Physica A 391 (2012) 55465556

Contents lists available at SciVerse ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa
Forecasting volatility of fuel oil futures in China: GARCH-type,
SV or realized volatility models?
Yu Wei
School of Economics and Management, Southwest Jiaotong University, The First Section of the Northern Second Ring Road, Chengdu 610031, China
a r t i c l e i n f o
Article history:
Received 24 May 2011
Received in revised form 23 July 2011
Available online 15 June 2012
Keywords:
Oil futures
Volatility forecasting
Realized volatility
a b s t r a c t
In most previous works on forecasting oil market volatility, squared daily returns were
taken as the proxy of unobserved actual volatility. However, as demonstrated by Andersen
and Bollerslev (1998) [22], this proxy with too high measurement noise could be perfectly
outperformed by a so-called realized volatility (RV) measure calculated by the cumulative
sum of squared intraday returns. With this motivation, we further extend earlier works
by employing intraday high-frequency data to compare the performance of three typical
volatility models in the daily out-of-sample volatility forecasting of fuel oil futures on the
Shanghai Futures Exchange (SHFE): the GARCH-type, stochastic volatility (SV) and realized
volatility models. By taking RV as the proxy of actual daily volatility and then computing
forecasting errors, we find that the realized volatility model based on intraday high-
frequency data produces significantly more accurate volatility forecasts than the GARCH-
type and SV models based on daily returns. Furthermore, the SV model outperforms many
linear and nonlinear GARCH-type models that capture long-memory volatility and/or the
asymmetric leverage effect in volatility. These results also prove that abundant volatility
information is available in intraday high-frequency data, and can be used to construct more
accurate oil volatility forecasting models.
2012 Elsevier B.V. All rights reserved.
1. Introduction
Modeling and forecasting of volatility in the oil markets is a key issue in such fields as oil derivative product pricing,
portfolio allocation and risk measurement. Engles [1] seminal paper kicked off the development of a large number of so-
called historical volatility models in which a time-varying volatility process is extracted from financial return data. Most
such models can be regarded as variants of the generalized autoregressive conditional heteroskedasticity (GARCH) models
developed by Bollerslev [2] and others [39]. An important competitive class of GARCH-type models is associated with the
stochastic volatility (SV) model [1020]. Both the GARCH-type and SV models are commonly employed in analysis of daily,
weekly and monthly returns. However, the recent widespread availability of intraday high-frequency prices for financial
assets, and the research thereon, has shed new light on the concept of volatility. The availability of intraday data has not
only led to the development of improved volatility measurements but has also inspired research into the potential value of
these data as an information source for volatility forecasting [21].
The majority of research carried out to date regards daily squared returns as a daily volatility measurement, although
Andersen and Bollerslev [22] and Barndorff-Nielsen and Shephard [23] demonstrated that the measurement noise in daily
squared returns is too high for observing the true underlying volatility process. They showed that realized volatility (a daily
volatility measure), as calculated by the cumulative sum of squared intraday returns, is less subject to measurement error
and thus less noisy. In fact, under certain assumptions, realized volatility (RV) is a consistent estimator of the quadratic
E-mail addresses: ywei@home.swjtu.edu.cn, weiyusy@126.com.
0378-4371/$ see front matter 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2011.08.071
Y. Wei / Physica A 391 (2012) 55465556 5547
variation in the underlying diffusion process [22]. In principle, the volatility measures derived fromintraday high-frequency
data should prove to be more accurate, thus allowing forecast efficiency gains [2427].
A great deal of research has focused on evaluating the forecasting performance of different volatility models in the
oil markets by using the GARCH-class ones [2846]. For example, Sadorsky [45] employed the closing futures prices for
WTI crude oil, heating oil, unleaded gasoline and natural gas to compare the volatility forecasting accuracy of the GARCH,
GARCH-in-mean and TGARCH (or GJR) models with that of a number of others under three loss functions. The empirical
results by DM test of Diebold and Mariano [47] indicated that the TGARCH model fitted well for heating oil and natural gas
volatility, and the GARCH model fitted well for crude oil and unleaded gasoline volatility. Despite their greater complexity,
such models as the state space, vector autoregression and bivariate GARCH did not appear to perform as well as the
single equation GARCH model. Narayan and Narayan [42] employed the exponential GARCH (EGARCH) model to examine
whether shocks had asymmetric and persistent effects on crude oil price volatility. They found inconsistent evidence of such
asymmetry and persistence in different sub-samples, although the evidence over their full sample period suggested that
shocks had both permanent and asymmetric effects on volatility. Kang et al. [38] evaluated the out-of-sample forecasting
accuracy of four GARCH-class models (GARCH, IGARCH, CGARCH, FIGARCH) using the DM test under two loss functions.
They found that the CGARCH and fractionally integrated GARCH (FIGARCH) models were able to capture the long-memory
volatility of three crude oil markets (WTI, Brent and Dubai) and exhibited superior performance to their GARCHand IGARCH
counterparts. In a study complementing those of earlier works [38,42,45], Cheong [32] investigated the out-of-sample
forecasting performance of four GARCH-class models (GARCH, APARCH, FIGARCH, FIAPARCH) under three loss functions,
finding that the simplest and most parsimonious GARCH model fitted the Brent crude oil data better than the other models
examined, although the FIAPARCH out-of-sample WTI forecasts exhibited superior performance. Agnolucci [29] compared
the predictive ability of GARCH-type models with the implied volatility (IV) obtained by inverting the Black equation in
WTI crude oil futures contracts quoted on the New York Mercantile Exchange (NYMEX). His empirical results indicated
that the GARCH-type models seemed to perform better than the implied volatility. Employing the weekly crude oil spot
