Comparison of Extreme Value Theory and GARCH models on

∗
Estimating and predicting of Value-at-Risk
Zuo-xiang Peng 1 2 Shi Li 1 Hao Pang 1
(1 School of Statistics, Southwestern University of Finance and Economics, Chengdu, 610074;
2 School of Mathematics and Finance, Southwestern University, Chongqing, 400715)

Abstract: Based on extreme value theory and General Pareto Distribution (GPD), the paper analyzes
and describes the performance of the thick-tail of the high frequency financial time series data with
Tail-Index which fitted by local fitness on tail distribution of the data. Both processes, the one is
procedures of estimating and testing of the tail index, the other is estimating and forecasting methods
of Value-at-Risk (VaR) are given systematically. The one-step forward forecasting results of the
Composite Index of Shanghai Stock Exchange by Extreme Value Theory and other well-known
modeling techniques, such as ARCH/GARCH models, are empirically compared and contrasted. The
empirical results argue that GPD method is superior to GARCH models on estimating and forecasting
of VaR.

Key Words: Value-at-Risk Tail-Index One-Step Forward Forecasting Extreme Value Theory
Empirical Analysis.

0 Introduction
The Value-at-Risk(VaR) method has two limitations in measuring the VaR of high frequency
financial time series data: on the one hand, the high frequency financial time series data often have the
characteristics of thick-tail, if the financial time data, such as stock returns, are still supposed
following normal distribution, the VaR would be underestimated. On the other hand, in the process of
empirical analysis, the hypothesis testing of returns time series follow the normal distribution, by
Jarque-Bera test, always be rejected, the problems of specification bias will be considered. Extreme
Value Theory (EVT) and ARCH/GARCH model have the different statistical characteristics to
describe the performance of the thick-tail properties of the high frequency financial time series data.
Embrechts, Kluppelberg and Mikosch(1997), Reiss and Thomas(1997)[1][2] have investigated the
application of the EVT on the financial and insurance industries systematically. In the series of articles,
Mcneil(1997,1998,1999, and 2000) studied the Tail Index and the estimation of the VaR for the
financial time series [3]-[7]. Silva & Mendes (2003), Agbeyegbe & Leon (2002)[8][9] estimated the Tail
Index and VaR for the Asian stock market and Jamaica stock market separately based on the EVT.
Gençay, Selçuk & Ulugülyaˇgci (2003) [10] , with the EVT, estimated the Tail Index of the ISE-100
returns series of Turkey stock market to gain the one-step estimating process of the VaR .
In general, the work mentioned above only limited their studies in the framework of EVT when
they studied the characteristics of the distribution function or density function with Tail Index, so it is
considered that the investigating angle of above literatures are less confined. If the analysis on tail
characteristics of the distribution function or the density function are used to evaluate the properties of
the random sampling data and the sensitivity to the large-sample observed numbers, the expected

∗
Project(70371061) supported by NSFC, Project (Grant: 03JB790011) supported by Doctorial Foundation of Ministry of
Education of the P.R. of China , Project supported by the tenth Five plan ‘211 project’ of Southwestern University of finance
& economics。
1
analysis will possess the existing conditions of the high order moments (or high order conditional
moments) in the unit root testing on financial high frequency time series with ARCH/ GARCH error
terms. In the meantime, the analysis on (high, lower) Tail-Index can not only reflect the intensity of
leverage effects, but also can be used to specify the fluctuating or volatility models. Based on the
those ideas, the paper attempts to establish the Tail Index estimation with EVT to describe the
thick-tail degree of the distribution function or density function, to estimate and forecast the VaR, and
to analyze the existing conditions and effects of the high order moments or high order conditional
moments.
The remainder of this paper is as follows: the second section offers the foundation for the EVT, the
process of estimating and testing for the Tail Index discussed in detail in the third section, the fourth
section is the estimating method and the forecasting method of the VaR; the fifth section empirically
analyzes the estimating and testing process of lower Tail-Index of the Shanghai stock market returns
rate series, and empirically estimate the VaR, in the fifth section, the one-step forward forecasting
outcomes separately based on the GARCH error term model and the EVT will be compared to draw
the conclusions.
1 The summary for the Extreme Value Theory
Suppose that X 1 , X 2 ,..., X n are independent identically distributed series data 1 , the common
distribution function is F ( x ) . If the constant an > 0 exists, bn ∈ R , and G ( x ) is non-degenerated
distribution function, then following limit exists:
lim P (max{ X i ,1 ≤ i ≤ n} ≤ an x + bn ) = G ( x )
n →∞
So function G ( x ) must be one a type of followings:
⎧ −
1
⎪exp(−(1 + ξ x) ξ ),1 + ξ x > 0, ξ ≠ 0
Gξ ( x) = ⎨
⎪ exp( −e − x ), x ∈ R, ξ =0
⎩
where ξ is called Tail Index, F ( x ) belongs to the attraction field Gξ ( x ) , and Gξ ( x ) is called the
types of extreme distribution function. [11][12]
Resnick(1987) proved that normal distribution, exponential distribution and lognormal

