Professional Documents
Culture Documents
Article information:
Access to this document was granted through an Emerald subscription provided by emerald-srm:213039 []
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.
The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1526-5943.htm
JRF
11,2
164
Alper Ozun
School of Management, Bradford University, Bradford, UK
Atilla Cifter
Sekerbank, Istanbul, Turkey, and
Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)
Sait Ylmazer
Tekstilbank, Istanbul, Turkey
Abstract
Purpose The purpose of this paper is to use filtered extreme-value theory (EVT) model to forecast
one of the main emerging market stock returns and compare the predictive performance of this model
with other conditional volatility models.
Design/methodology/approach This paper employs eight filtered EVT models created with
conditional quantile to estimate value-at-risk (VaR) for the Istanbul Stock Exchange. The
performances of the filtered EVT models are compared to those of generalized autoregressive
conditional heteroskedasticity (GARCH), GARCH with student-t distribution, GARCH with skewed
student-t distribution, and FIGARCH by using alternative back-testing algorithms, namely, Kupiec
test, Christoffersen test, Lopez test, Diebold and Mariano test, root mean squared error (RMSE), and
h-step ahead forecasting RMSE.
Findings The results indicate that filtered EVT performs better in terms of capturing fat-tails in
stock returns than parametric VaR models. An increase in the conditional quantile decreases h-step
ahead number of exceptions and this shows that filtered EVT with higher conditional quantile such as
40 days should be used for forward looking forecasting.
Originality/value The research results show that emerging market stock return should be
forecasted with filtered EVT and conditional quantile days lag length should also be estimated based
on forecasting performance.
Keywords Stock returns, Emerging markets, Risk assessment, Stock exchanges, Turkey
Paper type Research paper
Introduction
Estimating losses of financial instruments has become a crucial task for market risk
management in the current global economy. The importance of that task is more
critical in emerging financial markets where fluctuations in the volume of hot money
from international portfolio investments and hedge funds, unstable regulatory and
political environment, and the lack of informational efficiency create high volatility and
extremes in returns.
Complex and volatile market conditions in emerging markets require dynamic and
flexible econometric models that are able to capture extreme changes in financial
variables. In this research paper, we use filtered (conditional quantile) expected
shortfall as filtered extreme-value approach for value-at-risk (VaR) estimation to
capture the extremes in the returns. Extreme-value theory (EVT) do not follow the
central limit theorem in mathematics arguing that if the sum of the variables has a
finite variance, then it follows Gaussian distribution. EVT focuses on the extremes
rather than mean.
There are some important reasons to choose EVT against parametric volatility
models. First, the distribution of returns is heavy-tailed and asymmetric in most
financial time series. EVT which approximates the area under the tail asymptotically
might be more powerful than imposing an explicit functional form. What is
more, extremes in returns might be caused by mechanisms that are structurally
different from the usual dynamics of financial markets. For example, extremes might
be the result of a major default or a speculative bubble. In those extreme conditions, the
distributional characteristics of financial time series might shift and require the
separation of tail estimation from the estimation of the rest of the distribution
(Neftci, 2000).
EVT has been used in financial risk estimation in recent years. The originality of
our paper is that we use conditional quantile expected shortfall with different lags and
compare them to find the optimal model capturing the extremes. As another original
work, we also apply h-step-ahead root mean square error (RMSE) and number of
exceptions to measure performance of the filtered expected shortfall. The performance
of the model is also empirically compared with those of the parametric models with
Kupiec (1995), Lopez (1999), Christoffersen (1998) and Diebold and Mariano (1995)
back-test algorithm.
We use the time series of daily returns of the Istanbul Stock Index-100 (Istanbul
Stock Exchange (ISE)-100) from January 2, 2002 to April 18, 2007 for our empirical
research. As an emerging market with dramatic macroeconomic and regulatory
changes in recent years, the ISE-100 gives us an opportunity to work with a highly
volatile and heavy-tailed data set. The back-test results show that the filtered expected
shortfall has superior performance in estimating the extremes and presents a new,
dynamic, and flexible perspective in VaR estimation.
Theoretical framework
Estimating VaR has become a crucial task of risk management functions of banks and
financial institutions since the Basel Committee stated that banks should be able to
cover losses on their trading portfolios over a ten-day horizon, 99 percent of the time.
