International Review of Economics and Finance 30 (2014) 101119

Contents lists available at ScienceDirect

International Review of Economics and Finance

journal homepage: www.elsevier.com/locate/iref

Structural breaks and long memory in modeling and forecasting

volatility of foreign exchange markets of oil exporters: The importance
of scheduled and unscheduled news announcements
Walid Mensi a, Shawkat Hammoudeh b,, Seong-Min Yoon c,1
a
Department of Accounting and Finance, Faculty of Management and Economic Sciences of Tunis, El Manar University, B.P. 248, C.P. 2092 Tunis Cedex, Tunisia
b
Lebow College of Business, Drexel University, Philadelphia, PA 19104-2875, United States
c
Department of Economics, Pusan National University, Busan 609-735, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: This paper analyzes the dual long memory properties of four major foreign exchange markets of
Received 1 June 2013 the world oil exporter Saudi Arabia, using the ARFIMAFIGARCH model under several global
Received in revised form 19 October 2013 events. It discerns the impacts of both scheduled and unscheduled news announcements and
Accepted 21 October 2013
structural changes on changing persistence. The results show little evidence of long memory
Available online 6 November 2013
in the conditional mean but provide strong support for long memory in conditional volatility for
the four Saudi exchange rates versus major currencies. Moreover, scheduled news announce-
JEL classification: ments have no significant impact on both expectations and volatility, while unscheduled news
G14
announcements demonstrate significant effects on the conditional volatility for all exchange
G15
rates. Furthermore, we detect at least five structural changes for the exchange rate with the yen
and four for the rest of the exchange rates. The structural breaks seem to have greater impacts
Keywords: on changing persistence, and that the ARFIMAFIGARCH model coupled with the dummy
Dual long memory
variables of the unscheduled news announcements and the structural changes is the most
Structural breaks
News announcements
suitable for examining the long memory processes of these foreign exchange markets in
ARFIMAFIGARCH model in-sample. Finally, the out-of-sample forecasts provide mixed results and indicate that none of
Out-of-sample forecasts the specifications of the volatility model is appropriate for analyzing the LM dynamics in the
Saudi Arabian exchange market. Overall, our results have implications for portfolio managers
and policy makers in oil-producing countries.
2013 Elsevier Inc. All rights reserved.

1. Introduction

Modeling together long memory (LM), structural breaks and news announcements has been an important focus area in
the empirical economics and finance literature that deals with return and volatility persistence and correlations. A better
understanding of the LM process is a key component of determining the optimal investment strategies and portfolio management
because of its relevance to market efficiency.2 The presence of LM in asset returns contradicts the validity of EMH because it

Corresponding author. Tel.: +1 215 895 6673; fax: +1 215 895 6975.
E-mail addresses: walid.mensi@fsegt.rnu.tn (W. Mensi), hammousm@drexel.edu (S. Hammoudeh), smyoon@pusan.ac.kr (S.-M. Yoon).
1
The third author (S. M. Yoon) received support from the National Research Foundation of Korea in a grant provided by the Korean Government
(NRF-2011-330-B00044).
2
The efcient market hypothesis (EMH) in its weak-form suggests that the asset prices reect all available information, implying the impossibility of
generating abnormal returns. As a result, the return series should follow a random walk, meaning no benets from forecasting returns, rendering future asset
returns unpredictable.

1059-0560/$ see front matter 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.iref.2013.10.004
102 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

means the existence of significant correlations between observations that are widely separated in time. Moreover, evidence of the
LM in volatility implies volatility persistence, which suggests that uncertainty is an important determinant of the behavior of
exchange rates. Thus, this stylized fact is important to investors and forecasters but the literature has not reached a consensus on
how to examine it, particularly when it is related to emerging and frontier markets such as those of the oil-exporting countries
which will be the focus of this study. Still, the literature recognizes this fact and has used several methods such as the rescaled
range (R/S), the modified R/S test, the Geweke and Porter-Hudak (GPH) method, the Gaussian semi parametric (GSP) approach
and the exact maximum likelihood (EML) method, among others to test the long-memory hypothesis.
The most recognized model used to examine LM in the conditional mean is the fractional integrated autoregressive moving
average (ARFIMA) model. However, since the pioneering work of Bollerslev (1986), the GARCH-family processes become the
most popular specifications that capture stylized facts including persistence and volatility clustering, among others. In these
processes, the analysis of LM has been extended from focusing on persistence in the conditional mean to also examining
persistence in the conditional volatility of financial time series. Newer specifications of these processes such as the fractionally
integrated GARCH (FIGARCH) class models pay special attention to the LM property because this stylized fact is an important
distributional feature of financial variables like the ones indicated above. Previous studies on this topic examine the LM process in
mean and in variance separately, while shocks that occur in the markets affect both the conditional mean and the conditional
variance. To overcome this shortcoming, the joint ARFIMAFIGARCH model provides the useful way and the ability of capturing
the dual LM in mean and variance in financial time-series. On the other hand, ignoring structural changes begets spurious LM
effects in the data. That is, large variance persistence large variance persistence as in the Integrated GARCH (IGARCH) may be
induced by the failure to identify sudden changes.
When the LM property is prevalent after adjusting for sudden changes and news announcements, then the evidence supports the
hypothesis that structural changes and news announcements are not linked to the assessment of the LM property in the volatility of
the exchange markets under investigation, suggesting that the LM property is an inherent characteristic of the market under
consideration. The foreign exchange market of Saudi Arabia is a good candidate to examine this hypothesis. The kingdom's foreign
exchange reserves depend overwhelmingly on oil revenues which are affected by announcements from OPEC and structural breaks in
the world markets. Moreover, the Saudi Arabian riyal exchange rate (SAR) is pegged to the U.S. dollar but not to other major currencies
like euro, yen and British pound. The Saudi Forex market has experienced many episodes of upheavals that led then to devaluations of
the riyal. Nevertheless, the riyal has been fixed against the dollar at SR3.75:$1 since 1986, but has appreciated/depreciated against
other major floating currencies. The riyal's devaluation at that time was particularly due to lower global oil prices, widening Saudi
budget imbalances, falling Saudi foreign exchange reserves and a deteriorating Saudi balance of payments. In addition, the
devaluation of the riyal continues in 1997 after the total collapse of some East Asian currencies and the collapse in oil prices.3 With
little room for maneuvering, the government took a number of measures to curb spending and raise revenues. In this case, the LM
property should incorporate two important exogenous variables, the structural changes and macroeconomic news announcements
for the foreign exchange markets under consideration. These news announcements may be divided into scheduled and unscheduled
news announcements in order to take account of the potential asymmetric effects on modeling returns and volatility.
Only a few studies have focused on the impact of LM, structural breaks and announcements on the exchange rate markets and
those who have are related to developed markets, despite the sensitivity of exchange rate markets to new information. However, no
studies have been devoted to the LM and structural changes in the exchange market of Saudi Arabia or any other oil-exporting
countries. It will be interesting herein to investigate if the exchange rate markets of major oil exporting countries exhibit such
characteristics. This subject should be interesting since oil prices are sensitive to OPEC announcements and have been subjected to
major disruptions because of global events. Oil prices are also priced in U.S. dollars and pegged to the greenback. It will also be helpful if
this investigation on the link with the dollar exchange rate is done in comparison with the behavior of the exchange rates of other
major currencies. The broad objective of this paper aims to use the ARFIMAFIGARCH model to examine the LM properties in returns
and volatilities of the exchange rates of the world's largest oil exporter, Saudi Arabia. More importantly, we employ this model to test
the impact and the magnitude of scheduled (including OPEC interventions) and unscheduled news announcements on the persistence
of the Saudi exchange rates with respect to the U.S. dollar and other major currencies. Saudi Arabia is not only the largest-oil exporter
and the leading member in OPEC but also a frontier economy that pegs its currency to the dollar but not to the euro, the British pound
and the yen. Moreover, we examine the potential effects of the presence of LM, news announcements (scheduled news such measured
by OPEC announcements, and unscheduled news announcements of major events) and structural breaks on persistence of Saudi
exchange rates. The structural breaks will be detected by using the Iterated Cumulative Sum of Squares (ICSS) algorithm.
In this study, we find significant effects of unscheduled news announcements on the conditional volatility of the daily Saudi
Arabian exchange rates, while scheduled news announcements have no significant effect on both the conditional mean and the
variance of SAR. The ARFIMAFIGARCH model with dummy variables for unscheduled news and structural breaks is the most
suitable for analyzing the LM processes in those foreign exchange markets. Finally, using four symmetric forecasting error
statistics, the out-of-sample analysis provides mixed results and indicates that none of the volatility model specifications is
appropriate for analyzing the LM dynamics in the SAR exchange market.
This research differs from previous studies in at least four aspects. First, a vast amount of empirical works have assessed the
long memory in the first and second moments independently in stock and energy markets, while few studies focus on the dual
long memory in the foreign exchange markets, particularly for the oil-exporting countries. This study essentially examines the