distribution all belong to the attraction field G0 ( x) , but the necessary and sufficient conditions of

1
−
ξ
1 − F ( x ) belonging to the attraction field Gξ ( x ) is 1 − F ( x) = x L( x ) , where L ( x ) is a slow-
changing function. T-distribution and skewed-t distribution also belong to the attraction field

Gξ ( x), ξ > 0 . So the Tail Index ξ is the important parameter to measure tail attenuation of the

distribution. When ξ > 0 , the distribution has thick-tail character. How to estimate ξ is the important
composing aspect of the EVT, and also is the important field to study the existing conditions for high
order moments of ARCH/GARCH model and the estimation and forecast of the VaR in the risk
management industry.
2 The estimating and testing of Tail Index
The estimating methods of the parameter ξ include Hill-type estimating method ( ξ > 0 ),
[13]

1
Financial time series data, such as returns data, is no-independent, but under some weak-consistency conditions, the
arguments in this section will be tenable (Leadbetter, Lindgren & Rootzen (1983)), in which the stationary of time series is
requested.
2
Pickands-type estimating method [14]（ ξ ∈ R ）and moment estimating method [15]（ ξ ∈ R ）and so on.
As to the returns series, the tails often take the token of few extra income or loss. Theoretically the
value of returns can be infinite, and in general the Tail Index ξ > 0 [16].
Based on the scatter plot of H (m, n) 2 , the value of Tail Index ξ can only be estimated roughly, if
the reliable estimating value of the Tail Index ξ is wanted to gain, then it is requested that the
process of estimation and hypothesis must be systematical. Considering the tail of local fitting
distribution function, the relative efficiency estimation of the Tail-Index, and based on that the
institution investors and disperse investors take more care of the changing conditions of the financial
time series when the observed value exceeds the threshold u in the process of making investment
decision, the GPD method will be used to estimate the Tail Index. Considering
F (u + y ) − F (u )
Fu ( y ) = P( X − u ≤ y | X > u ) = ,y >0
1 − F (u )
where u > 0 is the threshold. As to the sufficient large u ，and following condition will exists [19]：

⎧ −1
y y
⎪1 − (1 + ξ ) ξ ,1 + ξ > 0, ξ ≠ 0;
⎪ σu σu
Fu ( y ) ≈ G ( y;σ u , ξ ) = ⎨
⎪ −
y
⎪ 1− e , σu
ξ =0
⎩

where G ( y, σ , ξ ) called GPD，its tail distribution can be approximately presented as following

−1
N x−u ξ
1 − F ( x) = u (1 + ξ )
n σu

where N u is the numbers that the observed values exceed the threshold u ,and ξ , σ are the
parameters to estimate. In the empirical analysis, the threshold u can be selected by the scatter plot of
the sample mean exceedance, which is defined as {(u , en (u )), X 1, n < u < X n ,n } , where en (u ) is
called sample mean excess function (SME).
The definition of en (u ) is
n
∑ ( X i − u )+ ⎧X − u X > u ⎧1 X > u
en (u ) = i =1
, where:( X − u ) + = ⎨ ， I{ X >u} = ⎨
n
⎩0 X ≤u ⎩0 X ≤ u
∑ I{ X >u}
i
i =1

Here, en (u ) is the estimation of population truncation mean M (u ) = E ( X | X > u ) .If the

truncation distribution function Fu ( x ),( x = y + u ) follows GPD ( y, ξ , σ ) , then
M (u ) = (σ + ξ u ) /(1 − ξ ), σ + ξ u > 0
The above analytic conclusions show that Hill scatter plot, to begin with, can be used to estimate
the Tail-Index initially, then the scatter plot of SME is employed to analyze the characters of the tail