However, classical VaR models focus on the whole empirical distribution of the returns
rather than that of extreme returns. On the other hand, managing extreme risk requires
estimation of quantiles that usually are not directly captured from the time series data.
EVT is presented as major alternative for generalized autoregressive conditional
heteroskedasticity (GARCH) models. GARCH approach dynamically improves itself in
recent years. Christoffersen and Jacobs (2004) empirically display that asymmetric
GARCH model that captures the leverage effect provides best estimation within all
GARCH models. Bekaert and Wu (2000) show that the leverage effect in equities
determines a strong negative correlation between returns and volatility, therefore,
creates a crucial source of skewness in returns. As a recent model, the normal mixture
GARCH (NM-GARCH) model is presented by Alexander and Lazar (2006). In the
NM-GARCH model, the individual variances are only tied with each other through
their dependence on the error term, thus the model captures switching shocks and
long-term memory in the stock returns (Ozun and Cifter, 2008).
165
JRF
11,2
166
with methodological concerns can be followed in the works of Embrechts et al. (1997)
and Focardi and Fabozzi (2003, 2004).
With Turkish data, Cifter et al. (2007) use conditional quantile expected shortfall
and generalized Pareto distribution (GPD) for interest rates and find that conditional
EVT improves forecasting. Gencay et al. (2003), Gencay and Selcuk (2004), Altay and
Kucukozmen (2006) and Eksi et al. (2006) use EVT with unconditional quantile EVT to
estimate fat-tails in the stock returns in Turkey and they find that EVT performs better
than classical VaR models.
Filtered extreme-value theory
VaR reflects the change in a portfolio with a confidence level on a time period. In this
description, DPDt, measures changes in the market value of portfolio P on time Dt
period with probability p (Dowd, 2004):
PDPDt # VaR p
0 # y # xF 2 u
u!1
For negative returns, under the assumptions that x u y, the tail estimation can be
computed with equation (3):
!
^
^Fx 1 2 N u 1 j x 2 u21=j^
3
s
n
In this equation, n is the total number of observations in the data set, Nu is the
violations (extremes) above u. For pre-defined p . F(u) distribution, VaR for one day,
VaR( p%), is calculated with the equation (4):
^ !
s
n 2j
VaRGPD u
p 21
4
Nu
j^
In equation (4), by defining u with either constant or conditional quantile, GPD with
constant or conditional quantile is obtained. From similar perspective, Artzner et al.
(1999) evaluate expected shortfall as an alternative for VaR. In expected shortfall, the
expected value of the portfolio return is taken into consideration if there is a violation.
167
JRF
11,2
Expected shortfall can be constructed with the following equation (Gilli and
Kellezi, 2002):
ESp VaRp EX 2 VaRp jX . VaRp
168
The second nomination on the right side of the equation is the mean of the violation
distribution of FVaRp( y) on the VaRp threshold. For GPD, we can express the mean
violation function with j , 1 parameters as in the equation (6):
eu EX 2 ujX . u
s ju
;
12j
s ju . 0
VaRp s^ 2 j^u
1 2 j^ 1 2 j^
If X is GPD; then for all r , 1/j integer, (r), the first moment of each r exists.
In this research paper, eight different filtered expected shortfall estimation with two,
three, four, five, ten, 15, 20, and 40 days rolling quantile are estimated. The rolling
quantile days are randomly selected and maximum rolling is estimated as 40 days
since conditional EVT approximate to unconditional EVT more than 40 days rolling.
We use parametric models like GARCH, GARCH-t, GARCH-skewed student-t, and
FIGARCH for performance comparison of the filtered expected shortfall. The
methodologies for GARCH models are not examined here, but detailed examinations
can be found in Chung (1999), Baillie et al. (1996), Davidson (2002) and Laurent and
Peters (2002).
Methodologies of alternative back-tests
Alternative back-testing algorithms are employed to compare the performance of the
models. Kupiec (1995) test, Christoffersen (1998) test, Lopez (1999) test, RMSE (70
days), and h-step ahead forecasting RMSE (70 days), number of exceptions, h-step
ahead number of exceptions and Diebold and Mariano (1995) encompassing test are
used as the back-tests.