prices in 11 international markets, Mohammadi and Su [40] compared the forecasting accuracy of four GARCH-class models
(GARCH, EGARCH, APARCH, FIGARCH) under two loss functions. The DM test showed the APARCH model to exhibit the best
performance. In a recent investigation, Nomikos and Pouliasis [43] concluded the superiority of Markov Regime Switching
GARCH (MRS-GARCH) model among the GARCH-type ones.
It is clear from the foregoing summary of the extant literature that the empirical evidence is mixed. Which volatility
model is superior in modeling and forecasting the volatility of oil markets remains obscure. Furthermore, all of the
conclusions in the earlier papers are based on a method that actual volatility is assessed by daily or weekly squared returns.
Although Chevallier [33] used intraday tick-by-tick data of carbon prices to measure realized volatility, his emphasis was to
detect the outliers and instability in the volatility of carbon prices instead of volatility forecasting. To the best of the authors
knowledge, no previous research evaluating the volatility forecasting performance of different models in the oil markets
has adopted the realized volatility measurement as a proxy for actual daily volatility. Hence, the conclusions reached in that
research may be unreliable and deserving of further investigation.
As discussed above, there are two aspects which deserve to be further extended in the earlier researches on forecasting
oil volatility. For one thing, most papers focused on the GARCH-type models, for another, noisy squared daily/weekly
returns were treated as actual volatility to evaluate forecasting accuracy. However, Hansen and Lunde [48] stated that the
substitution of a noisy proxy such as squared daily returns for the true but unobservable conditional variance could result
in an inferior model being chosen as the best one. The use of RV as a proxy variable, in contrast, did not lead to the favoring
of an inferior model.
Motivated by this, the aimof this paper is to evaluate which model commonly used in financial research, GARCH-type, SV
or RV, is more suitable for the volatility forecasting of fuel oil futures on the Shanghai Futures Exchange (SHFE). This paper
is different from the existing research on volatility forecasting of oil markets in the following respects. First, a scaled RV
measurement based on intraday high-frequency data proposed by Hansen and Lunde [49] is used as a proxy for actual daily
volatility to calculate forecasting error. Second, a larger scope of forecasting models is adopted, i.e., the GARCH-type models
commonly employed in earlier studies are compared not only with a SV model based on daily data, but also with a RV model
based on intraday high-frequency data. Finally, to obtain robust conclusions, this paper incorporates most of the GARCH-
type models proved to perform well in earlier studies, and all the evaluation criteria adopted by earlier literature [29,32,38,
40,42,46]. We also employ a technique for model comparison as in Ref. [46], namely, the superior predictive ability (SPA)
test by Hansen [50]. Hansen states that the stationary bootstrap procedure employed in the SPA test gives it good power
properties and makes it more robust than the approaches used in previous studies, such as the DM test and Whites [51]
reality check test. Furthermore, at the end of the empirical analysis, we also use squared daily returns as an alternative
benchmark of volatility forecasting to check the performance of different volatility models.
The remainder of this paper is organized as follows. Section 2 introduces the sample data and discusses the construction
of the daily andintraday returns. Section3shows howRVis derivedfromintraday returns andthe autoregressive fractionally
integrated moving average (ARFIMA) model, then Section 4 briefly describes the historical volatility models, i.e., the GARCH-
type and SV models. The methodology of out-of-sample forecasting and the SPA test are discussed in Section 5, and the
empirical forecasting results are presented in Section 6. Section 7 concludes the paper.
5548 Y. Wei / Physica A 391 (2012) 55465556
2. Data sample, daily and intraday returns
In recent years, people have witnessed Chinas increasing integration into the world economy, which has made closer
cooperative relationships with other economies in areas such as economic development and resource allocation essential.
China is currently the worlds second largest oil consumer and third largest importer. On August 25, 2004, with the approval
of the China Securities Regulatory Commission (CSRC), fuel oil contracts were listed on the Shanghai Futures Exchange
(SHFE) for trading. Together with gas oil, diesel and kerosene, fuel oil is one of the four major products processed fromcrude
oil, and China is now the largest fuel oil consumer in Asia.
Before the introduction of fuel oil futures to the SHFE, the producers and consumers of fuel oil in China had to use the oil
futures in overseas exchanges, such as the NYMEX, the International Petroleum Exchange (IPE), the Singapore Exchange
(SGX) and the Tokyo Commodities Exchange (TOCOM). Today, they can easily hedge risk without worrying about the
volatility of exchange rates and overseas interest rates. At the end of 2009, the trading volume of fuel oil futures on the
SHFE was the third largest internationally, lower only than that of light sweet crude oil futures on the NYMEX and Brent
crude oil futures on the IPE. The modeling and forecasting of the price volatility of fuel oil futures on the SHFE is thus of great
importance to oil hedgers and speculators globally.
Similar to the approach adopted by Wang and Yang [52], we choose the intraday high-frequency (every 15 min) price
quotes for fuel oil futures on the SHFE during November 1, 2007July 14, 2011, for a total of N = 879 trading days. There are
about four trading hours in a trading day on the SHFE, and fifteen 15 min quotes are recorded per day. These 15 min price
quotes are denoted as I
t,d
, t = 1, 2, . . . , N and d = 1, 2, . . . , 15, where I
t,15
indicates the close price quote. Hence, the daily
return R
t
is defined as
R
t
= 100(ln I
t,15
ln I
t1,15
), (1)
and the intraday 15 min return R
t,d
is defined as
R
t,d
= 100(ln I
t,d
ln I
t,d1
). (2)
3. Realized volatility measurement and the ARFIMA model
It is generally accepted that squared daily returns provide a poor approximation of actual daily volatility. Andersen and
Bollerslev [22] pointed out that more accurate estimates can be obtained by summing all squared intraday returns. If we
were to apply their method directly in this paper, then we would define the RV measurement as
RV