1 m
∑ log ( X n − k +1, n ) − log ( X n − m, n ) ，where X1, n ≤ X 2, n ≤ ... ≤ X n, n is
2
The definition of Hill’s estimator to estimate tail-Index ξ is H ( m, n ) =
m k =1

order statistic of X 1 , X 2 ,..., X n （Dekkers, Haan (1989) studied the asymptotic properties of H ( m, n) ）. The selection of m in the process of

empirical analysis is very difficult. In general, by observing time series figure of H ( m, n) , in which m = m(n) , then the appropriate estimator will

be selected in the range of which H (m, n) is roughly unchanged. Loretan and Phillips (1994) argued that m should be less than or equal to 0.1 × n .

3
distribution function. Based on those, if SME is the increasing function of u （because the slope
ξ > 0 ）, then the tail distribution can be estimated as thick tail roughly, and the relatively optimal
threshold value of u can be selected. In aid of this threshold u , the value of Tail-Index ξ can be
estimated with maximum likelihood method.
The testing of estimated Tail-Index with the GPD method can be gained by fitting methods,
such as fitting of truncation distribution, fitting of tail distribution and fitting of residual term, and so
on. Among those methods, (a) truncation distribution fitness is the fitting comparative analysis on

value of Fˆu ( x ) and G ( x, ξˆ,σˆ ) , where Fˆu ( x ) is the empirical distribution function corresponding to

the observed values exceed the threshold value of u . (b) Tail-distribution fitting analysis is the
−1 −1
N x − u ξˆ N x − u ξˆ
fitting comparisons of 1 − F ( x) = u (1 + ξˆ ) and G ( x, ξˆ, σˆ ) , where 1 − F ( x) = u (1 + ξˆ ) is
n σˆ u n σˆ u

an empirical tail distribution function, corresponding to the observed values exceed the threshold
value of u . Meanwhile, the fitting analysis of QQ plot, a plot of the observed data exceeding the

threshold u and function G ( x, ξˆ,σˆ ) , are used to compare and test the rationality of the tail

distribution specification. (c) The fitting testing of the residuals terms. The residual terms are defined
1 ξˆ x
as { log(1 + i ), xi > u} . If the population truncation distribution function is GPD distribution, then
ξˆ σˆ
the QQ plot of the residual term (compared with exponential distribution) should be close to a
straight line with the slope of 1.
3 The estimating and forecasting of the Value-at-Risk
The measuring of the risk can be described by the VaR, which is the greatest latent losses, given
the confidence level and the normal market. Suppose {rt } is the returns time series of some financial
asset or portfolio investment, α is significance level 3 , rt is the returns rate at time t 4 , then the
VaRt (α ) is defined as that it will meet 1-α of following probability:
P(rt ≥ VaRt (α ) Ψ t −1 ) = α
The 1 − α is the upper quantile, which reflects the abnormal liquidation of funds at financial
markets, the volatility of the financial asset prices and the latent loss. From view of statistics, the
value of VaR means the big fractile. Because the returns series often have the character of thick tail
and don’t obey the normal distribution, so the GPD method is adopted to estimate the VaR.
By estimating method of GPD for Tail-Index, as to the enough great threshold value u , the tail
distribution function of the returns rate data is:
−1
N x − u ξˆ
1 − F ( x) ≈ u (1 + ξˆ )
n σˆ

where F ( x) is the distribution function of the returns, N u is the sampling data numbers that the
observed data exceed the threshold value u with the sample size of n . So given the significance