Kupiec test
Kupiec (1995) test defines the failure ratio ( f ) as the excess values from VaR(x) to the
total observations (T). When we nominate the pre-defined VaR with a, likelihood ratio
statistics with x 2 distribution for the Kupiec (1995) test can be given in equation (9):
LR 2{logj f x 1 2 f T2x j 2 logja x 1 2 aT2x j}
Christoffersen test
According to Christoffersen (1998) test, the probability of failure rate in the VaR
estimation is the important point for back-testing. To conduct the test, one should first
define p a Pr yt , VaRt a and test H0 : p a a against H1 : p a a.
LR
22 InLa d
!x1
^
Lp
11
Lopez test
Lopez (1999) performs the back-test in three steps. In the first step, the proper
distribution of returns and statistics model is chosen. Second, employing the model
created in the first step and using historical losses/gains VaR and VaR(a) are
constructed and Li, a mean for losses/gains is obtained. In the last step, the process
described above is repeated many times, for example 10,000, in order to reach an
estimated value for a mean of loss distribution, Li 1i 10,000.
Lopez (1999) defines the violation function LVaRta ; xt;t1 that can be given in
equation (12):
1 xt;t1 2 VaRta 2
0
if
if
xt;t1 . VaRta
xt;t1 # VaRta
12
13
169
JRF
11,2
170
d
DM r
P
2
P 22 dt1 2 d
15
Data
ISE Rate (ISE-100 Index) is received from Bloomberg. Our dataset covers 1,325 daily
observations including 610 observations with negative returns from January 2, 2002 to
April 18, 2007. We constituted the series in log-differenced level. Figure 1 shows ISE
Index in log-differenced series where Figure 2 shows negative and positive returns
separately. By performing augmented Dickey-Fuller (Dickey and Fuller, 1981) and
0.15
Ise
0.1
0.05
0
54 107 160 213 266 319 372 425 478 531 584 637 690 743 796 849 902 955 1,0081,0611,1141,1671,2201,273
0.05
0.1
Figure 1.
ISE log-differenced series
0.15
0
0.14
0.02
0.12
0.04
0.1
0.06
0.08
0.08
0.06
0.1
Figure 2.
ISE negative and positive
returns
0.04
0.12
0.14
0.16
0.02
Ise negative returns
0
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714
Phillips-Peron test (Philips and Perron, 1988) we found that ISE Index is stationary at
log differenced level as shown in Table I.
Main statistical properties of the log-differenced index is shown in Table I.
Although x 2 normality test and Jargue-Bera statistics indicate that the index is
normally distributed kurtosis and skewness values show that distribution is skewed
and heavy-tailed. As a result, filtered EVT, like filtered expected shortfall, may capture
tail loss better compared to GARCH and alternative GARCH models.
171
Empirical evidence
Filtered expected shortfall with conditional quantile of two, three, four, five, ten, 15, 20,
and 40 days rolling and volatility models as GARCH(1,1), GARCH(1,1)-student-t,
GARCH(1,1)-skewed student-t, and FIGARCH(1,1) are estimated. Kupiec (1995), Lopez
(1999) and Christoffersen (1998) backtesting procedures and h-step ahead forecasting
of RMSE and number of exceptions are applied to compare predictive performance of
the models. Back-testing is done with 95 and 99 percent confidence intervals and Basel
guidelines require using a 99 percent confidence interval.
Table II reports GARCH(1,1), GARCH(1,1)-student-t, GARCH(1,1)-skewed student-t,
and FIGARCH(1,1) estimates for ISE index. All the parameters of GARCH models are
statistically significant and according to log-likelihood stat GARCH(1,1) fits better than
other GARCH models. Table III reports filtered expected shortfall shape (j) and scale
(b) parameters of lower and upper tail. In this paper, we only estimate lower tail of VaR
estimation based on GARCH and filtered expected shortfall models so that only lower
tail parameters are used to estimate filtered expected shortfall models[1].
Figures 3 and 4 show filtered expected shortfall and GARCH models graphs. Also
reported in Table IV Diebold-Mariano test (Diebold and Mariano, 1995) shows that
GARCH student-t and GARCH-skewed student-t are not statistically different from
GARCH with Gaussian distribution where FIGARCH, filtered expected shortfall with
two, three, four, and five days conditional quantile are statistically different at five
percent confidence interval and filtered expected shortfall with ten, 15, 20, and 40 days
ISE
Unit root tests
ADF test
P-P test
Main stats.