3
The reader can visit http://www.sama.gov.sa for further information.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 103

dual long memory in the return and volatility for a frontier foreign exchange market which has pegged exchange rate with
respect to the dollar but has flexible rates with respect to other currencies for a long time. Second, we discern the impacts of
scheduled and unscheduled announcements on changing return and volatility persistence of this foreign exchange market. Third,
to avoid an overestimation of persistence in the GARCH models, we incorporate the structural breaks as dummy variables in an
ARFIMAFIGARCH model. Finally, we assess the results of this study by evaluating the forecasting performance of five different
model specifications for SAR including a no dummy model, a scheduled news announcements model, an unscheduled news
announcements model, a joint scheduled news announcements and sudden changes model, and a joint unscheduled news
announcements and sudden changes model. We use the out-of-sample forecast analysis based for 6-months-ahead horizons.
This paper is organized as follows. Section 2 presents the literature review. Section 3 describes the econometric methodology
used in this study. Section 4 provides the data and descriptive statistics. Section 5 discusses the empirical results. Section 6
concludes the paper.

2. Literature review

Increasing integration of international financial markets has led to a voluminous literature on the behavior of exchange rates.
Studying the time series properties of currency returns is crucial for financial market participants and risk management. In recent
years, the question of structural breaks, LM and persistence in asset prices has attracted a great deal of attention in terms of both
the first and the second moment of a process. In particular, there has also been an increase in research on LM dynamics in asset
and currency returns. Several methods are used to test the LM properties in exchange rate markets.
To capture the long memory properties and asymmetry features in exchange rate returns, Hwang (2001) applies the asymmetric
fractionally integrated family generalized autoregressive conditional heteroskedasticity (FIFGARCH) for Canadian dollar/US dollar
exchange rate over the period 19731998. The author shows significant asymmetric long memories with mean-reverting impulse
response. Baillie, Cecen, Erkal, and Han (2004) examine the LM of the European exchange rates and show that high-frequency
currency returns volatility is well represented by a FIGARCH model. Dufrnot, Lardic, Mathieu, Mignon, and Pguin-Feissolle (2008)
use cointegration tests to study the adjustment dynamics of the real exchange rates towards the fundamentals for five European
countries (France, Germany, Netherlands, Portugal and UK). The authors obtain conflicting results. Furthermore, they try to
discriminate between the dynamics of the short memory non-linear adjustment (STAR processes) and the dynamics of the long
memory linear adjustment (ARFIMA processes). Dufrnot et al. (2008) use ARFIMA (p, d, q) processes, the Lo (1991) modified R/S
statistic, GPH method and the EML estimation to test the LM of five European exchange rate markets and find contradictory results.
The authors show that the ARFIMA models seem to outperform the ESTAR (exponential smooth transition autoregressive) models.
Gil-Alana (2006) investigates the long-run and the seasonal long memory effects in the monthly Japanese real exchange rate by
means of fractionally integrated techniques. The author shows that the root at zero plays a much more important role than the
seasonal one, though the latter root also plays a significant role. Gil-Alana (2006) concludes that in the presence of shocks stronger
policy actions must be required with respect to those affecting the long-run than those related with the seasonal structure to bring
the series back to its original level.
Recently, Kang, Cheong, and Yoon (2011) investigate the impacts of structural changes on volatility persistence, and then
incorporated these impacts into the bivariate estimation in order to understand the information flow and volatility transmission in
two crude oil markets. The authors use the bivariate GARCH framework with and without structural change dummies and show that
the degree of persistence of volatility was reduced by incorporating these structural changes into the volatility model. Moreover, the
inclusion of information regarding structural changes reduces volatility persistence and improves the understanding of volatility
transmission in crude oil markets. Following the same line, Arouri, Hammoudeh, Lahiani, and Nguyen (2012) examine the potential
of structural changes and LM properties in returns and volatility of four major precious metal commodities (gold, silver, platinum and
palladium) traded on the COMEX markets and show that dual long memory is found to be adequately captured by an ARFIMA
FIGARCH model. In addition, evidence shows that conditional volatility of precious metals is better explained by long memory than by
structural breaks. Chkili, Aloui, and Nguyen (2012) use univariate and multivariate GARCH-type models to investigate the properties
of conditional volatilities of stock returns and exchange rates and show strong evidence of asymmetry and long memory in the
conditional variances of all markets. Aloy, Boutahar, Gente, and Pguin-Feissolle (2012), using the U.S. Dollar, the UK pound and the
German Deutsche mark as numeraires, find that only few bilateral real exchange rates exhibit the LM mean-reverting properties.
More recently, using intra-day data (1, 2, 3, 5, and 10 min) for the USD/GBP spot exchange rate and for different sample periods,
Caporale and Gil-Alana (2013) examine the LM properties and find evidence that a lower degree of integration is associated with
lower data frequencies. When the data are extracted every 10 min, there are several cases with values of the LM parameter d strictly
smaller than 1, implying a mean-reverting behavior. Our literature survey does not find relevant studies on the exchange rate markets
of the oil-exporting countries.
On the other hand, in recent years, there has been a vast literature dedicated to the impact of news announcements or
macroeconomic events on international financial markets but few studies focus on the exchange rate markets. Dornbusch (1980)
and Frenkel (1981) suggest that exchange rate movements basically respond to new information. Laakkonen (2007) examines
the effect of economic announcements on exchange rate volatility while Nikkinen and Vhmaa (2009) investigate the impact of
news events such as central bank interventions on implied exchange rate correlations. Omrane and Heinen (2009) test the effect
of certain categories of news announcements on the quoting activity of individual foreign exchange dealers of the euro/dollar
exchange rate. Jiang, Konstantinidi, and Skiadopoulos (2012) investigate the effect of news announcements on the volatility
spillovers across US and European equity markets and show a significant transmission between these markets. Moreover, they
104 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

conclude that scheduled (unscheduled) news decrease (increase) the stock markets implied volatility. In the same vein, Marshall,
Musayev, Pinto, and Tang (2012) examine the effect of sixteen scheduled US macroeconomic announcements, the release of the
minutes of the Federal Open Market Committee (FOMC), the official US interest rate changes and the Bank of Japan (BOJ)
interventions on the foreign exchange of implied volatility on four major exchange rates USD/EUR, USD/GBP, USD/CHF and USD/
JPY and find a significant connection. The recent studies of Jiang et al. (2012) and Marshall et al. (2012) investigate the impact of
news announcements on implied volatility of international financial markets without the potential effect of both long memory
process and structural breaks. It is therefore important to examine the role played by the news announcements and structural
breaks to analyze the LM dynamics in the SAR exchange rates.

