria or data points.

5. Conclusion
We compare the GARCH model with six advanced GARCH models, namely, the
EGARCH, GJR-GARCH, TGARCH, AVGARCH, APARCH and NGARCH models.
Using a sample of 21 major stock indices of the world, for a period of about 13 years, we
find that the GARCH model provides the best one-step-ahead forecasts for the daily
conditional variance. Our results are robust to the choice of performance evaluation
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

criteria and different market conditions. The selection of a comprehensive dataset and
the application of SPA tests ensure that our results are free from data snooping bias. Our
results are consistent with the findings of Hansen and Lunde (2005), who compared 330
different forecasting models and found that none of them provides significantly better
forecasts than the GARCH(1,1) model. Advanced GARCH specifications are better at
modeling the complex patterns that are empirically observed in the conditional variance
process. However, they are also more likely to model the idiosyncrasies of the historical
data, which have no relation with the future realizations of the conditional variance
process. This has an adverse effect on their forecasting performance. Our empirical
results indicate that the latter effect dominates, and the basic GARCH model usually
provides the best one-step-ahead forecasts of conditional variance for the stock indices.
Our results are in line with the principle of parsimony, which suggests that in terms of
out-of-sample forecasting performance, parsimonious models are usually preferable to
the models that are more complex. We believe that despite their appealing theoretical
properties, the richly parameterized GARCH models are not ideal for the forecasting

Index GARCH TGARCH AVGARCH NGARCH APARCH GJR EGARCH

Mean R2 0.187 0.172 0.168 0.185 0.175 0.177 0.169

Median R2 0.192 0.181 0.177 0.188 0.181 0.184 0.176
Mean rank 1.81 4.476 5.286 2.81 4.524 3.524 5.571
Ranked best 14** 1 0 2 0 3 1
Binomial p-value (0.000) (0.961) (1.000) (0.823) (1.000) (0.594) (0.961)
Ranked worst 1 1 7* 0 3 1 8**
Binomial p-value (0.961) (0.961) (0.022) (1.000) (0.594) (0.961) (0.006) Table IV.
Wilcoxon test Statistic 201.0** 225.0** 125.0** 205.5** 192.5** 204.5** Comparing the
p-value (0.000) (0.000) (0.011) (0.000) (0.001) (0.000) forecast accuracy of
GARCH-models
Notes: For each index, we rank the models from one (best) to seven (worst). Ranked best (Worst) is the based on the R2 of
number of times a forecasting model is ranked as the best (worst). Wilcoxon signed-ranks test is used to the Mincer-
test the null hypothesis that the median R2 of the GARCH model is equal to that of a competing model. Zarnowitz
Statistical significance at the 1 and 5% level is indicated by ** and * , respectively regressions
 SEF Index MSE MAE HMSE HMAE
32,4
AEX 0.704 0.736 0.125 0.000
AORD 0.000 0.000 0.001 0.540
ATX 0.000 0.000 0.000 0.649
BEL20 0.951 0.890 0.600 0.238
458 SENSEX 0.942 0.782 0.867 0.000
IBOV 0.792 0.831 0.505 0.120
CAC 0.527 0.978 0.002 0.006
UKX 0.001 0.000 0.000 0.498
GD 0.001 0.004 0.001 0.000
DAX 0.946 0.622 0.001 0.092
SPX 0.000 0.000 0.005 0.511
SPTSX 0.002 0.000 0.829 0.002
HSI 0.002 0.000 0.002 0.490
JCI 0.990 0.920 0.547 0.125
KOSPI 0.955 0.468 0.040 0.000
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

MERVAL 0.138 0.090 0.942 0.037

NKY 0.000 0.000 0.000 0.960
SHCOMP 0.811 0.882 0.685 0.166
SMI 0.509 0.676 0.032 0.001
STI 0.003 0.117 0.639 0.667
TWSE 0.995 0.759 0.158 0.006
Benchmark preferred 12 13 10 12
Table V. Binomial p-value (0.000) (0.000) (0.000) (0.000)
Superior predictive
ability (SPA) tests Notes: This table reports the results for SPA tests with the GARCH model selected as the benchmark
with GARCH as the model. Benchmark preferred is the number of times the benchmark model is preferred over the
benchmark model competing models (at 5% level of significance, out of 21 indices)

applications. For instance, it is widely believed that the linear GARCH model is
inadequate to model the asymmetric and skewed return distributions. However, Gokcan
(2000) found that the GARCH (1,1) model provides better out-of-sample performance
than the EGARCH model, even when the returns had a skewed distribution. Our results
are consistent with a number of previous studies that found that the non-linear GARCH
models (like the EGARCH or GJR-GARCH models) do not provide better forecasting
performance than the standard GARCH model (Balaban, 2004; McMillan et al., 2000; Ng
and McAleer, 2004).