Asymptotic test: (x 2)
Normality test: (x 2)
Mean (m)
SD (s)
Skewness (S)
Kurtosis (K)
Minimum
Maximum
Jarque-Bera statistic
Notes: *Statistical significance at 5 percent level (at least); at-value
236.5121 *
236.5267 *
876.03 [0.0000]a
399.71 [0.0000]a
0.00091553 C
0.0212336
20.0707095
6.98092
20.133408 at obs. 289
0.11794 at obs. 215
876.031
Table I.
Unit root test and main
statistical properties
JRF
11,2
172
Table II.
Estimation results from
volatility models
v
a
b1
n-student-t
j-Ske.
n-Skew
d
a b1
Loglike
GARCH
GARCH-t
GARCH-skew
FIGARCH
0.001 * * (2.738)
0.089 * * (5.325)
0.889 * * (44.21)
0.97840
3,321.90
0.001 * * (3.395)
0.084 * * (4.500)
0.886 * * (37.29)
8.125 * * (5.314)
0.97146
3,349.99
0.0014 * (2.912)
0.085 * * (4.490)
0.884 * * (36.15)
2 0.0606 (1.505)
8.215 * * (5.266)
0.96958
3,351.13
0.001 * * (2.973)
0.204 * (2.853)
0.600 * * (6.051)
0.5043 (6.022)
0.80434
3,325.44
Lower tail
Shape (j)
Scale (b)
Models
Table III.
Extreme-value
parameters
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
ES
ES
ES
ES
ES
ES
ES
ES
2 days
3 days
4 days
5 days
10 days
15 days
20 days
40 days
0.064
0.081
0.114
0.034
0.170
0.131
0.139
0.184
0.011
0.011
0.009
0.011
0.010
0.011
0.010
0.009
Upper tail
Shape (j)
Scale (b)
0.072
0.115
0.134
0.187
0.300
0.238
0.226
0.348
0.011
0.010
0.010
0.008
0.007
0.009
0.009
0.008
0.0000000
0.0500000
0.1000000
0.1500000
Figure 3.
Filtered expected shortfall
models
0.2000000
Ise
FES 3 days
FES 5 days
FES 15 days
FES 40 days
FES 2 days
FES 4 days
FES 10 days
FES 20 days
0.2500000
0
1
23 45 67 89 111 133 155 177 199 221 243 265 287 309 331 353 375 397 419 441 463 485 507 529 551 573 595
0.02
0.04
0.06
173
0.08
0.1
Downloaded by UNIVERSITAT DE VALENCIA At 05:03 17 July 2015 (PT)
Ise
0.12
Garch
Garch-t
0.14
Garch-skew
Figure 4.
GARCH models
Figarch
0.16
Models
GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days
Ratio DM
0.5092
0.5194
2 0.09754 *
0.094686 *
0.148612 *
0.191513 *
0.185461 *
0.32498 *
0.047025 * *
0.077021 * *
0.085852 * *
40 days conditional quantile performs best based on Christoffersen and Kupiec tests.
According to 99 percent confidence interval, filtered expected shortfall with ten and
20 days conditional quantile performs best based on Lopez test, filtered expected
shortfall with two days conditional quantile performs better based on Christoffersen
test, filtered expected shortfall with 15 and 40 days conditional quantile are the best
ones based on Kupiec tests. Since 99 percent confidence interval is more significant and
Basel guidelines require using 99 percent confidence interval, we also consider
99 percent confidence interval for back testing results. According to all back testing
procedures filtered expected shortfall models predictive performance is better than
GARCH models. There is not one filtered expected shortfall model that beats other
models based on Lopez, Christoffersen; and Kupiec tests therefore we applied h-step
ahead forecasting of RMSE and number of exceptions. Based on h-step ahead
forecasting of RMSE GARCH(1,1) is the best one[2]. Table VII shows that based on
h-step ahead forecasting of number of exceptions up to 70 days filtered expected
Table IV.
Comparing predictive
accuracy with the
Diebold-Mariano statistic
JRF
11,2
174
Table V.