3. Methodology

In the present section, we discuss briefly the different econometric techniques used in this work. First, we present long
memory tests, and second we describe the ARFIMAFIGARCH model. Third, we present the Incln and Tiao (1994) ICSS algorithm
to detect sudden changes in variance. In final step, we analyze the different statistics employed to evaluate the out-of-sample
forecasts, using different evaluation criteria.

3.1. Long memory tests

Several tests have been developed to evaluate the long memory in financial time series. In this study, we develop the analysis
by employing the modified rescaled range (R/S) method and two most popular semi-parametric tests of long memory given by
Geweke and Porter-Hudak (1983) and Robinson and Henry (1999).

3.1.1. Modied rescaled range analysis (R/S)

The rescaled range (R/S) was developed by Hurst (1951). Examining the long-range dependence for the U.S. stock market
returns, Lo (1991) concludes that the result of rescaled range (R/S) analysis was affected by the presence of short memory and
thus might lead to a spurious estimation of long memory. To overcome the drawback of the R/S method, Lo (1991) proposed a
modified version of R/S method. According to Cajueiro and Tabak (2008), let P(t) be the price of an asset at time t and rt be the
logarithmic return denoted by rt PPt1
t
So, consider a sample of continuously compounded asset returns (r1,r2, ,r) and let r
denotes the sample mean (r 1 r ) where is the time span considered. Then, the modified R/S statistic is given by

1 X X
R=S max 1 t rtr min 1 t rtr  1
S t1 t1

where S is the usual standard deviation estimator

 X 1
1 2 2
S r t r : 2
t

According to Hurst (1951), the rescaled range (R/S), for many records in time is determined as follows

H
R=S : 3
2

However, different values of the Hurst exponent imply fundamentally different price behaviors as follows.
- 0 b H() b 0.5 indicates that the time series is anti-persistent. The price behavior follows a mean reverting process.
- A value of H() = 0.5 exhibits that the price behavior is referred to geometric Brownian motion (random walk or short
memory).
- 0.5 b H() b 1 denotes that the time series is persistent with long-term memory (or more generally, long-range dependency).

3.1.2. The GPH test

The GPH estimator of the long memory parameter, d, is based on the following periodogram
h  i h  i
2
ln I w j 0 1 ln 4sin w j =2 e j 4

p
where wj = 2/T, j = 1,2,n. ej is the residual term; wj refers to the n ^
T Fourier frequency; d GPH estimates of d, being equal to
^ and I(wj) the sample periodogram defined as
1

  1 XT w 2

I wj  t1 rt e j;t  : 5
2T
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 105

3.1.3. The GSP test

The GSP test assesses the long memory in a covariance stationary series using a semi-parametric approach. The GSP estimator
of the long memory parameter for a covariance stationary series is expressed as follows

12H
f w GW as w0 6

where 0.5 b H b 1 and 0 b G b , f(w) is the spectral density of t.

Similar to the estimation procedure of the GPH estimator, the periodogram with respect to the observations t, t = 1,T is
given by

1 XT 
it 2
I  t1 r t e  : 7
2T

The estimate of the long memory parameter H is then given by

H argmin1 H 2 RH
8
0 b 1 b 2 b 1

8  9
<1 X I j = 1 Xm
m
Rh log 2H1 log j 9
:m j1 12h
j
; m j1

 

where m 0; n 2 and j = 2j/T.
According to Robinson and Henry (1999), the semi-parametric estimator of the long memory parameter is consistent and
asymptotically normal.

3.2. ARFIMAFIGARCH model

In this research, we adopt the ARFIMAFIGARCH model which allows for a long range memory.
The ARMA (m, n) model with news dummy variables is specified as follows.
Xm Xn
rt k1
k r tk
h1 h th
Newsi;t 10

The ARFIMA (m, d, n) can be expressed as follows:

 
d
L1L r t Newsi;t Lt ; 11

q
t zt ht ; zt eN0; 1 12

where rt is the return for exchange rates at time t and t is the independently distributed error with variance 2t , d is the LM
parameter, L, denotes the lag operator, and (L) = 1 1L 2L2 mLm and (L) = 1 + 1L + 2L2 + nLn are,
respectively, the autoregressive (AR) and the moving-average (MA) polynomials. {zt} a sequence of independent standard normal
variables, {ht} a positive time dependent conditional variance.
The ARFIMA process is non-stationary when d 0.5, while it exhibits an LM for 0 b d b 0.5. This process exhibits a short
(intermediate) memory for d = 0 (d b 0.5).
We follow Baillie, Bollerslev, and Mikkelsen (1996) by introducing the FIGARCH (p, , q) process

2
L1L t 0 1Lvt Newsi;t : 13

This model provides a greater flexibility for modeling the conditional variance and can distinguish between the covariance
stationary GARCH model for = 0. If = 1, we obtain the non stationary IGARCH model while for 0 b b 1 the FIGARCH model
is sufficiently flexible to allow an intermediate range of persistence.
The parameters of the ARFIMAFIGARCH model can be estimated using nonlinear optimization procedures to maximize
the logarithm of the Gaussian likelihood function. The log-likelihood of normal distribution (LNorm) is given by the following
equation

T h i
1X 2
LNorm ln 2 ln ht zt : 14
2 t1
106 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

The estimation procedure of ARFIMAFIGARCH family models requires a minimum number of observations. This minimum
number is related to the truncation order of the fractional differencing operators (1 L)d and (1 L) (Kang & Yoon, 2013).
Following the standard procedure described in the econometric literature, the truncation order of the infinite (1 L)d and (1 L) is
set to 1000 lags:

1X
000
d kd k
1L L : 15
k0
k 1 d

where (.) is the gamma function. The fractional differencing operator (1 L) follows the same procedure as in the conditional
variance specification in Eq. (13).

3.3. ICSS algorithm to detect sudden changes in variance

To investigate the existence of structural breaks in the SAR exchange rate data series, we employ the Incln and Tiao (1994)
ICSS algorithm for each series. The ICSS has been extensively used to identify the points of shocks/sudden changes in the volatility
of return series including, among others, Aggarwal, Inclan, and Leal (1999), Arouri et al. (2012), Hammoudeh and Li (2008), Kang
et al. (2011), Kumar and Maheswaran (2013), and Vivian and Wohar (2012). The ICSS algorithm allows us to detect both the
beginning and the ending of volatility regimes for the series.
Let us assume that t denotes a time series with zero mean and with the unconditional variance 2t . For each interval, the
variance is given by j or j = 1, 2, , NT where NT is the total number of changes or jumps in the variance in the T observations.
The set of those points is given by 1bK 1 bK 2 bbK NT . The variance over the NT intervals is defined as follows:
8 2
> 0 ; 1b t b K 1
>
>
>
< 2
t 1; K1 b t b K2 :
2
16
>
>
>
>
: 2 ; K b t b T
NT KN
T

We use the cumulative sum of squares procedure in order to assess the number of changes or jumps in the variance and the
time point for each variance shift. The cumulative sum of the squared observations from the start of the series to the kth point in
time is defined as follows:
Xk 2
Ck ;
t1 t
where k 1; 2; ; T: 17

The Dk statistic is given by


Ck k
Dk ; where D0 DT 0; 18
CT T

and CT is the sum of squared residuals from the whole sample period.
If no changes or jumps in the variance occur, the Dk statistic will oscillate around zero. In contrast, if there is at least one
sudden change in the variance of the series, the Dk statistic deviates from zero. These critical values define the upper and
lower limits for the drifts. If the maximum of the absolute value of the statistic Dk is greater than the critical value, then the
null hypothesis of no sudden change in variance is rejected. In this case, by letting k be the value at which maxk|Dk| is reached,
q
  q
 
and if maxk T = jD j exceeds the critical value, then k taken as an estimate of the change or jump point. The term T=
2 k 2 is
used to standardize the distribution.
q
 
The critical value of 1.358 is the 95th percentile of the asymptotic distribution of maxk T = jD j. Therefore, the upper and
2 k

lower boundaries can be established at 1.358 in the Dk plot. A change or jump point in variance is identified, if Dk exceeds these
boundaries. However, if the series harbors multiple change points, the Dk function alone will not be sufficiently powerful to detect
the change points at different intervals. To overcome this shortcoming, Incln and Tiao (1994) amended an algorithm that uses
the function to search systematically for change points at different points in the series. This algorithm works by evaluating the
Dk function over different time periods, and those different periods are determined by the breakpoints, which are themselves
identified by the Dk plot.