Notes
1. The GARCH (1,1) model represents the conditional variance as a function of infinite number
of forecast errors (innovations) with geometrically declining weights. Therefore, the GARCH
(1,1) model imposes a single smoothly decaying lag pattern, which may not be optimal.
2. We tested three alternative conditional mean specifications: zero mean, constant mean and
AR(1). The key results of this analysis remain robust under different specifications of the
conditional mean.
3. This analysis restricts itself to comparing only the one-step-ahead variance forecasts for the
various GARCH models. However, multi-period forecasts may be required for the risk
 management and option pricing applications. For instance, under the Basel capital adequacy Stock index
framework, the capital requirements for the banks are based on a forecast of 10-day 99% VaR.
volatility with
For recent applications of multi-period VaR forecasts, see Louzis et al. (2013) and Kinateder
and Wagner (2014). GARCH

References
Akgül, I. and Sayyan, H. (2008), “Modelling and forecasting long memory in exchange rate
459
volatility vs. stable and integrated GARCH models”, Applied Financial Economics, Vol. 18
No. 6, pp. 463-483.
Alberg, D., Shalit, H. and Yosef, R. (2008), “Estimating stock market volatility using asymmetric
GARCH models”, Applied Financial Economics, Vol. 18 No. 15, pp. 1201-1208.
Alizadeh, S., Brandt, M.W. and Diebold, F.X. (2002), “Range-based estimation of stochastic
volatility models”, The Journal of Finance, Vol. 57 No. 3, pp. 1047-1091.
Andersen, T.G., Bollerslev, T., Diebold, F.X. and Ebens, H. (2001a), “The distribution of realized
stock return volatility”, Journal of Financial Economics, Vol. 61 No. 1, pp. 43-76.
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

Andersen, T.G., Bollerslev, T., Diebold, F.X. and Labys, P. (2001b), “The distribution of realized
exchange rate volatility”, Journal of the American Statistical Association, Vol. 96 No. 453,
pp. 42-55.
Andersen, T.G., Bollerslev, T., Diebold, F.X. and Labys, P. (2003), “Modeling and forecasting
realized volatility”, Econometrica, Vol. 71 No. 2, pp. 579-625.
Antonakakis, N. and Darby, J. (2013), “Forecasting volatility in developing countries’ nominal
exchange returns”, Applied Financial Economics, Vol. 23 No. 21, pp. 1675-1691.
Awartani, B.M.A. and Corradi, V. (2005), “Predicting the volatility of the S&P-500 stock index via
GARCH models: the role of asymmetries”, International Journal of Forecasting, Vol. 21
No. 1, pp. 167-183.
Balaban, E. (2004), “Comparative forecasting performance of symmetric and asymmetric
conditional volatility models of an exchange rate”, Economics Letters, Vol. 83 No. 1,
pp. 99-105.
Bali, T.G. (2000), “Modeling the conditional mean and variance of the short rate using diffusion,
GARCH, and moving average models”, Journal of Futures Markets, Vol. 20 No. 8,
pp. 717-751.
Bali, T.G. and Demirtas, K.O. (2008), “Testing mean reversion in financial market volatility:
evidence from S&P 500 index futures”, Journal of Futures Markets, Vol. 28 No. 1, pp. 1-33.
Barndorff-Nielsen, O.E. (1997), “Normal inverse Gaussian distributions and stochastic volatility
modelling”, Scandinavian Journal of Statistics, Vol. 24 No. 1, pp. 1-13.
Barndorff-Nielsen, O.E. (2002), “Econometric analysis of realized volatility and its use in
estimating stochastic volatility models”, Journal of the Royal Statistical Society: Series B
(Statistical Methodology), Vol. 64 No. 2, pp. 253-280.
Barndorff-Nielsen, O.E., Nicolato, E. and Shephard, N. (2002), “Some recent developments in
stochastic volatility modelling”, Quantitative Finance, Vol. 2 No. 1, pp. 11-23.
Barndorff-Nielsen, O.E. and Shephard, N. (2002), “Estimating quadratic variation using realized
variance”, Journal of Applied Econometrics, Vol. 17 No. 5, pp. 457-477.
Bollerslev, T. (1986), “Generalized autoregressive conditional heteroskedasticity”, Journal of
Econometrics, Vol. 31 No. 3, pp. 307-327.
Brailsford, T.J. and Faff, R.W. (1996), “An evaluation of volatility forecasting techniques”, Journal
of Banking & Finance, Vol. 20 No. 3, pp. 419-438.
 SEF Brooks, C. and Burke, S.P. (1998), “Forecasting exchange rate volatility using conditional variance
models selected by information criteria”, Economics Letters, Vol. 61 No. 3, pp. 273-278.
32,4
Cao, C.Q. and Tsay, R.S. (1992), “Nonlinear time-series analysis of stock volatilities”, Journal of
Applied Econometrics, Vol. 7 No. S1, pp. S165-S185.
Chiang, M.H. and Huang, H.Y. (2011), “Stock market momentum, business conditions, and
GARCH option pricing models”, Journal of Empirical Finance, Vol. 18 No. 3,
460 pp. 488-505.
Christoffersen, P. and Jacobs, K. (2004), “Which GARCH model for option valuation?”,
Management Science, Vol. 50 No. 9, pp. 1204-1221.
Da Veiga, B., Chan, F. and McAleer, M. (2012), “It pays to violate: how effective are the Basel
accord penalties in encouraging risk management?”, Accounting & Finance, Vol. 52 No. 1,
pp. 95-116.
Ding, Z., Granger, C.W.J. and Engle, R.F. (1993), “A long memory property of stock market returns
and a new model”, Journal of Empirical Finance, Vol. 1 No. 1, pp. 83-106.
Ederington, L.H. and Guan, W. (2005), “Forecasting volatility”, Journal of Futures Markets, Vol. 25
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