t
=

R
2
t,d
. (3)
However, this definition ignores the information contained in overnight returns. To address this problem, Hansen and
Lunde [49] suggested scaling the RV measurement in the following way.
RV
t
= RV

t
, (4)
where the so-called scale parameter is defined as
=
N
1
N

t=1
R
2
t
N
1
N

t=1
RV

t
. (5)
Andersen et al. [53] found the distribution of RV to be highly non-normal and skewed, but its logarithms to be approxi-
mately normal. Accordingly, they suggested that the natural logarithms of a RV measurement series, denoted as lnRV, could
be modeled by a Gaussian dynamic process. They adopted an ARFIMA to model this process. The ARFIMA(1, d, 1) model
with mean for lnRV can be given by
(1 L)(1 L)
d
(ln RV
t
) = (1 + L)
t
, (6)
where L is the lag operator, coefficients d, and are fixed and unknown parameters, and
t
is Gaussian white noise with
mean zero and variance
2

. Because different ARFIMA model specifications produce rather similar forecasting results [21],
we consider the ARFIMA(1, d, 1) model in the following empirical study and denote it as ARFIMA-lnRV.
4. Historical volatility models
4.1. GARCH-type models
The most commonly used volatility model is the GARCH model proposed by Bollerslev [2] and based on the work of
Engle [1]. Then many extensions of the basic GARCH model are developed to depict the stylized facts of financial markets.
Briefly, six linear and nonlinear GARCH-type models, i.e., the GARCH, IGARCH, GJR, EGARCH, FIGARCH and FIAPARCH, are
employed in this paper to forecast the volatility of fuel oil futures on the SHFE. More technical details about these GARCH-
type models can be found in Ref. [46] among others.
Y. Wei / Physica A 391 (2012) 55465556 5549
4.2. Stochastic volatility model
The stochastic volatility model (SV) provides an alternative to the GARCH-type models, and for daily returns it is given by
R
t
=
t

t
,
t
NID(0, 1),

2
t
=
2
exp(h
t
),
h
t
= h
t1
+

t1
,
t
NID(0, 1), h
1
NID(0,
2

/{1
2
}), (7)
where
2
may be treated as a scale parameter, and the persistence parameter is restricted to a positive value less than
1 to ensure stationarity;
t
and
t
are assumed to be mutually uncorrelated. Shephard [54] provides additional theoretical
details of the SV model.
5. Forecasting methodology and the SPA test
In this section, we evaluate the forecasting performance of the ARFIMA-lnRV, SV and six GARCH-type models. The
ARFIMA-lnRV model is based on intraday high-frequency data, whereas the SV and GARCH-type models are based on daily
returns. The forecasting process is handled as follows.
The November 1, 2007July 14, 2011 observations of fuel oil futures on the SHFE are classified into two subgroups: (1)
in-sample data for volatility modeling, covering 579 trading days, and (2) out-of-sample data for model evaluation, covering
the last 300 trading days of the total data sample. The estimation period is then rolled forward by adding one new day and
dropping the most distant day. In this way, the sample size used to estimate the models remains at a fixed length and the
forecasts do not overlap, thereby allowing daily (one-day-ahead) out-of-sample volatility forecasts to be obtained.
The volatility forecasts obtained by the RV, SV and GARCH-type models are indicated by
2
t
, and the scaled RV
measurement denoted as RV
t
is taken as a proxy for actual daily volatility (the forecasting benchmark). Various forecasting
criteria or loss functions can be considered in assessing the predictive accuracy of a volatility model, although, as Lopez [55]
noted, it is not obvious which loss function is most appropriate for the evaluation of such models. Rather than making a
single choice, we thus employ the following six accuracy statistics or loss functions as our forecasting criteria.
MSE = n
1
n

t=1
(RV
t

2
t
)
2
, (8)
MAE = n
1
n

t=1
|RV
t

2
t
|, (9)
HMSE = n
1
n

t=1
(1
2
t
/RV
t
)
2
, (10)
HMAE = n
1
n

t=1
|1
2
t
/RV
t
|, (11)
where n is the number of forecasting data points; MSE and MAE are the mean square error and mean absolute error; HMSE
and HMAE are the MSE and MAE, respectively, adjusted for heteroskedasticity. Different criteria serve different practical
purposes. For example, in the case of Value-at-Risk applications, greater interest may lie in the accurate forecasting of a
high rather than a low level of volatility, which implies that the MSE criterion is the most relevant loss function in risk
management applications. Additional discussion of these criteria can be found in Ref. [56].
Whena particular loss functionis smaller for model Athanit is for model B, it is impossible toconclude that the forecasting
performance of the former is superior to that of the latter. Such a conclusion cannot be made on the basis of a single loss
function and a single sample. Recent work has focused on a testing framework that can determine whether one particular
model is outperformed by another [47,51,57]. As discussed in the Introduction, the SPA test, an extension of the White
framework proposed by Hansen and Lunde [49], has been shown to possess good power properties and to be more robust
than previous approaches.
In contrast to other evaluation techniques, the SPA test can be used to compare the performance of two or more
forecasting models at the same time. Forecasts are evaluated employing a pre-specified loss function, and the best
forecasting model is the one that produces the smallest expected loss. In the SPA test, the loss function relative to the
benchmark model is defined as X
(0,i)
t,l
= L
(0)
t,l
L
(i)
t,l
, where L
(0)
t,l
is the value of loss function l at time t for benchmark model
M
0
, and L
(i)
t,l
is the value of loss function l at time t for competing model M
i
, for i = 1, . . . , K. The SPA test is used to compare
the forecasting performance of a benchmark model against its K competitors. The null hypothesis that the benchmark or
base model is not outperformed by any of the competing models can be expressed as H
0
: max
i=1,...,K
E(X
(0,i)
t,l
) 0. It is
tested with the statistic T
SPA
l
= max
i=1,...,K
(

X
i,l
/

lim
n
var(

X
i,l
)), where n is the number of forecast data points
5550 Y. Wei / Physica A 391 (2012) 55465556
Table 1
Descriptive statistics of time series of fuel oil futures on the SHFE.
R
t
R
t
2
lnRV
t
Mean (%) 0.089 1.980 0.025
Standard deviation (%) 1.405 3.682 1.301
Skewness 0.263 3.139 0.099
Excess kurtosis 1.546 10.968 0.209
JarqueBera 97.646
a
5849.652
a
30.043
a
Q(20) 34.682
b
677.987
a
4871.860
a
ADF 32.187
a
7.367
a
3.123
b
PP 32.223
a
24.084
a
17.140
a
Notes: The JarqueBera statistic tests for the null hypothesis of normality for the distribution of the series.
Q(20) is the LjungBox statistic for up to 20th order serial correlation. ADF and PP are statistics of the
Augmented DickeyFuller and PhillipsPerron unit root tests based on the least AIC criterion.
a
Denote rejection of the hypothesis at the 1% level.
b
Denote rejection of the hypothesis at the 5% level.
and

X
i,l
= n
1

n
t=1
X
(0,i)
t,l
. The estimation of lim
n
var(

X
i,l
) and the p-value of T
SPA
l
are obtained using the stationary
bootstrap procedure discussed by Politis and Romano [58]. To save space, this paper includes no further technical details of
the SPA test, but more in-depth discussions can be found in the studies of Hansen and Lunde [49] and Koopman et al. [21]. In
summary, the p-value of a SPA test indicates the relative performance of a base model (M
0
) in comparison with alternative
models (M
i
). A high p-value means that the null hypothesis that the base model is not outperformed cannot be rejected.
6. Empirical results
6.1. Descriptive statistics of the data
Table 1 provides descriptive statistics of the three time series of fuel oil futures on the SHFE. The sample mean of the
return series is quite small in comparison with the standard deviation. It is thus reasonable to set the conditional mean of
the return series to zero. In addition, for all three series, the JarqueBera statistic shows that the null hypothesis of normality
is rejected at the 1% significance level, as also evidenced by the high degree of excess kurtosis and skewness. The LjungBox
statistic for serial correlation shows that the null hypothesis of no autocorrelation up to the 20th order is rejected, thus
confirming serial correlation in the three fuel oil futures series. The Augmented DickeyFuller and PhillipsPerron unit root
tests both support the rejection of the null hypothesis of a unit root at the 5% or 1% significance level, thus implying that all
of the series are stationary and may be modeled directly without further transforms.
6.2. Detrended cross-correlation analysis (DCCA) of volatility series
Recently, Podobnik and Stanley [59] propose a detrended cross-correlation analysis which can be used to detect the long-
range cross-correlation between two time series. This method is widely applied to financial time series [59,60] and can be
described as follows.
Step 1. Consider the two time series, {x
t
, t = 1, . . . , N} and {y
t
, t = 1, . . . , N}, where N is the equal length of these two
series. Then we describe the profile of each series and get two newseries, xx
k
=