3.4. Measures of forecasting accuracy

Forecasting financial volatility is of particular concern for derivative pricing, portfolio optimization and value-at-risk analysis.
In accordance with the relevant literature (e.g., Brailsford & Faff, 1996; Brooks & Persand, 2003; Degiannakis, 2004), the daily
ex post volatility (variance) is measured by the squared returns.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 107

To measure the forecasting accuracy, we calculate the mean of absolute errors (MAE), the root mean squared errors (MSE), the
Theil inequality coefficient (TIC), and the logarithmic loss errors (LL) of the volatility forecasts, as follows:

1X T
MAE j F At j; 19
T t1 t

1X T
2
MSE F At ; 20
T t1 t
v
u T
u1 X 2
t F At
T t1 t
v;
TIC v 21
u u T
u X
u1 T 2 u 1X 2
t Ft t A
T t1 T t1 t

T  2
1X F
LL ln t ; 22
T t1 At

in which T is the number of forecasting data points and Ft denotes the returns (volatility) forecast for day t, whereas At signifies
actual returns (volatility) on day t. A smaller forecast error statistic means the superior forecasting ability of a given model.
The four error statistics used in this study are effective for comparing forecasts but suffer from the lack of a statistical
framework. More precisely, as the difference between two forecasts may not be statistically significantly different from zero,
the smaller forecast error statistic is not necessarily superior to the other statistics of competing models (Chortareas, Jiang, &
Nankervis, 2011). To take such consideration into account, it is important to show that any decrease in forecast errors is statistically
significant instead of comparing the forecast error statistics among the forecasting models. In that regard, we use the equal
accuracy test initially proposed by Diebold and Mariano (1995) for comparing the forecast accuracy for at two forecasts.
Having generated n, h-step-ahead forecasts from two different forecasting models, the forecast errors from two competing
models e1,t + h and e2,t + hwhere t = 1, 2, , n and h is the number of trading days in a given forecast horizon starting on day t,
corresponding to 180-day-ahead horizons. L(e1,t + h) and L(e2,t + h), represent their associated loss functions. The null hypothesis
of equal forecast accuracy for two forecasts can be represented as E(L(e1,t + h)) = E(L(e2,t + h)) or E[dt + h] = 0 where dt + h =
L(e1,t + h) L(e2,t + h) is the loss differential. Thus, the equal accuracy null hypothesis is equivalent to the zero expected loss
differential.
The mean of differences between forecast errors d = n1 nt = 1dt + h as the approximate asymptotic variance
  h Xk1 i
1
V d n 0 2 k1 k 23

where k the k auto-covariance of dt + h which can be estimated as

Xn   
^ n1
dt d dtk d ; 24
tk1

The Diebold and Mariano (1995) test statistic for the null hypothesis of equal forecast accuracy is
h i0:5
DM Vd
d d: 25

The DM statistic follows an asymptotic standard normal distribution under the null hypothesis. In this research, the DM test is
conducted using the loss differential based on the MAE, MSE, TIC and LL of the different forecasting models.

4. Data and descriptive statistics

4.1. Data

In this study, we use daily closing spot values for the four exchange rates SAR/USD, SAR/EUR, SAR/GBP and SAR/JPY. The Saudi
currency is measured in terms of these major currencies that are the most liquid and widely traded (Bekiros & Marcellino, 2013).
The sample spans the period December 13, 2003December 31, 2009 as determined by the data availability of announcements.
The choice of this period is also motivated by its high volatility and the presence of major global events in it. More reasons for
selecting the Saudi Arabian exchange rate market include the facts that the country's market is the largest Forex market in the
Gulf Cooperation Council (GCC) and the Middle East region. The country has very large foreign exchange reserves which enable
it to maintain a stable peg, is ranked as 23rd largest economy of the world and is a member of the G20 countries. As indicated
108 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

earlier, this largest global oil exporter has a fixed exchange rate regime, having its riyal pegged to the U.S. dollar but flexible to
other major currencies. The Saudi government is also the principal source of foreign exchange proceeds. These considerations add
to the fact that the world's largest oil exporter is subject to scheduled announcements from OPEC and structural changes in the
world's economies and financial markets as explained before.
The exchange rate data is extracted from the Oanada website. We also consider scheduled news announcements measured by
OPEC conferences. The data related to these announcements is compiled from Press Releases published by the OPEC Secretariat
and the compiled list of official announcements on production decisions. In contrast, the unscheduled news announcements
include general but major financial and geopolitical events such as the subprime crisis, Iran sanctions, Gulf wars, and terrorism.4
Both scheduled and unscheduled news may affect significantly currency markets. The SAR exchange rates are used for at least two
reasons. First, the scheduled news announcements integrate OPEC's interventions and Saudi Arabia is the largest oil exporter in
OPEC. Cairns and Calfucura (2012) discuss the actions of OPEC and Saudi Arabia in terms of their objectives and their technical
and social constraints. Second, Saudi Arabia engages in a great volume of transactions with US, Europe and Japan. The
continuously compounded daily returns are computed by taking the difference in the logarithm of two consecutive prices or rates.
Specifically, the foreign exchange returns are given by

r t ln pt =pt1  100 26

where rt is the return of each price or rate at time t, Pt is the current price or rate, and Pt 1 is the price or rate on the previous day.
Each of scheduled and unscheduled news announcements is considered as an event represented by a dummy variable, coded
one or zero for each event. We specify these variables as follows


1 OPEC news announced on day t
dsheduled;t 27
0 otherwise


1 unscheduled news announced on day t
dunscheduled;t : 28
0 otherwise

Figs. 1 and 2 illustrate the time series of rates and returns of the four SAR exchange rates. As seen in these figures, the exchange
rate series exhibit heteroskedastic characteristics (shown formally in Table 1). More precisely, the plot of exchange return series
displays that heteroskedaticity is significantly present for all these series. On the other word, big shocks are clustered together in
the Saudi foreign exchange market (periods of high volatility as well as those of lower volatility occur in clusters). Intuitively, we
choose to model the series as a GARCH type model for capturing the dynamics of exchange rate volatility.