No. 5, pp. 465-490.

Engle, R.F., Lilien, D.M. and Robins, R.P. (1987), “Estimating time varying risk premia in the term
structure: the Arch-M model”, Econometrica, Vol. 55 No. 2, pp. 391-407.
Ferulano, R. (2009), “A mixed historical formula to forecast volatility”, Journal of Asset
Management, Vol. 10 No. 2, pp. 124-136.
Fleming, J. and Kirby, C. (2003), “A closer look at the relation between GARCH and stochastic
autoregressive volatility”, Journal of Financial Econometrics, Vol. 1 No. 3, pp. 365-419.
Franses, P.H. and Van Dijk, D. (1996), “Forecasting stock market volatility using (non-linear)
Garch models”, Journal of Forecasting, Vol. 15 No. 3, pp. 229-235.
Garman, M.B. and Klass, M.J. (1980), “On the estimation of security price volatilities from
historical data”, The Journal of Business, Vol. 53 No. 1, pp. 67-78.
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”, The Journal of Finance,
Vol. 48 No. 5, pp. 1779-1801.
Gokcan, S. (2000), “Forecasting volatility of emerging stock markets: linear versus non-linear
GARCH models”, Journal of Forecasting, Vol. 19 No. 6, pp. 499-504.
Hansen, P.R. (2005), “A test for superior predictive ability”, Journal of Business & Economic
Statistics, Vol. 23 No. 4, pp. 365-380.
Hansen, P.R. and Lunde, A. (2005), “A forecast comparison of volatility models: does anything
beat a GARCH(1,1)?”, Journal of Applied Econometrics, Vol. 20 No. 7, pp. 873-889.
Hansen, P.R. and Lunde, A. (2006), “Consistent ranking of volatility models”, Journal of
Econometrics, Vol. 131 Nos 1/2, pp. 97-121.
Harvey, A.C. and Shephard, N. (1996), “Estimation of an asymmetric stochastic volatility model
for asset returns”, Journal of Business & Economic Statistics, Vol. 14 No. 4, pp. 429-434.
Hentschel, L. (1995), “All in the family Nesting symmetric and asymmetric GARCH models”,
Journal of Financial Economics, Vol. 39 No. 1, pp. 71-104.
Higgins, M.L. and Bera, A.K. (1992), “A class of nonlinear arch models”, International Economic
Review, Vol. 33 No. 1, pp. 137-158.
Hwang, S. and Satchell, S.E. (2005), “GARCH model with cross-sectional volatility: GARCHX
models”, Applied Financial Economics, Vol. 15 No. 3, pp. 203-216.
 Jacob, J. and Vipul (2008), “Estimation and forecasting of stock volatility with range-based Stock index
estimators”, Journal of Futures Markets, Vol. 28 No. 6, pp. 561-581.
volatility with
Karmakar, M. and Shukla, G.K. (2015), “Managing extreme risk in some major stock markets: an
extreme value approach”, International Review of Economics & Finance, Vol. 35 No. 1,
GARCH
pp. 1-25.
Kinateder, H. and Wagner, N.F. (2014), “Multiple-period market risk prediction under long
memory: when VaR is higher than expected”, Journal of Risk Finance, Vol. 15 No. 1, pp. 4-32. 461
Kışınbay, T. (2010), “Predictive ability of asymmetric volatility models at medium-term horizons”,
Applied Economics, Vol. 42 No. 30, pp. 3813-3829.
Koopman, S.J., Jungbacker, B. and Hol, E. (2005), “Forecasting daily variability of the S&P 100
stock index using historical, realised and implied volatility measurements”, Journal of
Empirical Finance, Vol. 12 No. 3, pp. 445-475.
Lee, K.Y. (1991), “Are the GARCH models best in out-of-sample performance?”, Economics Letters,
Vol. 37 No. 3, pp. 305-308.
Lehar, A., Scheicher, M. and Schittenkopf, C. (2002), “GARCH vs. stochastic volatility: option
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