k
t=1
(x
t
x) and yy
k
=

k
t=1
(y
t
y), k =
1, . . . , N.
Step 2. Divide both the profiles {xx
k
} and {yy
k
} into N
s
= int(N/s) nonoverlapping segments of equal length s. Since
the length N of the series is not often a multiple of the considered time scale s, a short part at the end of each profile may
remain. In order not to disregard this part of the series, the same procedure is repeated, starting from the opposite end of
each profile. Therefore, 2N
s
segments are obtained together. We set 10 < s < N/5.
Step 3. We calculate the local trends xx
(1)s+j
and yy
(1)s+j
for each of the 2N
s
segments by a least-square fit of each
series. Then we determine the co-moved variance
F
2
(s, )
1
s
s

j=1
[xx
(1)s+j
xx
(1)s+j
][yy
(1)s+j
yy
(1)s+j
] (12)
for = 1, 2, . . . , N
s
and
F
2
(s, )
1
s
s

j=1
[xx
N(N
s
)s+j
xx
N(N
s
)s+j
][yy
N(N
s
)s+j
yy
N(N
s
)s+j
] (13)
for = N
s
+1, N
s
+2, . . . , 2N
s
. The trends xx
(1)s+j
and yy
(1)s+j
can be computed from a linear, quadratic or high order
polynomial fit of each profile for segment .
Step 4. Average over all segments to get the fluctuation function
F
q
(s) =

1
2N
s
2N
s

=1
[F
2
(s, )]

1/2
. (14)
Y. Wei / Physica A 391 (2012) 55465556 5551
Fig. 1. Loglog plot between F(s) and s for the RV and squared return series.
Step 5. Analyze the scaling behavior of the fluctuation function through the loglog plots of F
q
(s) versus s. If two series
are long-range cross-correlated, as a power-law
F(s) s
H
, (15)
the scaling exponent H can be estimated by analyzing the loglog relationships between fluctuation function, F(s), and time
scale, s. If H > 0.5, we can say that two time series are long-range cross-correlated. An increase of one series is likely to be
followed by an increase of the other series.
Fig. 1 shows the loglog plot between F(s) and time scale s for the RV and squared return series. We can find a general
power-law relationship between F(s) and time scale s. The DCCA scaling exponent (H = 0.7847) is obviously larger than
0.5, implying the existence of strong long-range cross-correlations between the RV and squared return series.
6.3. Estimation results for different volatility models
Table 2 presents the in-sample estimation results for the volatility models discussed in Sections 3 and 4. The lower parts
of the table show the results of the diagnostic test on the standardized residuals.
First, the results reported in Table 2 show that is close to 1 and significant at the 1% level for the GARCH, IGARCH,
GJR, EGARCH and FIGARCH models. The parameter for the SV model is also close to 1 and significant at the 1% level.
These results indicate a high degree of volatility persistence in the fuel oil futures on the SHFE. Second, the asymmetric
leverage coefficients reported in this table are not significantly different from0 in the GJR, EGARCHand FIAPARCHmodels,
thus indicating that there is no obvious leverage effect in the price volatility of these futures. Third, the value of the power
coefficient is 2.597 for the FIAPARCH model. The null hypothesis is rejected for = 1, but not for = 2, at the 10%
significance level. This result indicates that fuel oil futures series are better modeled with conditional variance rather than
conditional standard deviation. Finally, the fractional difference parameters d reported in Table 2 are all significant in the
FIGARCH, FIAPARCH and ARFIMA-lnRV models, which indicate a significant degree of long-memory volatility in the fuel oil
futures.
The results of the diagnostic tests are reported in the lower parts of Table 2. In general, the Log(L) and AIC values are
very close to each other within the GARCH-type models. None of the LjungBox tests can reject the null hypothesis of no
serial correlation in the standardized residuals and squared standardized residuals at the 10% level. The ARCH tests are also
unable to reject non-heteroskedasticity at the 10% level.
Clive W. J. Granger, the 2003 Nobel Laureate in Economics, suggested that in the face of uncertainty over the true model,
model selection should be based on forecasting performance rather than in-sample fit [61]. In line with this view, we rely
on out-of-sample forecasting performance to evaluate the various volatility models.
6.4. Forecasting results
The out-of-sample daily volatility forecasts are constructed using the rolling forecasting methodology discussed in
Section 5. For clarity, Fig. 2 shows the forecasting results of three models, i.e., the ARFIMA-lnRV, SV and FIGARCH. The RV
measures (the forecasting benchmark) are displayed as dots. It is clear that, on the whole, SV and FIGARCH models make
higher volatility forecasts than that of the ARFIMA-lnRV model.
Table 3 presents the loss function values of the eight competing volatility models. It can be seen that in all four loss
functions, the ARFIMA-lnRV model based on intraday high-frequency data achieves the smallest forecasting error. Table 3
5552 Y. Wei / Physica A 391 (2012) 55465556
T
a
b
l
e
2
E
s
t
i
m
a
t
i
o
n
r
e
s
u
l
t
s
o
f
d
i
f
f
e
r
e
n
t
v
o
l
a
t
i
l
i
t
y
m
o
d
e
l
s
.
G
A
R
C
H
I
G
A
R
C
H
G
J
R
E
G
A
R
C
H
F
I
G
A
R
C
H
F
I
A
P
A
R
C
H
S
V
A
R
F
I
M
A
-
l
n
R
V

0
.
0
1
3
(
0
.
0
0
8
)
0
.
0
1
3
(
0
.
0
1
8
)
0
.
0
1
3
(
0
.
0
0
8
)
0
.
7
1
7
b
(
0
.
3
0
1
)
0
.
0
7
2
(
0
.
0
3
8
)
0
.
0
0
1
(
0
.
0
4
9
)

2
1
.
1
5
7
a
(
0
.
4
1
9
)

0
.
9
4
6
a
(
0
.
0
4
4
)

0
.
1
1
6
a
(
0
.
0
2
4
)
0
.
1
1
6
a
(
0
.
0
2
2
)
0
.
1
3
1
a
(
0
.
0
3
0
)
0
.
4
1
6
(
0
.
5
7
0
)

0
.
9
8
0
a
(
0
.
0
1
1
)