4.2. Descriptive statistics

Table 1 presents the statistical properties summary of the SAR/USD, SAR/EUR, SAR/GBP, and SAR/JPY exchange rate returns
over the period under consideration. As shown in Panel A of Table 1, only the SAR/GBP rate has a positive average return, implying
depreciation of the target currency while for the rest of rates the average returns are negative, suggesting appreciation of the
Saudi currency. More exactly, the SAR/GBP rate experiences both the highest average return (i.e., currency depreciation) and
standard deviation over the study period. Moreover, all return series are found to be leptokurtic (i.e., fat tails), with the exception
of the SAR/USD which shows a Kurtosis with a value less than three (i.e., a platykurtic distribution for the SAR/USD rate and
leptokurtic distributions for the remainder of the rates). The SAR/USD, SAR/EUR and SAR/JPY return series are skewed to the left,
whereas the SAR/GBP rate is skewed to the right. As indicated by the Jarque-Bera test for normality, none of the exchange rate
return series is normally distributed. In addition, we test the validity of the hypothesis of a white-noise process for sample
returns, using the LjungBox test statistic of the return residuals (Q(10)) and the squared residual return (Qs(10)). We show
significant autocorrelation in all cases, implying that the past behavior of the market may be more relevant.
Before fitting the series, it is helpful to check the stationarity of the return series. We apply the standard unit root tests based on
the augmented DickeyFuller (ADF), and PhillipsPeron (PP) and the stationarity test based on the Kwiatkowski, Phillips, Schmidt,
and Shin (KPSS) test. The null hypothesis of the ADF and PP tests is that a time series contains a unit root, whereas the KPSS test has
the null hypothesis of a stationary process. The results of these tests are given in Panel B of Table 1. Both ADF and PP (KPSS) tests
reject (do not reject) the null hypothesis of a unit root (stationarity) at the 1% significance level. Thus, the exchange rate returns are
a stationary series. We also examine the existence of autoregressive conditional heteroskedasticity (ARCH) using the test described
in Engle (1982). Additionally, Panel C of Table 1 reports the Lagrange Multiplier test for conditional heteroskedasticity. All series
exhibit the ARCH effects, and therefore the estimation of a GARCH model is appropriate for modeling some stylized facts such as
fat-tails, clustering volatility and persistence for the foreign exchange rate (SAR) returns.

4
The list of scheduled and unscheduled news announcements is available from the authors upon request.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 109

Fig. 1. Daily spot exchange rates.

5. Empirical results

5.1. Long memory test

To check whether there is long memory in exchange rate return and return volatility series, we first use a battery of LM tests:
the HurstMandelbrot R/S, Lo's modified R/S, GPH, and GSP tests as described earlier. Table 2 reports the estimation results of the
long memory tests applied to the returns and squared returns (as proxy of volatility). As shown for the (simple) returns in this
table, we should not reject the null hypothesis of random walk or short memory for three exchange rate return series including
the SAR/EUR, SAR/JPY and SAR/GBP rates, while it's rejected for the SAR/USD rate which is the only fixed exchange rate for the
Saudi currency. This implies long range dependence or persistence in return (i.e., LM in means) which violates the weak efficiency
hypothesis. The investors may exploit this inefficiency to generate abnormal returns. For the squared returns (volatility), we
strongly reject the null hypothesis of no persistence at the 1% level whatever the considered exchange rate is. According to the
four tests, all series exhibit a long-range dependence or persistence in volatility. For example, the estimated values of the Hurst
exponent exceed 0.5 and close to one, which suggests high persistence in these exchange rate series. On the other hand, for both
semi-parametric (GPH and GSP) tests, the estimation values of the LM parameter d are significantly different from zero for all
series, indicating that the volatility process exhibits an LM property. As a result, the volatility of exchange market decays slowly
and can be better modeled by the FIGARCH approach due to its ability to test the LM process in financial time-series. In the whole,
we find dual LM for the SAR/USD rate and also strong evidence of long-range dependency in volatility for the other three rates for
each of the applied tests.

5.2. Estimation of dual long memory model without and with news announcements

The present subsection discusses the empirical results of the ARFIMAFIGARCH model without and with the dummy exogenous
variables as measured by both the scheduled and unscheduled news announcements.
110 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

Fig. 2. Sudden changes in return dynamics of the four foreign exchange markets. Note: Bands (dot lines) are at 3 standard deviations around where structural
change points are estimated by the ICSS algorithm.

Table 1
Statistical properties for exchange rate return series.

Exchange rate return series SAR/USD SAR/EUR SAR/GBP SAR/JPY

Panel A: Descriptive statistics

Mean 0.000063 0.008018 0.005064 0.007958
Minimum 0.8571 3.4660 3.1541 3.5312
Maximum 0.5218 2.5554 4.0579 2.7763
Std. dev 0.137696 0.525207 0.555100 0.536360
Skewness 0.808592 0.102541 0.534324 0.237615
Excess Kurtosis 2.697837 3.608499 5.273588 3.129314
JarqueBera 246.93*** 1040.17*** 2190.75*** 816.50***
Q(10) 159.23*** 44.18*** 102.47*** 53.15***
Qs(10) 151.00*** 344.25*** 622.14*** 233.19***

Panel B: Unit root tests

ADF 24.88*** 38.15*** 34.93*** 38.32***
PP 54.66*** 38.09*** 34.50*** 38.05***
KPSS 0.048 0.055 0.207 0.184

Panel C: Heteroskedasticity test

ARCH-LM(10) test 12.464*** 21.522*** 32.815*** 13.549***

Notes: Q(10) and Qs(10) refer to the empirical statistics of the LjungBox test for autocorrelation of returns and squared returns series, respectively. ADF and PP
are the acronyms for the Augmented Dickey-Fuller and the Phillips-Perron unit root tests' statistics, respectively, while KPSS refers to the Kwiatkowski, Phillips,
Schmidt and Shin stationarity test statistic. Mackinnon's 1% critical value is 3.432 for the ADF and PP tests. The critical value for the KPSS test is 0.739 at the 1%
significance level. ***Significance level at the 1% level.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 111

Table 2
Results of the long memory tests for returns and squared returns.

Returns Squared returns

SAR/USD SAR/EUR SAR/GBP SAR/JPY SAR/USD SAR/EUR SAR/GBP SAR/JPY

Panel A: HurstMandelbrot R/S test

Hurst exponent H 0.5154 0.5595 0.5516 0.5542 0.7824 0.8330 0.8274 0.7418
(0.0048) (0.0028) (0.0030) (0.0029) (0.0037) (0.0069) (0.0080) (0.0042)
Test statistic 0.4964*** 1.1468 1.7171 1.2776 2.4420*** 4.6862*** 4.9358*** 3.9048***

Panel B: Lo's modified R/S test

Test statistic (q = 1) 0.6562*** 1.0767 1.5705 1.1959 2,2566*** 4.2054*** 4.3978*** 3.6204***
Test statistic (q = 5) 0.9951 1.0629 1.5036 1.1790 1.9305** 3.5364*** 3.5689*** 3.1254***

Panel C: GSP test

d ( = 0.45) 0.5680 0.0732 0.1274 0.0277 0.2152 0.8194 0.8517 0.9614
(0.1213)*** (0.0928) (0.0928) (0.0913) (0.1213)* (0.0928)*** (0.0928)*** (0.0913)***
d ( = 0.50) 0.4496 0.0349 0.1146 0.0862 0.3663 0.6253 0.6768 0.6538
(0.1021)*** (0.0762) (0.0772) (0.0754) (0.1021)*** (0.0762)*** (0.0772)*** (0.0754)***
d ( = 0.55) 0.2707 0.0332 0.0932 0.0169 0.5316 0.4805 0.5178 0.5299
(0.0870)*** (0.0630) (0.0635) (0.0625) (0.0870)*** (0.0630)*** (0.0635)*** (0.0625)***
d ( = 0.60) 0.2033 0.0087 0.0650 0.0155 0.4201 0.4579 0.5259 0.3352
(0.0737)*** (0.0518) (0.0527) (0.0516) (0.0737)*** (0.0518)*** (0.0527)*** (0.0516)***