3 Generally, the values of 1 − α are taken by 0.95、0.99、0.999。

4
When rt > 0 ， rt = 100 * log( pt / pt −1 ) , but when rt < 0 ， rt takes as {− rt } 。
4
level α , the estimation of the VaR is:
σˆ nα ˆ
VaR(α ) = u + [( ) −ξ − 1]
ξˆ N u
According to the different threshold value, the change process of the VaR can be estimated, and
the point and interval estimation of VaR can also be gained.
The other method to estimate and forecast the VaR is ARCH/GARCH type models, for instance,
J.P. Morgen Bank uses this method to estimate and forecast the risks. This paper compares the
ARCH/GARCH type model to the GPD method in process of estimating and forecasting the VaR. In
detail, when comparing one-step forward prediction of the VaR, both methods, the GARCH-type
models and GPD method, adopt the method of moving windows or slide windows [10]
、[20]-[22]
and
expanding the window width step by step separately.
In the paper, volatile-ratio is taken as testing indictor in the examining process of inspecting the
performances of the forecasting of the VaR, which is defined as follows:
T
∑ I{ri >VaRi (α )}
⎧0, if ri < VaRi (α )
i = n +1
v − ratio = , I{ri >VaRi (α )} = ⎨
T −n ⎩1, if ri ≥ VaRi (α )
If the model specification is correct, then the ratio, which measures the realized returns rate
larger than the VaR at same time, will close to the given significance level α .
The model specification of the GARCH-type models of forecasting the VaR is defined:
rt = ε t ht , ht = ω + α1 r 2t −1 + β1 ht −1
where ε t is i.i.d. series, the corresponding distribution function is among those distributions, such as
normal distribution, generalized error distribution (GED) and standardized t -distribution, and so on.
On the one hand, the one-step forward forecasting value of VaR with the GARCH-type models is
VaRt (α ) = Ψ ← (1 − α ) * ht −1 , t = n + 1, n + 2, ,T
where Ψ ( x ) is the distribution function of ε t .
On the other hand, the one-step forward forecasting value of VaR with the GPD model is
σˆ N ˆ
VaRm+1 (α ) = u + [( u )ξ − 1], m = n, n + 1,…, T − 1
ξˆ mα
Nu
where u is the best optimal threshold value, (m = n, n + 1, , T − 1) is the ratio that the
m

observed sampling numbers exceeds the threshold value in the window, the parameter estimators σ̂ 、

ξˆ is the estimators using the whole sampling data .

4 The empirical comparative analysis
(1) The selection of data and stationary testing
The percentage of log-returns series of Composite Index of Shanghai Stock Exchange（closing
price, the date from July 1, 1999 to May 10,2005 , the sampling size is 1408）is selected as the objects ,
which definition is shr 96t = 100 × {log( Pt ) − log( Pt −1 )} , where Pt is the closing price of day t .
Considering the special statistical characters of the high frequency financial series, such as
clustering of volatility, long-memory, the duration character of the volatility and the structure
changing character of the time series and so on, can influence the standard testing size and power in
5
the ADF/PP unit root testing of the high frequency financial time series, so both method, fixed and
slide window width, are employed to test the stationary property of shr 96 , and this test will produce
the statistic values of different ADF unit root test , and by those statistic, the stationary property of the
time series can be determined.
In the paper, the model is specified as : shr 96t = ρ shr 96t −1 + αΔshr 96t −1 + ε t , the testing estimator is
ρˆ − 1
zt = . In order to assure the efficient of the testing, the window width is specified as 500. The
σˆ ρˆ

results show as Fig.1, the hypothesis that shr96 is non-stationary is rejected (α =0.01), the hypothesis
that shr 96 is I (0) can be accepted.