1
0
.
9
1
9
a
(
0
.
0
5
3
)

0
.
8
8
4
a
(
0
.
0
2
2
)
0
.
9
1
4
a
(
0
.
0
2
2
)
0
.
8
8
5
a
(
0
.
0
2
1
)
0
.
9
7
6
a
(
0
.
0
1
0
)
0
.
5
1
8
a
(
0
.
1
4
5
)
0
.
3
4
1
(
0
.
2
2
8
)

0
.
0
4
4
a
(
0
.
0
1
8
)
d

0
.
0
5
3
b
(
0
.
0
2
6
)

0
.
0
3
3
(
0
.
0
3
1
)
0
.
0
0
7
(
0
.
0
1
6
)

0
.
0
6
2
(
0
.
0
6
6
)

2
1
.
9
4
6

2
.
5
9
7
a
(
0
.
3
4
0
)

0
.
0
9
5
(
0
.
1
0
1
)
0
.
0
7
6
(
0
.
1
7
1
)
d
0
.
5
1
2
a
(
0
.
1
0
2
)
0
.
3
5
1
a
(
0
.
1
2
5
)
D
i
a
g
n
o
s
t
i
c
L
o
g
(
L
)

1
4
2
6
.
4
3
0

1
4
2
6
.
4
3
1

1
4
2
5
.
8
6
0

1
4
2
5
.
3
8
9

1
4
2
4
.
8
1
7

1
4
2
2
.
9
4
6

1
4
1
1
.
7
1
8

1
5
3
9
.
9
0
4
A
I
C
3
.
2
5
2
3
.
2
5
0
3
.
2
5
3
3
.
2
5
4
3
.
2
5
1
3
.
2
5
1
2
.
7
8
7
3
.
5
1
2
Q
(
2
0
)
2
1
.
1
7
6
[
0
.
3
8
7
]
2
1
.
1
8
3
[
0
.
3
8
6
]
2
1
.
7
2
0
[
0
.
3
5
6
]
2
1
.
7
8
0
[
0
.
3
5
2
]
2
0
.
6
6
7
[
0
.
4
1
7
]
2
0
.
1
2
0
[
0
.
4
5
0
]
2
1
.
6
4
0
[
0
.
3
8
4
]
2
2
.
8
2
4
[
0
.
2
4
2
]
Q
2
(
2
0
)
1
6
.
7
2
2
[
0
.
5
4
2
]
1
6
.
7
3
3
[
0
.
5
4
2
]
1
7
.
3
7
0
[
0
.
4
9
8
]
1
7
.
5
9
0
[
0
.
4
8
3
]
1
4
.
7
3
1
[
0
.
6
8
0
]
1
5
.
0
7
8
[
0
.
6
5
7
]
1
7
.
6
2
3
[
0
.
4
7
5
]
2
4
.
6
2
3
[
0
.
1
6
5
]
A
R
C
H
(
2
0
)
0
.
7
9
0
[
0
.
7
2
8
]
0
.
7
9
0
[
0
.
7
2
7
]
0
.
8
3
1
[
0
.
6
7
6
]
0
.
8
5
2
[
0
.
6
5
0
]
0
.
7
1
8
[
0
.
8
1
0
]
0
.
7
5
5
[
0
.
7
6
9
]
0
.
7
9
2
[
0
.
7
2
4
]
0
.
8
3
2
[
0
.
6
7
2
]
N
o
t
e
s
:
T
h
e
n
u
m
b
e
r
s
i
n
p
a
r
e
n
t
h
e
s
e
s
a
r
e
t
h
e
s
t
a
n
d
a
r
d
e
r
r
o
r
s
o
f
t
h
e
e
s
t
i
m
a
t
i
o
n
s
.
L
o
g
(
L
)
i
s
t
h
e
l
o
g
a
r
i
t
h
m
m
a
x
i
m
u
m
l
i
k
e
l
i
h
o
o
d
f
u
n
c
t
i
o
n
v
a
l
u
e
.
A
I
C
i
s
t
h
e
a
v
e
r
a
g
e
A
k
a
i
k
e
i
n
f
o
r
m
a
t
i
o
n
c
r
i
t
e
r
i
o
n
.
Q
(
2
0
)
a
n
d
Q
2
(
2
0
)
a
r
e
t
h
e
L
j
u
n
g