pricing and risk management”, Journal of Banking & Finance, Vol. 26 Nos 2/3, pp. 323-345.
Louzis, D.P., Xanthopoulos-Sisinis, S. and Refenes, A.P. (2013), “The role of high-frequency
intra-daily data, daily range and implied volatility in multi-period value-at-risk
forecasting”, Journal of Forecasting, Vol. 32 No. 6, pp. 561-576.
McMillan, D., Speight, A. and Apgwilym, O. (2000), “Forecasting UK stock market volatility”,
Applied Financial Economics, Vol. 10 No. 4, pp. 435-448.
McMillan, D.G. and Garcia, R.Q. (2009), “Intra-day volatility forecasts”, Applied Financial
Economics, Vol. 19 No. 8, pp. 611-623.
Mincer, J. and Zarnowitz, V. (1969), “The evaluation of economic forecasts”, in Mincer, J. (Ed.),
Economic Forecasts and Expectations, National Bureau of Economic Research, New York,
NY, pp. 3-46.
Molnár, P. (2012), “Properties of range-based volatility estimators”, International Review of
Financial Analysis, Complexity and Non-Linearities in Financial Markets: Perspectives from
Econophysics, Vol. 23 No. 1, pp. 20-29.
Nelson, D.B. (1991), “Conditional heteroskedasticity in asset returns: a new approach”,
Econometrica, Vol. 59 No. 2, pp. 347-370.
Ng, H.G. and McAleer, M. (2004), “Recursive modelling of symmetric and asymmetric volatility in
the presence of extreme observations”, International Journal of Forecasting, Vol. 20 No. 1,
pp. 115-129.
Ñíguez, T.M. (2008), “Volatility and VaR forecasting in the Madrid Stock Exchange”, Spanish
Economic Review, Vol. 10 No. 3, pp. 169-196.
Pagan, A.R. and Schwert, G.W. (1990), “Alternative models for conditional stock volatility”,
Journal of Econometrics, Vol. 45 Nos 1/2, pp. 267-290.
Parkinson, M. (1980), “The extreme value method for estimating the variance of the rate of return”,
The Journal of Business, Vol. 53 No. 1, pp. 61-65.
Patton, A.J. (2011), “Volatility forecast comparison using imperfect volatility proxies”, Journal of
Econometrics, Vol. 160 No. 1, pp. 246-256.
Politis, D.N. and Romano, J.P. (1994), “The stationary bootstrap”, Journal of the American
Statistical Association, Vol. 89 No. 428, pp. 1303-1313.
Poon, S.H. and Granger, C.W.J. (2003), “Forecasting volatility in financial markets: a review”,
Journal of Economic Literature, Vol. 41 No. 2, pp. 478-539.
 SEF Rogers, L.C.G. and Satchell, S.E. (1991), “Estimating variance from high, low and closing prices”,
The Annals of Applied Probability, Vol. 1 No. 4, pp. 504-512.
32,4
Schwert, G.W. (1989), “Why does stock market volatility change over time?”, The Journal of
Finance, Vol. 44 No. 5, pp. 1115-1153.
Shu, J. and Zhang, J.E. (2006), “Testing range estimators of historical volatility”, Journal of Futures
Markets, Vol. 26 No. 3, pp. 297-313.
462 Taylor, J.W. (2004), “Volatility forecasting with smooth transition exponential smoothing”,
International Journal of Forecasting, Vol. 20 No. 2, pp. 273-286.
Taylor, S. (1986), Modelling Financial Time Series, John Wiley and Sons, Chichester.
Taylor, S.J. (1994), “Modeling stochastic volatility: a review and comparative study”,
Mathematical Finance, Vol. 4 No. 2, pp. 183-204.
Todorova, N. (2012), “Volatility estimators based on daily price ranges versus the realized range”,
Applied Financial Economics, Vol. 22 No. 3, pp. 215-229.
Ulu, Y. (2005), “Out-of-sample forecasting performance of the QGARCH model”, Applied Financial
Economics Letters, Vol. 1 No. 6, pp. 387-392.
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