Panel D: GPH test

d ( = 0.45) 0.5532 0.1194 0.2713 0.0066 0.1687 0.8574 0.8982 1.0798
(0.1142)*** (0.2441) (0.1373)** (0.1423) (0.1350) (0.0965)*** (0.0915)*** (0.1350)***
d ( = 0.50) 0.3221 0.0488 0.1841 0.1022 0.3270 0.6646 0.7413 0.7307
(0.1576)*** (0.1748) (0.1093)* (0.1073) (0.1101)*** (0.0797)*** (0.0881)*** (0.1229)***
d ( = 0.55) 0.2285 0.0786 0.1050 0.0562 0.4758 0.4996 0.4948 0.5243
(0.1182)** (0.1262) (0.0852) (0.0928) (0.0957)*** (0.0705)*** (0.0749)*** (0.1024)***
d ( = 0.60) 0.1850 0.0580 0.0532 0.0196 0.4003 0.4760 0.5217 0.2301
(0.1017)** (0.0982) (0.0672) (0.0788) (0.0763)*** (0.0582)** (0.0649)*** (0.0876)***

Notes: The numbers in parentheses are standard deviation of estimates. q in Lo's modified R/S test is number of lag of autocorrelation.

Tables 35 report the results of the ARFIMAFIGARCH model without news announcements, with scheduled news
announcements and with unscheduled news announcements, respectively. We find little evidence of persistence in the
conditional mean for all exchange rates under all these model specifications, with the exception of the SAR/USD rate. However,
the long memory parameter d is negative and statistically significant at the 1% significance level for the fixed SAR/USD rate, which
shows evidence of anti-persistence (negative dependence) for the SAR/USD both without and with scheduled and unscheduled
news announcements. More exactly, the negative estimated values of the parameter d is 0.403 for the model without dummy
variables, 0.440 for the model with scheduled news, and 0.486 for the model with unscheduled news. For the remainder of
returns of the exchange rates, the LM parameter d is not statistically significant, suggesting the validity of weak-form efficiency or
lack of predictability of those exchange rates. On the other hand, both scheduled and unscheduled news announcements have no
significant effects on the conditional means of all four exchange rates.
Turning now to the variance equation, the LM parameters are positive and highly significant for all four exchange rates and
for each of the model specifications, suggesting the presence of long-range dependence or persistence in volatility and confirming
the results of the LM tests for squared returns. For the model without news, the values range from 0.427 (for the SAR/JPY) to 0.878
(the SAR/EUR), suggesting that the exchange rates display little evidence to revert toward the variance mean. This result is
consistent with Arouri et al. (2012) for the precious metal markets. Note that the LM parameter in the conditional variance
decreases when we consider news announcements, and the highest decrease is recorded in the unscheduled news specification of
the ARFIMAFIGARCH model. This implies that including dummies in the model reduces the persistence and increases the
reversion towards the variance mean. Taking for example the SAR/JPY rate, the value of parameter is low (i.e., 0.5), reaching
0.427 for the model without the dummy news announcement variables while decreasing to 0.369 when the unscheduled news
announcements are incorporated. On the other hand, we show that the scheduled news has no significant effect on both the
conditional mean and conditional volatility process for the four exchange rates. In contrast, the unscheduled news has a
significant impact on all rates with the exception of the SAR/EUR rate. However, the coefficient is negative for the managed SAR/
USD rate. That is, the unscheduled news announcements contribute to decrease volatility while they increase the conditional
volatility for both the SAR/GBP and SAR/JPY rates. Tables 35 also give the diagnostic tests that measure the accuracy of the model
specifications. We apply the residual tests and the model selection criteria to measure the specification accuracy. For the residual
tests, we use the BoxPierce test, Q(5) for up to the 5th-order and Q(20) up to the 20th-order serial correlations in the residuals,
the LM ARCH (5) and the LM ARCH (10) tests for the presence of ARCH effects in the residuals up to lag 10 and 20.5 On the other

5
There is no common way regarding the determination of the optimal time lags in both the BoxPierce test and LM ARCH test statistics. In this study, we
according to the practical approach consider the 5 lags (one week business trading days) and 20 lags (monthly business trading days) for Q-statistic test, and the
5 lags (one week business trading days) and 10 lags (two week business trading days) for the LM ARCH test.
112 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

Table 3
Estimation results of the ARFIMA (m, d, n)FIGARCH (p, , q) models without news announcements.

SAR/USD SAR/EUR SAR/GBP SAR/JPY

Mean equation
.0005 0.0212 0.0073 0.0043
(0.0002)** (0.0079)*** (0.0095) (0.0093)
1 .985 0.213 0.220 0.223
(0.007)*** (0.048)*** (0.055)*** (0.049)***
1 516
(0.152)***
2. .422
(0.152)***
d .403 0.056 0.042 0.051
(0.144)*** (0.037) (0.045) (0.036)

Variance equation
0 0030 0.0004 0.0018 0.0053
(0.0023) (0.0003) (0.0013) (0.0029)*
1 0.053 0.155 0.308 0.495
(0.093) (0.075)** (0.133)** (0.084)***
1 0.429 0.950 0.838 0.765
(0.204)** (0.013)*** (0.092)*** (0.056)***
565 0.878 0.639 0.427
(0.162)*** (0.067)*** (0.219)*** (0.088)***

Diagnostic test
logL 4.352 1249.722 1233.438 1398.225
AIC 1.9210417 1.315251 1.366121 1.436835
BIC 1.8550027 1.335600 1.387338 1.456800
Q(5) 28.7369*** 2.30839 2.93362 4.02333
Q(20) 42.2606*** 12.8527 19.3884 18.5570
Qs(5) 10.5721** 3.17388 1.85459 16.7039***
Qs(5) 17.7857 10.0977 10.8497 33.2822**
LM(5) 2.0081** 0.64213 0.36171 3.0856***
LM(10) 1.7425* 0.58252 0.35030 1.9483**
P(40) 140.5659*** 73.2805*** 49.6828** 78.0286***

Notes: This table reports the results of the quasi-maximum likelihood estimation of the ARFIMAFIGARCH class model for the daily exchange rate returns. We
select the ARFIMA(2, d, 1)-FIGARCH(1, , 1) model for the SAR/USD rate returns and the ARFIMA(1, d, 0)-FIGARCH(1, , 1) model for the SAR/EUR, SAR/GBP, SAR/
JPY rate returns. logL and BIC are the maximum likelihood and Bayesian information criterion, respectively. Q(20) and Qs(20) are the Box-Pierce i.i.d. tests on
standardized and squared standardized returns. The LM(10) is the LM test for ARCH effects. The P(40) is Pearson goodness-of-fit statistic. P-values are given in the
parentheses. ***, ** and * indicate rejection of the null hypothesis at the 1%, 5% and 10% significance levels, respectively.

hand, both two popular model selections namely the Akaike information criterion (AIC) and the Bayesian information criterion
(BIC) are used to select the best model specification. These tables also provide the Pearson goodness-of-fit-statistic P (40) test up
to lag 40.
As indicated in the third panel of Table 3, the results of BoxPierce Q(5) and Q(20) show no significant serial correlations for
three exchange rates, namely the SAR/EUR, SAR/GBP and SAR/JPY rates, LM(5) and LM(10) show no remaining ARCH effect for
both the SAR/EUR and SAR/GBP rates. Thus, the results are valid when the dummy news variables are considered in the model
specification. Overall, these diagnostic and model selection tests demonstrate that the ARFIMAFIGARCH model are suitable for
investigating the long memory volatility process in the SAR/EUR and SAR/GBP rates and there is no significant evidence of
misspecification in the ARFIMAFIGARCH model. Furthermore, the minimum or low values of the AIC and BIC are obtained from
the ARFIMAFIGARCH model with unscheduled news, indicating that this is the best model that captures the LM dynamics in
these exchange rate markets.
In the next subsections, we investigate the presence of sudden changes in the Saudi Arabia exchange market. We then model
the conditional volatility in the presence of structural breaks. To be more convincing, we end this investigation by carrying out an
out-of-sample forecast analysis.