B
o
x
Q
-
s
t
a
t
i
s
t
i
c
s
o
f
o
r
d
e
r
2
0
c
o
m
p
u
t
e
d
o
n
t
h
e
s
t
a
n
d
a
r
d
i
z
e
d
r
e
s
i
d
u
a
l
s
a
n
d
s
q
u
a
r
e
d
s
t
a
n
d
a
r
d
i
z
e
d
r
e
s
i
d
u
a
l
s
,
r
e
s
p
e
c
t
i
v
e
l
y
.
A
R
C
H
(
2
0
)
i
s
t
h
e
n
o
n
-
h
e
t
e
r
o
s
k
e
d
a
s
t
i
c
i
t
y
s
t
a
t
i
s
t
i
c
o
f
o
r
d
e
r
2
0
.
T
h
e
p
-
v
a
l
u
e
s
o
f
t
h
e
s
t
a
t
i
s
t
i
c
s
a
r
e
r
e
p
o
r
t
e
d
i
n
s
q
u
a
r
e
b
r
a
c
k
e
t
s
.
a
D
e
n
o
t
e
r
e
j
e
c
t
i
o
n
o
f
t
h
e
h
y
p
o
t
h
e
s
i
s
a
t
t
h
e
1
%
l
e
v
e
l
.
b
D
e
n
o
t
e
r
e
j
e
c
t
i
o
n
o
f
t
h
e
h
y
p
o
t
h
e
s
i
s
a
t
t
h
e
5
%
l
e
v
e
l
.
Y. Wei / Physica A 391 (2012) 55465556 5553
Fig. 2. One-day ahead volatility forecasts of fuel oil futures on the SHFE and realized volatility measures (as dots).
Table 3
Loss function values for out-of-sample daily volatility forecasts of fuel oil futures on the SHFE (RV as benchmark).
MSE MAE HMSE HMAE
ARFIMA-lnRV1 7.895
a
0.673
a
5.187
a
1.011
a
SV 7.982
b
0.848
b
10.408
b
1.766
b
GARCH 8.022
c
0.873 13.369 1.922
IGARCH 8.043 0.882 13.083
c
1.911
c
GJR 8.040 0.881 13.741 1.940
EGARCH 8.032 0.935 16.577 2.182
FIGARCH 8.032 0.868
c
13.090 1.921
FIAPARCH 8.024 0.889 14.184 2.010
a
Denotes the smallest loss function value under a specific loss function.
b
Denotes the second smallest loss function value under a specific loss function.
c
Denotes the third smallest loss function value under a specific loss function.
also shows that the SV model based on daily returns achieves the second smallest forecasting error in the four loss functions.
However, no GARCH-type model is found to be superior to the ARFIMA-lnRV or SV models. Furthermore, among the GARCH-
type specifications, no model outperforms all of the others across the different loss functions.
The empirical results reported in Table 3 are in accordance with the findings of Koopman et al. [21], Wei et al. [46], among
others. Although long memory in volatility is a well-known stylized fact in financial and oil markets [6267], long-memory
GARCH-type models (e.g., FIGARCH and FIAPARCH) do not show overwhelming superiority to other GARCH-type ones in
financial or oil markets as far as volatility forecasting is concerned [32,37,40,46,68].
6.5. SPA tests
Although the forecasting results in Table 3 clearly prove the superiority of the ARFIMA-lnRV model, they apply only
to one select sample. To assess whether the same results can be obtained for similar samples, we carry out the SPA tests
discussed in Section 5. Table 4 presents the p-values of the SPA tests on the eight rival volatility models. The first column in
the table lists the name of the base model (M
0
) in the SPAtest, and thus the seven remaining models are treated as competing
models (M
i
). Every number in Table 4 is a SPA p-value obtained through 10,000 times of bootstraps under a specified loss
function. As noted in Section 5, the larger the p-value, the less likely it is that the SPA null hypothesis the base model is not
outperformed by all competing models can be rejected, that is, the better the forecasting performance of the base model
(M
0
) is relative to its alternatives (M
i
). The values in bold are the largest p-values under a specific loss function.
Table 4 confirms the findings in Table 3. The ARFIMA-lnRV model achieves the largest p-values under all loss functions.
This result constitutes strong evidence that the ARFIMA-lnRVmodel is capable of achieving muchbetter volatility forecasting
accuracy for fuel oil futures on the SHFE than such historical volatility models as the SV and GARCH-type models. Within the
historical models, the SV performs quite well under the MSE loss function, and no GARCH-type model achieves the largest
p-value. In short, it can be concluded that the RV model based on intraday data produces better forecasting accuracy than
the historical models.
The implication of these findings is interesting and worthy of examination in future studies. First, volatility measurement
and forecasting by the RV method have higher accuracy than the traditional GARCH-class and SV method. Thus Value-at-
Risk forecasting and prediction of option prices of oil products, which are heavily dependent on the precision of volatility
forecasting, may benefit from the RV model. Second, although the GARCH-class models are popular among academic
researchers and practitioners, the SV model are obviously superior to the GARCH-class ones. If only daily data is available,
5554 Y. Wei / Physica A 391 (2012) 55465556
Table 4
SPA p-values for out-of-sample daily volatility forecasts of fuel oil futures on the SHFE (RV as benchmark).
Base model MSE MAE HMSE HMAE
ARFIMA-lnRV 0.837 0.518 0.555 0.559
SV 0.578 0.000 0.005 0.000
GARCH 0.018 0.000 0.036 0.000
IGARCH 0.001 0.000 0.026 0.000
GJR 0.011 0.000 0.040 0.000
EGARCH 0.344 0.000 0.002 0.000
FIGARCH 0.142 0.000 0.003 0.000
FIAPARCH 0.229 0.000 0.009 0.000
Notes: The values in bold are the largest p-values under a specific loss function. The larger the p-value, the less likely it is
that the SPA null hypothesis the base model is not outperformed by all competing models can be rejected.
Table 5
Loss function values for out-of-sample daily volatility forecasts of fuel oil futures on the SHFE (Squared daily returns as
benchmark).
MSE MAE HMSE HMAE
ARFIMA-lnRV 2.793
a
0.778
a
15033.152
a
23.421
a
SV 2.885
b
0.956
b
32944.070
b
37.481
b
GARCH 2.914
c
0.982 35838.158
c
39.671
IGARCH 2.936 0.987 36082.823 39.349
c
GJR 2.924 0.987 38000.103 40.304
EGARCH 2.937 1.024 42072.681 43.265
FIGARCH 2.941 0.978
c
36289.412 39.735
FIAPARCH 2.932 0.993 38232.672 40.970
a
Denotes the smallest loss function value under a specific loss function.
b
Denotes the second smallest loss function value under a specific loss function.
c
Denotes the third smallest loss function value under a specific loss function.
Table 6
SPAp-values for out-of-sample daily volatility forecasts of fuel oil futures on the SHFE (Squared daily returns as benchmark).
Base model MSE MAE HMSE HMAE
ARFIMA-lnRV 0.829 0.527 0.578 0.562
SV 0.506 0.000 0.004 0.000
GARCH 0.109 0.000 0.031 0.000
IGARCH 0.016 0.000 0.024 0.000
GJR 0.097 0.000 0.000 0.000
EGARCH 0.298 0.000 0.004 0.000
FIGARCH 0.097 0.000 0.010 0.000
FIAPARCH 0.160 0.000 0.014 0.000
Notes: The values in bold are the largest p-values under a specific loss function. The larger the p-value, the less likely it is
that the SPA null hypothesis the base model is not outperformed by all competing models can be rejected.
investors in oil markets should rely on the results of SV model. Finally, the multivariable GARCH (MV-GARCH) is usually
adopted in earlier papers to describe the volatility spillover between spot and futures markets, or to calculate the hedge