Back testing
Models
GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days
Lopez
Christoffersen
Kupiec
0.68934
1.00409
0.78774
0.89262
0.15879
0.52972
0.85641
0.98265
1.41345
1.74404
1.41345
1.74404
0.00315
0.00025
0.00146
0.00063
0.17591
0.00958
0.00065
0.00021
0.00000
0.00000
0.00000
0.00000
0.00249
0.00023
0.00120
0.00054
0.11457
0.00766
0.00061
0.00022
0.00000
0.00000
0.00000
0.00000
Lopez
Christoffersen
Kupiec
0.32134
0.15609
0.25970
0.20462
0.66599
0.23692
0.08361
0.04985
0.00065
0.01124
0.00065
0.01124
0.00080
0.01469
0.00226
0.00595
0.00000
0.00396
0.06909
0.15074
0.85707
0.42083
0.85707
0.42083
0.99980
0.99601
0.99943
0.99846
0.99999
0.99907
0.98078
0.95528
0.68437
0.32276
0.68437
0.32276
Models
Table VI.
Back testing
GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days
Note: 99 percent confidence interval
Models
Table VII.
H-step ahead forecasting
GARCH
GARCH-t
GARCH-skew
FIGARCH
Filtered ES 2 days
Filtered ES 3 days
Filtered ES 4 days
Filtered ES 5 days
Filtered ES 10 days
Filtered ES 15 days
Filtered ES 20 days
Filtered ES 40 days
No. of exceptions
t1
Avrg.
15
13
14
14
22
16
13
12
8
5
8
4
15.136
11.454
13.500
13.257
25.196
17.575
14.515
13.712
7.696
5.7272
6.500
3.0151
RMSE
t1
Avrg.
0.0373
0.0405
0.0383
0.0389
0.0449
0.0481
0.0501
0.0498
0.0586
0.0602
0.06051
0.0614
0.0402
0.0433
0.0411
0.0419
0.0479
0.0515
0.0537
0.0533
0.0617
0.0630
0.0633
0.0640
Note: Average no. of exceptions and RMSE is estimated with 70 days step-ahead forecasting
175
Conclusions
As a result of the dynamic and chaotic features of financial markets in emerging
economies, financial forecasting is almost impossible with parametric models.
Observed extremes and fat tails in returns need to be estimated with relatively more
sophisticated models. Parametric models have certain strict assumptions on the
distribution function of returns. These assumptions, either normality or asymmetric
distribution, are not able to make statistically significant estimations.
40
35
Garch
Figarch
FES4
FES15
Garch-t
FES2
FES5
FES20
Garch-skew
FES3
FES10
FES40
30
No. of exceptions
shortfall with 40 days conditional quantile is the best one. We observed that an
increase in the conditional quantile decreases h-step ahead forecasting of number of
exceptions and this shows that filtered expected shortfall with 40 days conditional
quantile should be used for forward looking forecasting such as more than one month
forecasting. Christoffersen and Diebold (2000) show that volatility models such as
GARCH and other GARCH models can be used for forecasting up to 15-20 days ahead
for the USA financial instruments and Cifter (2004) shows that volatility models can be
used for forecasting up to 10-14 days ahead for Turkish interest rates.
Figure 5 shows h-step ahead forecasting of number of exceptions and Figures 6 and 7
show h-step ahead forecasting of RMSE up to 70 steps. The h-step ahead forecasting of
number of exceptions shows that filtered expected shortfall from 15 to 40 days
conditional quantile beats all GARCH models and filtered expected shortfall less than
15 days conditional quantile. Diebold-Mariano test of equal forecast accuracy (Diebold
and Mariano, 1995) is also applied to reveal statistical difference between filtered
expected shortfall models and found that filtered expected shortfall models with less
than ten days conditional quantile are not statistically different than two days
conditional quantile estimation (Table VIII), thus indicating that filtering less than ten
days conditional quantile may imitate GARCH models.
25
20
15
10
5
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
H-step ahead forecasting
Note: No. of exceptions
Figure 5.
H-step ahead forecasting
JRF
11,2
0.05
0.045
0.04
176
RMSE
0.035
0.03
0.025
0.02
Garch
Garch-t
0.01
Garch-skew
0.005
Figarch
0
1
Figure 6.
H-step ahead forecasting
7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70
H-step ahead forecasting
RMSE
0.015
0.04
0.03
0.02
0.01
FES2
FES3
FES4
FES5
FES10
FES15
FES20
FES40
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 67 69
Figure 7.
H-step ahead forecasting
EVT, on the other hand, employs the central limit theorem for risk estimation.