5.3. Detecting sudden changes in variance

Evidently, adjusting the volatility models by considering structural breaks leads to more accuracy by reducing the degree of
volatility persistence. To reinforce this aim, we investigate the presence of structural breaks in four exchange rates using the test
proposed by Incln and Tiao (1994). We apply this test for the daily period spanning from 2003 to 2009. Nevertheless, Fig. 2
displays the return behavior for the four exchange rates, and points out the points of structural changes and the 3 standard
deviation bands around them. On the other hand, Table 6 presents the estimation results of the structural breaks in volatility of
the SAR exchange rates as detected by the ICSS algorithm. As shown in this table, we undoubtedly reject the null hypothesis of no
structural breaks. More exactly, we show five structural breaks for the SAR/JPY rate and at least four breaks for the other rates.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 113

Table 4
Estimation results of the ARFIMA (m, d, n)FIGARCH (p, , q) model with scheduled news announcements dummies.

SAR/USD SAR/EUR SAR/GBP SAR/JPY

Mean equation
.0005 0.0207 0.0070 0.0060
(0.0003)* (0.0081)** (0.0099) (0.0094)
1 501 0.210 0.215 0.223
(0.194)*** (0.048)*** (0.054)*** (0.049)***
1 .024
(0.213)***
2 320
(0.100)***
d .440 0.055 0.037 0.050
(0.091)*** (0.037) (0.044) (0.036)
Scheduled news 0.0087 0.0070 0.0358 0.0902
(0.0127) (0.0778) (0.0911) (0.0759)

Variance equation
0 0029 0.0002 0.0015 0.0060
(0.0018) (0.0003) (0.0012) (0.0034)*
1 0.136 0.174 0.337 0.490
(0.128) (0.086)** (0.094)** (0.080)***
1 0.460 0.945 0.822 0.757
(0.080)*** (0.016)*** (0.062)*** (0.064)***
555 0.856 0.595 0.422
(0.122)*** (0.079)*** (0.144)*** (0.086)***
Scheduled news 0.0025 0.0174 0.0606 0.0439
(0.0025) (0.0225) (0.0478) (0.0560)

Diagnostic test
logL 0.195 1249.284 1232.305 1397.219
AIC 1.867094 1.316885 1.367075 1.437851
BIC 1.786380 1.343049 1.394355 1.463520
Q(5) 32.3302*** 2.39399 3.04594 4.04092
Q(20) 54.6168*** 13.1203 19.1834 18.1755
Qs(5) 21.0217*** 3.09223 1.75741 16.3063***
Qs(20) 25.5658 9.92322 10.9924 32.3603**
LM(5) 4.0603*** 0.62553 0.34175 3.0262***
LM(10) 2.4124** 0.58667 0.35347 1.8695**
P(40) 97.293*** 75.3736*** 51.3128*** 77.4560***

Notes: See the under-table-notes of Table 3.

Additionally, two increases in the standard deviation in sudden changes are detected for the SAR/USD and SAR/JPY rates, while
one structural change increase is for the SAR/EUR and the SAR/GBP. The date for increases in structural breaks corresponds to the
20072008 financial crisis and the increases in commodity prices during the period 20032007 which starts with the 2003 Iraq
war. Thus, the shifts in volatility for all exchange rates correspond to extreme events. Our results are consistent with previous
works such as Kang et al. (2011), among others.

5.4. Estimation results of dual long memory model with news announcements and sudden changes

Previous works show the crucial role plays by the structural breaks in modeling volatility. The pioneering study of Lastrapes
(1989) finds the importance of incorporating structural breaks in volatility models to get an exactitude level of persistence and
avoid the overestimation of persistence. Since this date, several studies still acknowledge the relevance of this specification to get
an accurate degree of volatility persistence (Kang et al., 2011; Ewing & Malik, 2013, among others).
The estimation results of the ARFIMAFIGARCH model with the dummy variables for scheduled and unscheduled news
announcements and structural changes are reported in Tables 7 and 8, respectively. As presented in these tables, we find a dual
long memory in the mean and variance for both the SAR/USD and SAR/EUR exchange rates as explained earlier. The evidence of
the LM property in volatility shows that uncertainty is an important determinant of the behavior of exchange rates. We conclude
that the SAR markets are not stable, and hence it offers an opportunity to USD and EUR investors and traders to add some risk to
their risk management strategies in return for greater expected profits. On the whole, the scheduled news announcements have
no significant impacts on the return and volatility for all exchange rates and as a result on the market reactions. These results are
explained by the fact the foreign exchange markets react before the official date of announcements. Additionally, the unscheduled
news announcements decrease significantly the conditional volatility of the managed SAR/USD exchange rate while they increase
it for the flexible SAR/EUR and SAR/GBP rates. This result is explained by the fact that the SAR exchange markets do not anticipate
(guess) this kind of news, and thus their effects are pronounced in the market as traders may use this set of information to earn
more profit. According to these results, the investors react to the unscheduled news announcements than the scheduled news in
114 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

Table 5
Estimation results of the ARFIMA (m, d, n)FIGARCH (p, , q) model with unscheduled news announcements dummies.

SAR/USD SAR/EUR SAR/GBP SAR/JPY

Mean equation
.0008 0.0207 0.0052 0.0040
(0.0004)** (0.0079)*** (0.0099) (0.0098)
1 519 0.214 0.215 0.207
(0.170)*** (0.048)*** (0.052)*** (0.050)***
1 .004
(0.194)***
2 313
(0.087)***
d .486 0.059 0.038 0.043
(0.140)*** (0.037) (0.042) (0.038)
Unscheduled news 0.0024 0.0086 0.0497 0.0277
(0.0038) (0.0463) (0.0591) (0.0704)

Variance equation
0 0024 0.0000 0.00004 0.0020
(0.0015) (0.0003) (0.00103) (0.0022)
1 0.138 0.196 0.358 0.567
(0.103) (0.098)** (0.140)** (0.083)***
1 0.441 0.942 0.846 0.792
(0.083)*** (0.019)*** (0.076)*** (0.041)***
. 543 0.832 0.602 0.369
(0.108)*** (0.093)*** (0.204)*** (0.078)***
Unscheduled news 0.0006 0.0170 0.0664 0.1309
(0.0000)*** (0.0104) (0.0289)** (0.0399)***

Diagnostic test
logL 0.580 1246.173 1226.5 1383.281
AIC 1.868380 1.313630 1.360687 1.423601
BIC 1.787666 1.339793 1.387966 1.449269
Q(5) 33.5747*** 2.26164 2.22520 4.27771
Q(20) 58.6204*** 12.7236 19.4454 18.2636
Qs(5) 19.2943*** 3.61580 3.71839 10.9827**
Qs(20) 24.4497 10.7945 11.1145 24.1981
LM(5) 3.7400*** 0.73390 0.72624 2.1072 *
LM(10) 2.2480** 0.64747 0.50359 1.3800
P(40) 78.0618*** 60.6797*** 66.6872*** 83.3865***

Notes: See the notes of Table 3.

order to beat the market. Furthermore, we show that the incorporation of structural changes in the ARFIMAFIGARCH model
reduces significantly the degree of persistence for all SAR rates. This result is consistent with those of previous studies (i.e., Arouri
et al., 2012, for precious metal markets, Ewing & Malik, 2013, for gold and oil futures markets, Kang & Yoon, 2013, for the crude oil

Table 6
Sudden changes in volatility as detected by the ICSS algorithm.