2、The estimating and testing of Tail-Index

Based on the accepting stationary testing above, the EVT can be used to estimate and test the
Tail-Index of the distribution function for shr 96 time series. From the initial estimating of Hill’s
scatter graph, the outcomes show that the shr 96' s lower Tail-Index have the values around 0.4, as
shown in the left of Fig. 2.
The GPD method is adapted to estimate and forecast of the shr96’s lower Tail-Index. The value
of threshold will be confirmed firstly. As to the right of Fig.2 shown, the threshold value can be
selected among 1.4, 1.45, 1.5 and 2.1 with GPD method, then the estimating values of the shr96’s
lower Tail-Index can be gained as shown in Table 1. Furthermore, due to the less change of the
shr96’s lower Tail-Index, the value of lower Tail-Index selected is of 0.2604, which corresponded to
more extreme threshold value 1.45. Fig.3.1 to Fig.3.3 offers the fitting testing results the truncated
distribution, tail distribution and the residual. Obviously, the outcomes are satisfactory.
3、The estimating and forecasting of the Value-at-Risk
Fig.4 puts forward the point estimation and the interval estimation of the VaR by means of GPD
method given the significance level α = 0.001 . The outcomes show the advantages of estimating the VaR
with GPD method, which mean the GPD method can realize the estimation of the VaR outside of the
sampling interval. The one-step forward estimating process of the VaR is concretely compared as follows:
(1) when forecasting the value of VaR with GARCH-type models, the common models include separately
GARCH-normal, GARCH-t and GARCH-GED models. In the process of one-step forward predicting, the
preceding two models used the methods of the expanding window width, but the third model used the
sliding window width method. The initial window widths are all specified as 1000. Considering the GPD
method pays attention to the changing situation of tail in the returns series, but the abnormal alterations of
the returns series are no so much, so the fixed parameters estimators of ξˆ = 0.267 and σˆ = 1.0679 are
adopted, when the empirical analysis of the forecasting the VaR are executed. The ratio
N u / m ,(m = 1000,1000 + 1, ,1407) , which shows in the window widths the proportion rate of sampling
data exceed the threshold value, takes as 0.015、0.01 and 0.005 separately, when the significance level is
specified as 0.05、0.01 and 0.0014, then the threshold values are determined by the corresponding ratios.
The one-step forward forecasting outcomes (volatile-ratio) of the VaR on shr96 with two types of models
are compared as Table 2 shows.
From the comparative analysis of the volatile-ratio, the GPD model is superior to the
GARCH-normal、GARCH-GED and GARCH-t models as to the one-step forward forecasting outcomes
from different types of models as mentioned above. But only discussing the one-step forward forecasting
with GARCH-type models, the GARCH-t model is superior to the GARCH- GED model, and the last one
is the GARCH-normal model.
6
 In addition ,the Fig.5.1and Fig.5.2 show that, the one-step forward forecasting outcomes from
the GARCH models have greater fluctuating character, which will lead to the over-forecasting
outcomes of the VaR’s values. The over-forecasting outcomes will lead the investors (especially the
short-term investors) to react to the market in excess; and lead to more manipulating risks for
financial institution and institutional investors. But the forecasting outcomes from GPD model are
more stationary, which will be in favor of financial institution, institutional investors and the
supervision departments to regulate and control financial risks, evade financial risks and intervene in
time. In the real world, the two types of models are combined to use, the outcomes will come to
more ideal.
Table 1 the lower Tail index estimation of shr96 : GPD method Table 2 the comparison of one-step forward forecasting
Probability α
Extreme 0.05 0.01 0.001
of Less method
u value ξlower σ lower
than GARCH-normal 0.0883 0.0762 0.0516
numbers
threshold GARCH-t 0.0835 0.0418 0.0147
1.4 204 0.2812(0.0967) 1.0167(0.1198) 0.8550 GARCH-GED 0.0983 0.0713 0.0369
1.45 192 0.2604(0.0975) 1.0692(0.1198) 0.8635 GPD 0.054 0.027 0.0074
Note: When forecasting the VaR’s values, GPD,GARCH-normal
1.5 184 0.2670(0.1010) 1.0697(0.1321) 0.8692
and GARCH-t models all adopt the method of expanding
2.1 110 0.2714(0.1457) 1.2148(0.2092) 0.9218 window width step by step, but GARCH-GED model adopts the
moving window .
Note: The number in the brackets are standard errors of parameters

Fig. 1 ADF unit-root testing of sliding window of shr 96 time series

Fig.2 Hill’s plot and the scatter plot of SME

7
 Fig 3.1 Fig. 3.2

Fig. 3 the different fitting testing :GPD method

(1) Fig.3.1 The fitting outcome of truncated distribution,

u = 1.45 , ξˆ = 0.2642 , σ = 1.0692 , Fu ( x − u ) adopts empirical

distribution function

(2)Fig. 3.2 The fitting testing of tail distribution , 1 − F ( x ) adopts

the empirical distribution function

(3)Fig.3.3 QQ plot of the residuals (to exponential distribution

function).

Fig. 3.3

Fig.4 The point estimation and interval estimation of VaR ( α = 0.001 .the significance interval is 0.95, ξ = 0.2604, σ = 1 ), GPD method

(shr96)

8
 Fig. 5.1 The one-step forward forecasting outcome of VaR(shr96), (the initial window width is 1000, the significance level is 0.01,
sliding window method step by step )

Fig. 5.2 The one-step forward forecasting outcome of VaR (shr96)

Note: the left plot is the outcome from GARCH(1,1)-GED model by the method of sliding window , the right plot is the outcome by the method of
expanded the window width, the initial window is specified as 1000, and the significance levels are all 0.01.