According to the theorem, if the sum of the variables has a finite variance, then it
follows Gaussian distribution. The distribution of extremes in returns is limited to
having the same form without relying on the distribution of the parent variable.
In this research paper, we investigate filtered EVT with different rolling quantile to
estimate VaR. By using daily returns of the ISE National 100 Index, we estimate risk
with filtered EVT. For comparison of the model performance, we also estimate VaR with
Models
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Filtered
Ratio DM
ES
ES
ES
ES
ES
ES
ES
ES
2 days rolling
3 days rolling
4 days rolling
5 days rolling
10 days rolling
15 days rolling
20 days rolling
40 days rolling
1.2342
0.612275
0.627426
0.323022 *
0.264849 *
0.309795 *
0.297323 *
Notes: Benchmark model is filtered expected shortfall with two days conditional quantile;
*significance at 5 percent confidence level
177
Table VIII.
Comparing predictive
accuracy with the
Diebold-Mariano statistic
JRF
11,2
178
Blum, P., Dias, A. and Embrechts, P. (2002), The art of dependence modeling: the latest
advances in correlation analysis, in Lane, M. (Ed.), Alternative Risk Strategies, Risk
Books, London.
Christoffersen, P. (1998), Evaluating interval forecasts, International Economic Review, No. 39,
pp. 841-62.
Christoffersen, P. and Diebold, F.X. (2000), How relevant is volatility forecasting for financial
risk management?, Review of Economics and Statistics, Vol. 82, pp. 12-22.
Christoffersen, P. and Jacobs, K. (2004), Which GARCH model for option valuation?,
Management Science, Vol. 50, pp. 1204-21.
Chung, C.F. (1999), Estimating the fractionally integrated GARCH model, working paper,
National Taiwan University, Taipei.
Cifter, A. (2004), Asymmetric and fractionally integrated GARCH models with (skewed)
student-t and GED distribution in risk management, Proceedings of 8th National Finance
Symposium, Istanbul Technical University, Istanbul.
Cifter, A., Ozun, A. and Yilmazer, S. (2007), Value-at-risk with expected shortfall and
generalized Pareto distribution: some comparative evidence for Turkish interest rate
markets, Banks Association of Turkey-Bankers Journal, Vol. 60, pp. 3-16.
Davidson, J. (2002), Moment and memory properties of linear conditional heteroscedasticity
models, working paper, available at: www.cf.ac.uk/carbs/econ/davidsonje/ (accessed on
April 10, 2007).
Dickey, D.A. and Fuller, W.A. (1981), Likelihood ratio statistics for autoregressive time series
with a unit root, Econometrica, Vol. 49, pp. 1057-72.
Diebold, F. and Mariano, R.S. (1995), Comparing predictive accuracy, Journal of Business &
Economics Statistics, Vol. 13 No. 3, pp. 253-63.
Dowd, K. (2004), FOMC forecasts of macroeconomic risks, Occasional Papers, 12, Industrial
Economics Division.
Eksi, Z., Yildirim, I. and Yildirak, K. (2006), Alternative risk measures and extreme value theory
in finance: implementation on ISE 100 index, Proceedings of ICBME, Yasar University, Izmir.
Embrechts, P., Kluppelberg, C. and Mikosch, T. (1997), Modeling Extremal Events for Insurance
and Finance, Springer, Berlin.
Focardi, S. and Fabozzi, F. (2003), Fat tails, scaling and stable laws, Journal of Risk Finance,
Vol. 5 No. 1, pp. 25-36.
Focardi, S. and Fabozzi, F. (2004), The Mathematics of Financial Modeling and Investment
Management, Wiley, New York, NY.
Gencay, R. and Selcuk, F. (2004), Extreme value theory and value-at-risk: relative performance
in emerging markets, International Journal of Forecasting, Vol. 20, pp. 287-303.
Gencay, R., Selcuk, F. and Ulugulyagci, A. (2003), High volatility, thick tails and extreme value
theory in value-at-risk estimation, Insurance: Mathematics and Economics, Vol. 33,
pp. 337-56.
Gilli, K. and Kellezi, E. (2002), Portfolio optimization with VaR and expected shortfall,
in Kontoghiorghes, E.J., Rustem, B. and Siokos, S. (Eds), Computational Methods in
Decision-making, Economics and Finance, Kluwer Applied Optimization Series, Kluwer
Academic, New York, NY, pp. 167-83.