Wei, W. (2002), “Forecasting stock market volatility with non-linear GARCH models: a case for
China”, Applied Economics Letters, Vol. 9 No. 3, pp. 163-166.
Yang, D. and Zhang, Q. (2000), “Drift independent volatility estimation based on high, low, open,
and close prices”, The Journal of Business, Vol. 73 No. 3, pp. 477-492.
Zakoian, J.M. (1994), “Threshold heteroskedastic models”, Journal of Economic Dynamics and
Control, Vol. 18 No. 5, pp. 931-955.
 Appendix 1 Stock index
volatility with
GARCH
Index name Country Ticker Region

Amsterdam Exchange Index The Netherlands AEX Europe

ASX all ordinaries index Australia AORD Asia-Pacific
Austrian Traded Index Austria ATX Europe
463
Euronext BEL-20 index Belgium BEL20 Europe
Bombay Stock Exchange Sensitive Index India SENSEX Asia-Pacific
Bovespa Index Brazil IBOV Americas
Euronext Paris CAC40 Index France CAC Europe
London Stock Exchange FTSE 100 U.K. UKX Europe
Athex Composite Share Price Index Greece GD Europe
Deutsche Borse AG German Stock Index Germany DAX Europe
S&P 500 Index USA SPX Americas
S&P TSX Composite Index Canada SPTSX Americas
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

Hang Seng Index Hong Kong HSI Asia-Pacific

Jakarta Stock Exchange Composite Index Indonesia JCI Asia-Pacific
Korea Composite Stock Price Index South Korea KOSPI Asia-Pacific
Merval Index Argentina MERVAL Americas
Nikkei 225 Index Japan NKY Asia-Pacific
Shanghai Composite Index China SHCOMP Asia-Pacific
Swiss market Index Switzerland SMI Europe
Straits times Index Singapore STI Asia-Pacific Table AI.
Taiwan Stock Exchange Weighted Index Taiwan TWSE Asia-Pacific Sample stock indices

About the authors

Prateek Sharma is a Doctoral Student at the Indian Institute of Management, Lucknow. Prior to
that, worked with the consumer banking division at Citibank. He has a BTech in Computer
Science and Masters in Management from the Indian Institute of Management, Lucknow. He is a
certified FRM holder and a member of the Global Association of Risk Professionals (GARP).
Prateek Sharma is the corresponding author and can be contacted at: prateeksharma1985@
gmail.com
Vipul, after doing his Masters in Physics and Masters in Management (Indian Institute of
Management, Ahmedabad, India), worked for six years in the industry. Subsequently, he acquired
professional qualification in Accounting, and Doctorate in Finance, and changed over to the
Academics 31 years back. Currently, he is a Professor of Finance at the Indian Institute of
Management, Lucknow, India, where he has been for the past 23 years.

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: [email protected]
 This article has been cited by:

1. Paul Quang, Tony Klein, Nam Nguyen, Thomas Walther. 2018. Value-at-Risk for South-East
Asian Stock Markets: Stochastic Volatility vs. GARCH. Journal of Risk and Financial Management
11:2, 18. [Crossref]
2. Olena Liashenko, Tetyana Kravets, Kateryna Krytsun. Software Packages for Econometrics:
Financial Time Series Modeling 188-208. [Crossref]
3. PaulSamit, Samit Paul, SharmaPrateek, Prateek Sharma. 2017. Improved VaR forecasts using
extreme value theory with the Realized GARCH model. Studies in Economics and Finance 34:2,
238-259. [Abstract] [Full Text] [PDF]
4. WaltherThomas, Thomas Walther. 2017. Expected shortfall in the presence of asymmetry and
long memory. Pacific Accounting Review 29:2, 132-151. [Abstract] [Full Text] [PDF]
5. LiHe, He Li, YuZhixiang, Zhixiang Yu, ZhangChuanjie, Chuanjie Zhang, ZhangZhuang, Zhuang
Zhang. 2017. Determination of China’s foreign exchange intervention: evidence from the Yuan/
Dollar market. Studies in Economics and Finance 34:1, 62-81. [Abstract] [Full Text] [PDF]
6. Mikhail Stolbov, Maria Shchepeleva. 2016. Financial stress in emerging markets: Patterns, real
Downloaded by INSEAD At 06:41 23 June 2018 (PT)

effects, and cross-country spillovers. Review of Development Finance 6:1, 71-81. [Crossref]