References
[1] Embrechts, P., Kluppelberg, C. & Mikosch, T. (1997) Modeling Extreme Events for insurance and finance, [M] Springer,
New York, 1997.
[2] Reiss, R.D. & Thomas, M. (1997) Statistical analysis of extreme values—from insurance, finance, hydrology and other
fields, [M] Springer, Berlin,1997.
[3] McNeil, A.J. (1997) Estimating the tails of loss severity distributions using extreme value theory, [J] ASTIN Bulletin, Vol.
27, 1997, pp117-137
[4] McNeil, A.J. (1998) Calculating quantile risk measures for financial time series using extreme value theory, [R]
manuscript, Department of Mathematics, ETH, Swiss Federal Technical University,1998.
[5] McNeil, A.J. (1999) Extreme value theory for risk managers, [C] in Internal Modeling and CAD II, Risk Books, 1999,
pp93-113.
[6] McNeil, A.J. & Frey, R. (2000) Estimation of tail-related risk measures for heteroscedastic financial time series: an
Extreme value approach, [J] Journal of Empirical Finance, Vol. 7, pp271-300.
[7] McNeil, A.J. & Saladin, T. (1997) The peak over thresholds method for estimating high quantiles of loss distributions, [R]
discussion paper, Department of Mathematics, ETH, Swiss Federal Technical University, 1997.
[8] Silva, A. C. & B. V. Mendes (2003), Value-at-risk and Extreme Returns in Asian Stock Markets, [J] International Journal
of Business, Vol. 8, No. 1, 2003.
9
[9] Terence D. Agbeyegbe and Gene Leon (2002), The tail behavior of stock index return on the Jamaican Stock Exchange,
[Z] http://www.centralbank.org.bb/ Publications/ WP2002pp103_110.pdf.
[10] R. Gençay, F. Selçuk & A. Ulugülyaˇgci (2003) High volatility, thick tails and extreme value theory in value-at-risk
estimation, [J] Insurance: Mathematics and Economics 33 (2003) 337–356
[11] Leadbetter, M. R., Lindgren, G., & Rootzen, H. (1983) Extremes and related properties of random sequences and
processes, [M] Springer, New York, 1983.
[12] Resnick, S. I. (1987) Extreme value, regular variation, and point processes, [M] Springer, New York,1987.
[13] Hill, B. M. (1975) A sample general approach to inference about the tail of a distribution, [J] Ann Statist. Vol. 3, 1975,
pp1163-1174
[14] J. Pickands (1975) Statistical inference using order statistics, [J] Ann Statist., Vol. 3, 1975, pp119-131.
[15] Dekkers, A. L. M., Einmahl, J. H. J., & Haan, L. de (1989) A moment estimator for the index of an extreme- value
distribution, [J] Ann Statist., Vol. 17, 1989, pp1833-1855
[16] Mills, T.C. (1999) The econometric modeling of financial time series (2nd edition), [J] Cambridge University Press,
1999,Cambridge
[17] Dekkers, A. L. M., Haan, L. (1989) On the estimation of the extreme value index and large quantile estimation, Ann
Statist , 17, 1795-1832
[18] Loretan, M. and Phillips, P. C. B. (1994) Testing the covariance stationarity of heavy-tailed time series: An overview of
the theory with applications to several financial datasets, [J] Journal of Empirical Finance, Vol. 1, 1994, pp211-248
[19] Balkema, A. A., and Haan, L. de (1974) Residual lifetime at great age, [J] Ann. Prob., Vol. 2, 1974, pp792-804.
[20] Gencay, R. & Selcuk, F. (2004) Extreme value theory and value-at-risk: Relative performance in emerging markets, [J]
International Journal of Forecasting Vol. 20, 2004, pp287– 303.
[21] Gencay, R., Selcuk, F. and Ulugulyagci, A. (2001) EVIM: A software package for extreme value analysis in MATLAB,
[J] Studies in Nonlinear Dynamics and Econometrics, Vol. 5 (3), 2001, pp213- 239.
[22] Gencay, R. and Selcuk, F. (2001) Software Review: Matlab Neural Networks Toolbox, [J] International Journal of
Forecasting, Vol. 17, 2001, pp305-317.
[23] Zuo-xiang, Peng (2003), Analysis on high-frequency financial time series econometric modeling，[D] Doctoral Thesis,
Southwestern University of Finance and Economics, July, 2003, Chengdu.

10