ratio and hedging effectiveness of oil futures contracts. According to our findings here, multivariable RV or SV model may
be more attractive in solving these problems.
6.6. Robust test using squared daily returns as forecasting benchmark
To assess the robustness of forecasting performance of different volatility models, we use squared daily returns, R
t
2
, as
an alternative forecasting benchmark. Tables 5 and 6 are the corresponding empirical results.
Table 5 shows that, even if the noisy squared daily returns are used as forecasting benchmark, the ARFIMA-lnRV model
also achieves the smallest forecasting error under all four loss functions. In addition, the superiority of the SV model to
the GARCH-class ones is still significant. The SPA results in Table 6 confirm the findings above. In general, even if the noisy
squared daily returns are employed as forecasting benchmark, the performance of the ARFIMA-lnRV model is acceptable.
7. Conclusion
In this article, we adopt intraday high-frequency data to compare the performance of three typical volatility models in the
out-of-sample volatility forecasting of fuel oil futures on the SHFE: the GARCH-type, SVand RVmodels. Different fromearlier
studies on oil markets, we employ a scaled RV measurement based on intraday returns as a proxy for actual daily volatility
and to compute the forecasting error. Our empirical results from the SPA tests indicate that the RV model, ARFIMA-lnRV,
Y. Wei / Physica A 391 (2012) 55465556 5555
based on intraday high-frequency data produces much better volatility forecasting accuracy than the historical volatility
models based on daily returns, particularly the GARCH-type models. Furthermore, among the historical volatility models,
the SV model is clearly superior to many linear and nonlinear GARCH-type models that are able to capture long-memory
volatility and/or the asymmetric leverage effect in volatility. These results remind us that abundant volatility information
is available in intraday high-frequency data, and can be used to construct more accurate volatility forecasting models in oil
markets.
According to the findings presented in this paper, energy economists, policymakers and financial practitioners should
make a volatility forecasting model based on intraday data as their first choice when intraday high-frequency data are
available. Otherwise, if intraday information is not obtainable, then the SV model may be more suitable than the commonly
used GARCH-type ones. Further research on oil markets in connection with problems, such as Value-at-Risk forecasting,
prediction of option prices, volatility spillover and hedging effectiveness of futures, may profit from the use of the
(multivariable) RV or SV models.
Acknowledgments
The author thanks the main editor H. E. Stanley and the five anonymous reviewers for their helpful comments and
suggestions. The author is also grateful for the financial support of the National Natural Science Foundation of China (Nos.
70771097, 71071131 and 71090402), the Program for New Century Excellent Talents in University (No. NCET-08-0826),
the Program for Changjiang Scholars and Innovative Research Team in University (No. PCSIRT0860) and the Fundamental
Research Funds for the Central Universities (Nos. SWJTU11ZT30 and SWJTU11CX137).
References
[1] R.F. Engle, Autoregressive conditional heteroskedasticity with estimates of the variance of United Kingdom inflation, Econometrica 50 (1982)
9871007.
[2] T. Bollerslev, Generalized autoregressive conditional heteroskedasticity, Journal of Econometrics 31 (1986) 307327.
[3] Z.R. Wang, X.H. Chen, Y.B. Jin, Y.J. Zhou, Estimating risk of foreign exchange portfolio: using VaRand CVaRbased on GARCHEVT-Copula model, Physica
A 389 (2010) 49184928.
[4] C.W. Cheong, A.H.S.M. Nor, Z. Isa, Asymmetry and long-memory volatility: some empirical evidence using GARCH, Physica A 373 (2007) 651664.
[5] I. Varga-Haszonits, I. Kondor, Noise sensitivity of portfolio selection in constant conditional correlation GARCHmodels, Physica A385 (2007) 307318.
[6] S. Cook, The impact of GARCH on asymmetric unit root tests, Physica A 369 (2006) 745752.
[7] C.W.S. Chen, M.J. Yang, R. Gerlach, H.J. Lo, The asymmetric reactions of mean and volatility of stock returns to domestic and international information
based on a four-regime double-threshold GARCH model, Physica A 366 (2006) 401418.
[8] B. Podobnik, P.C. Ivanov, I. Grosse, K. Matia, H.E. Stanley, ARCHGARCH approaches to modeling high-frequency financial data, Physica A 344 (2004)
216220.
[9] R.N. Mantegna, H.E. Stanley, Modeling of financial data: comparison of the truncated Lvy flight and the ARCH(1) and GARCH(1, 1) processes, Physica
A 254 (1998) 7784.
[10] H. Chen, C. Wu, Forecasting volatility in Shanghai and Shenzhen markets based on multifractal analysis, Physica A 390 (2011) 29262935.
http://dx.doi.org/10.1016/j.physa.2011.03.035.
[11] M.-K. Lee, J.-H. Kim, J. Kim, A delay financial model with stochastic volatility; martingale method, Physica A 390 (2011) 29092919.
http://dx.doi.org/10.1016/j.physa.2011.03.032.
[12] F. Slanina, A contribution to the systematics of stochastic volatility models, Physica A 389 (2010) 32303239.
[13] I. Florescu, C.G. Psric, A study about the existence of the leverage effect in stochastic volatility models, Physica A 388 (2009) 419432.
[14] P. Jizba, H. Kleinert, P. Haener, Perturbation expansion for option pricing with stochastic volatility, Physica A 388 (2009) 35033520.
[15] A.C. Silva, V.M. Yakovenko, Stochastic volatility of financial markets as the fluctuating rate of trading: an empirical study, Physica A 382 (2007)
278285.
[16] J. Perell, Market memory and fat tail consequences in option pricing on the expOU stochastic volatility model, Physica A 382 (2007) 213218.
[17] G.L. Buchbinder, K.M. Chistilin, Multiple time scales and the empirical models for stochastic volatility, Physica A 379 (2007) 168178.
[18] M.K.P. So, S.W.Y. Kwok, A multivariate long memory stochastic volatility model, Physica A 362 (2006) 450464.
[19] F. Seluk, Free float and stochastic volatility: the experience of a small open economy, Physica A 342 (2004) 693700.
[20] J. Perell, J. Masoliver, N. Anento, A comparison between several correlated stochastic volatility models, Physica A 344 (2004) 134137.
[21] S.J. Koopman, B. Jungbacker, E. Hol, Forecasting daily variability of the S&P100 stock index using historical, realized and implied volatility
measurements, Journal of Empirical Finance 12 (2005) 445475.
[22] T.G. Andersen, T. Bollerslev, Answering the skeptics: yes, standard volatility models do provide accurate forecasts, International Economic Review 39
(1998) 885905.
[23] O.E. Barndorff-Nielsen, N. Shephard, Non-Gaussian OU based models and some of their uses in financial economics (with discussion), Journal of the