Gupta, A. and Liang, B. (2005), Do hedge funds have enough capital? A value-at-risk approach,
Journal of Financial Economics, Vol. 77, pp. 35-47.
Inui, K. and Kijima, M. (2005), On the significance of expected shortfall as a coherent risk
measure, Journal of Banking & Finance, Vol. 29, pp. 853-64.
Kalyvas, L., Steftos, A., Sriopoulos, C. and Georgopoulos, A. (2007), An investigation of
riskiness in South and Eastern European markets, International Journal of Financial
Services Management, Vol. 2 Nos 1/2, pp. 13-21.
Kuester, K., Mittik, S. and Paolella, M.S. (2005), Value-at-risk prediction: a comparison of
alternative strategies, Journal of Financial Econometrics, Vol. 4 No. 1, pp. 53-89.
Kupiec, P.H. (1995), Techniques for verifying the accuracy of risk measurement models,
Journal of Derivatives, Vol. 3, pp. 73-84.
Laurent, S. and Peters, J.P. (2002), G@rch 2.2: an Ox package for estimating and forecasting
various ARCH models, Journal of Economic Survey, Vol. 16, pp. 447-85.
Lhabitant, F.S. (2002), Hedge Funds: Myths and Limits, Wiley, New York, NY.
Liang, B. and Park, H. (2007), Risk measures for hedge funds: a cross-sectional approach,
European Financial Management, Vol. 13 No. 2, pp. 333-70.
Longin, F.M. (2000), From value at risk to stress testing: the extreme value approach, Journal of
Banking & Finance, Vol. 24, pp. 1097-130.
Lopez, J.A. (1999), Methods for evaluating value-at-risk models, Federal Reserve Bank of
San Francisco Economic Review, Vol. 2, pp. 3-17.
McNeil, A.J. and Frey, R. (2000), Estimation of tail-related risk measures for heteroscedastic
financial time series: an extreme value approach, Journal of Empirical Finance, Vol. 7,
pp. 271-300.
Martinz, F.C. and Yao, F. (2006), Estimation of value-at-risk and expected shortfall based on
nonlinear models of return dynamics and extreme value theory, Studies in Nonlinear
Dynamics and Econometrics, Vol. 10 No. 2, pp. 107-49.
Neftci, S. (2000), Value at risk calculations, extreme events, and tail estimation, Journal of
Derivatives, Spring, pp. 12-19.
Ozun, A. and Cifter, A. (2008), A signal processing model for time series analysis: the effects of
international F/X markets on domestic currencies using wavelet networks, International
Review of Electrical Engineering, Vol. 3 No. 3, pp. 580-91.
Philips, P.C.B. and Perron, P. (1988), Testing for a unit root in time series regression,
Biometrica, Vol. 75, pp. 335-446.
Poon, S., Rockinger, M. and Tawn, J. (2004), Extreme value dependence in financial markets:
diagnostics, models and financial implications, The Review of Financial Studies, Vol. 17
No. 2, pp. 586-610.
Saltoglu, B. (2003), A High Frequency Analysis of Financial Risk and Crisis: An Empirical Study
on Turkish Financial Market, Yaylim Press, Istanbul.
Tolikas, K. and Brown, R. (2006), The distribution of the extreme daily share returns in the
Athens Stock Exchange, The European Journal of Finance, Vol. 12 No. 1, pp. 1-22.
Yamai, Y. and Yoshiba, T. (2005), Value-at-risk versus expected shortfall: a practical
perspective, Journal of Banking & Finance, Vol. 29 No. 4, pp. 997-1015.
Corresponding author
Alper Ozun can be contacted at: A.ozun@Brad.ac.uk
To purchase reprints of this article please e-mail: reprints@emeraldinsight.com
Or visit our web site for further details: www.emeraldinsight.com/reprints
179
1. Pilar Abad, Sonia Benito, Carmen Lpez. 2014. A comprehensive review of Value at Risk methodologies.
The Spanish Review of Financial Economics 12, 15-32. [CrossRef]
2. Stavros Degiannakis, Christos Floros, Alexandra Livada. 2012. Evaluating valueatrisk models before
and after the financial crisis of 2008. Managerial Finance 38:4, 436-452. [Abstract] [Full Text] [PDF]