Royal Statistical Society. Series B 63 (2001) 167241.
[24] A.B. Schmid, Detrending the realized volatility in the global FX market, Physica A 388 (2009) 18871892.
[25] F. Ren, G.-F. Gu, W.-X. Zhou, Scaling and memory in the return intervals of realized volatility, Physica A 388 (2009) 47874796.
[26] M. Linden, Estimating the distribution of volatility of realized stock returns and exchange rate changes, Physica A 352 (2005) 573583.
[27] D.D. Thomakos, T. Wang, L.T. Wille, Modeling daily realized futures volatility with singular spectrum analysis, Physica A 312 (2002) 505519.
[28] B. Adrangi, A. Chatrath, K.K. Dhanda, K. Raffiee, Chaos in oil prices? evidence from futures markets, Energy Economics 23 (2001) 405425.
[29] P. Agnolucci, Volatility in crude oil futures: a comparison of the predictive ability of GARCHand implied volatility models, Energy Economics 31 (2009)
316321.
[30] C. Aloui, S. Mabrouk, Value-at-risk estimations of energy commodities via long-memory, asymmetry and fat-tailed GARCH models, Energy Policy 38
(2010) 23262339.
[31] J.D. Cabedo, I. Moya, Estimating oil price Value at Risk using the historical simulation approach, Energy Economics 25 (2003) 239253.
[32] C.W. Cheong, Modeling and forecasting crude oil markets using ARCH-type models, Energy Policy 37 (2009) 23462355.
[33] J. Chevallier, Detecting instability in the volatility of carbon prices, Energy Economics 33 (2011) 99110
http://dx.doi.org/10.1016/j.eneco.2010.09.006.
[34] W.M. Fong, K.H. See, A Markov switching model of the conditional volatility of crude oil futures prices, Energy Economics 24 (2002) 7195.
[35] P. Giot, S. Laurent, Market risk in commodity markets: a VaR approach, Energy Economics 25 (2003) 435457.
5556 Y. Wei / Physica A 391 (2012) 55465556
[36] S. Hammoudeh, Y. Yuan, T. Chiang, M. Nandha, Symmetric and asymmetric US sector return volatilities in presence of oil, financial and economic risks,
Energy Policy 38 (2010) 39223932.
[37] I. Henriques, P. Sadorsky, The effect of oil price volatility on strategic investment, Energy Economics 33 (2011) 7987
http://dx.doi.org/10.1016/j.eneco.2010.09.001.
[38] S.H. Kang, S.-M. Kang, S.-M. Yoon, Forecasting volatility of crude oil markets, Energy Economics 31 (2009) 119125.
[39] S.P. Kolos, E.I. Ronn, Estimating the commodity market price of risk for energy prices, Energy Economics 30 (2008) 621641.
[40] H. Mohammadi, L. Su, International evidence on crude oil price dynamics: applications of ARIMAGARCH models, Energy Economics 32 (2010)
10011008.
[41] C. Morana, A semiparametric approach to short-term oil price forecasting, Energy Economics 23 (2001) 325338.
[42] P.K. Narayan, S. Narayan, Modelling oil price volatility, Energy Policy 35 (2007) 65496553.
[43] N.K. Nomikos, P.K. Pouliasis, Forecasting petroleum futures markets volatility: the role of regimes & market conditions, Energy Economics 33 (2011)
321337. http://dx.doi.org/10.1016/j.eneco.2010.11.013.
[44] M. Sadeghi, S. Shavvalpour, Energy risk management and value at risk modeling, Energy Policy 34 (2006) 33673373.
[45] P. Sadorsky, Modeling and forecasting petroleum futures volatility, Energy Economics 28 (2006) 467488.
[46] Y. Wei, Y. Wang, D. Huang, Forecasting crude oil market volatility: further evidence using GARCH-class models, Energy Economics 32 (2010)
14771484.
[47] F.X. Diebold, R.S. Mariano, Comparing predictive accuracy, Journal of Business & Economic Statistics 13 (1995) 253263.
[48] P.R. Hansen, A. Lunde, Consistent ranking of volatility models, Journal of Econometrics 131 (2006) 97121.
[49] P.R. Hansen, A. Lunde, A forecast comparison of volatility models: does anything beat a GARCH(1, 1)? Journal of Applied Econometrics 20 (2005)
873889.
[50] P.R. Hansen, A test for superior predictive ability, Journal of Business & Economic Statistics 23 (2005) 365380.
[51] H. White, A reality check for data snooping, Econometrica 68 (2000) 10971126.
[52] T. Wang, J. Yang, Nonlinearity and intraday efficiency tests on energy futures markets, Energy Economics 32 (2010) 496503.
[53] T.G. Andersen, T. Bollerslev, F.X. Diebold, H. Ebens, The distribution of realized stock return volatility, Journal of Financial Economics 61 (2001) 4376.
[54] N. Shephard, Statistical aspects of ARCH and stochastic volatility, in: D.R. Cox, D.V. Hinkley, O.E. Barndorff-Nielsen (Eds.), Time Series Models in
Econometrics, Finance and Other Fields, in: Monographs on Statistics and Applied Probability, vol. 65, Chapman and Hall, London, 1996, pp. 167.
[55] J.A. Lopez, Evaluation of predictive accuracy of volatility models, Journal of Forecasting 20 (2001) 87109.
[56] T. Bollerslev, R.F. Engle, D. Nelson, ARCH models, in: R.F. Engle, D.L. McFadden (Eds.), Handbook of Econometrics, vol. IV, Elsevier Science B.V.,
Amsterdam, 1994, pp. 29613038.
[57] K.D. West, Asymptotic inference about predictive ability, Econometrica 64 (1996) 10671084.
[58] D.N. Politis, J.P. Romano, The stationary bootstrap, Journal of the American Statistical Association 89 (1994) 13031313.
[59] B. Podobnik, H.E. Stanley, Detrended cross-correlation analysis: a new method for analyzing two non-stationary time series, Physical Review Letters
100 (2008) 084102.
[60] B. Podobnik, D. Horvatic, A.M. Petersen, H.E. Stanley, Cross-correlations between volume change and price change, Proceedings of the National
Academy of Sciences of the United States of America 106 (2009) 2207922084.
[61] N.R. Swanson, G. Elliott, E. Ghysels, J. Gonzalo, Predictive methodology and application in economics and finance: volume in honor of the
accomplishments of Clive W.J. Granger, Journal of Econometrics 135 (2006) 19.
[62] D.O. Cajueiro, B.M. Tabak, Evidence of long range dependence in Asian equity markets: the role of liquidity and market restrictions, Physica A 342
(2004) 656664.
[63] D.O. Cajueiro, B.M. Tabak, Testing for time-varying long-range dependence in volatility for emerging markets, Physica A 346 (2005) 577588.
[64] B. Podobnik, D. Fu, T. Jagric, I. Grosse, H.E. Stanley, Fractionally integrated process for transition economics, Physica A 362 (2006) 465470.
[65] C.W. Cheong, A.H.S.M. Nor, Z. Isa, Asymmetry and long-memory volatility: some empirical evidence using GARCH, Physica A 373 (2007) 651664.
[66] S.H. Kang, S.-M. Yoon, Long memory properties in return and volatility: evidence from the Korean stock market, Physica A 385 (2007) 591600.
[67] S.R. Bentes, R. Menezes, D.A. Mendes, Long memory and volatility clustering: is the empirical evidence consistent across stock markets? Physica A
387 (2008) 38263830.
[68] I. Vodenska-Chitkushev, F.Z. Wang, P. Weber, K. Yamasaki, S. Havlin, H.E. Stanley, Comparison between volatility return intervals of the S&P 500 index
and two common models, European Physical Journal B 61 (2008) 217223.