Series Number of change points Time period Standard deviation

SAR/USD 1 14 December 200319 September 2007 0.044

2 21 September 200720 November 2007 0.114
3 21 November 20074 December 2007 0.392
4 5 December 200731 December 2009 0.146
SAR/EUR 1 14 December 20035 November2005 0.561
2 6 November20058 August 2008 0.363
3 9 August 200820 March2009 0.900
4 21 March200931 December 2009 0.481
SAR/GBP 1 14 December 20035 November2005 0.531
2 6 November200531 August 2008 0.363
3 1 September200813February2009 1.152
4 14 February 200931 December 2009 0.587
SAR/JPY 1 14 December 200312 August 2005 0.529
2 13 August 200527July 2007 0.373
3 28 July 200715 September 2008 0.543
4 16 September 200825 March 2009 0.910
5 26 March 200931 December 2009 0.541

Notes: Number of sudden changes and the corresponding time periods were detected by the ICSS algorithm.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 115

Table 7
Estimation results of ARFIMAFIGARCH model with dummy variables for scheduled news announcements and sudden changes in volatility.

SAR/USD SAR/EUR SAR/GBP SAR/JPY

Mean equation
.0002 0.0184 0.00410 0.0060
(0.0004) (0.0079)** (0.0106) (0.0099)
1 538 0.208 0.202 0.206
(0.240)** (0.047)*** (0.050)*** (0.048)***
1 .000
(0.269)***
2. 300
(0.098)***
d .420 0.058 0.057 0.039
(0.135)*** (0.035)* (0.082) (0.036)
Scheduled news 0.0167 0.0009 0.0575 0.0397
(0.0118) (0.0742) (0.0820) (0.0905)

Variance equation
0 . 0015 0.0057 0.0191 0.2836
(0.0011) (0.0066) (0.0140) (0.0623)***
1 0.013 0.473 0.510 0.637
(0.142) (0.115)** (0.074)*** (0.089)***
1 0.497 0.844 0.704 0.759
(0.170)*** (0.075)*** (0.072)*** (0.061)***
552 0.457 0.250 0.251
(0.267)** (0.204)** (0.090)*** (0.091)***
Scheduled news 0.0021 0.0140 0.0479 0.0820
(0.0023) (0.0309) (0.0576) (0.0257)***

Diagnostic test
logL 7.501 1240.814 1216.179 1371.378
AIC 1.948251 1.311161 1.352620 1.415519
BIC 1.845524 1.346046 1.388992 1.452596
Q(5) 28.3415*** 3.39170 2.73582 3.87106
Q(20) 56.5116*** 13.4796 17.6579 17.7119
Qs(5) 17.9688*** 3.22321 2.09942 11.2034**
Qs(20) 26.6902* 10.2150 7.88739 24.7264
LM(5) 3.4888*** 0.63972 0.42484 2.1832*
LM(10) 2.4065*** 0.57913 0.24688 1.4160
P(40) 136.1586*** 63.5264*** 52.7225* 70.3395***

Notes: See the notes of Table 3.

markets, among others). In conclusion, we support the general evidence in the literature of changing persistence of volatility of
international financial markets.
Looking now at the diagnostic tests, we find that the residual tests and the model selection criteria show the fitting ability and
suitability of the ARFIMAFIGARCH model with dummy variables of the unscheduled news announcements and the structural
breaks in volatility for detecting the LM process.
Different reasons may explain the presence of LM for the largest foreign market in the GCC and the Middle East areas. In the
whole, the presence of the LM process in the Saudi foreign exchange market has several significant economic implications. The
behavior of LM in the Saudi exchange rates is in fact explained by standard fundamental factors such as the international trade
flows, prices of tradable goods, prices of foreign exchange futures/options, and international asset portfolios (Cheung, 1993). In
addition, it is also linked to relative price movements, relative stock price movements, intervention of monetary authorities,
relative currency supplies and relative incomes.6
Cheung (1993) and Diebold and Rudebusch (1989) show that these fundamental variables are fractionally integrated. Hence,
the variability in these macroeconomic variables is due to exchange rate dynamics. As a result, the persistence in these variables is
the essential factor for the presence of LM in Saudi foreign exchange market. The obtained result can be also related to the
aggregation (Granger, 1980) and the accumulation of different arrival information process.

5.5. Out-of-sample forecasts

The results of evaluation of the forecasting accuracy of one-step-ahead forecasts generated during recent 6-months from the
ARMAGARCH and ARFIMAFIGARCH model with/without sudden changes in volatility are reported in Tables 9 and 10. We

6
For more information on those fundamental factors, see Al-Khazali et al. (2012), Grossman and Orlov (2012), Kodongo and Ojah (2012), Lai and Fang (2012),
Lee and Chou (2013), Lin (2012) and Verheyen (2012).
116 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

Table 8
Estimation results of ARFIMAFIGARCH model with dummy variables for unscheduled news announcements and sudden changes in volatility.

SAR/USD SAR/EUR SAR/GBP SAR/JPY

Mean equation
.00001 0.0186 0.0032 0.0049
(0.0004) (0.0077)** (0.0108) (0.0098)
1 .984 0.210 0.200 0.207
(0.006)*** (0.047)*** (0.049)*** (0.049)***
1 493
(0.136)***
2 .457
(0.133)***
d .323 0.062 0.026 0.043
(0.123)*** (0.035)* (0.039) (0.038)
Unscheduled news 0.0017 0.0017 0.0310 0.0094
(0.0042) (0.0478) (0.0598) (0.0773)

Variance equation
0 0021 0.0060 0.0127 0.2874
(0.0016) (0.0053) (0.0099) (0.0605)***
1 0.000 0.493 0.549 0.635
(0.136) (0.098)*** (0.090)*** (0.090)***
1 0.606 0.845 0.760 0.742
(0.294)** (0.050)*** (0.064)*** (0.075)***
641 0.440 0.286 0.231
(0.306)** (0.147)*** (0.099)*** (0.092)**
Unscheduled news 0.0008 0.0387 0.0696 0.0290
(0.0000)*** (0.0205)* (0.0306)** (0.0441)

Diagnostic test
logL 4.545 1236.918 1213.153 1373.632
AIC 2.005158 1.307083 1.349288 1.417824
BIC 1.902430 1.341967 1.385661 1.454901
Q(5) 18.1128*** 3.26361 2.5755 3.65086
Q(20) 34.023*** 13.2979 17.9871 17.2385
Qs(5) 5.96207 3.09270 3.29462 12.1126***
Qs(20) 18.1997 10.8547 8.65927 27.0690*
LM(5) 1.1454 0.61136 0.65718 2.3796**
LM(10) 1.4381 0.62787 0.36454 1.5364
P(40) 168.3456*** 59.8006** 50.3877 71.5256***

Notes: See the notes of Table 3.

Table 9
Evaluation of forecasting accuracy for returns.

ARMAGARCH(1,1) ARFIMAFIGARCH(1,1) Scheduled news Unscheduled news Scheduled news Unscheduled news
model model announcements announcements announcements and announcements and
sudden changes sudden changes

SAR/USD
MAE 0.10387 0.10808 [2.06]** 0.10510 [0.46] 0.10680 [1.34]* 0.10340 [0.22] 0.10483 [0.58]
MSE 0.01500 0.01475 [0.39] 0.01383 [1.53]* 0.01428 [1.09] 0.01421 [1.41]* 0.01468 [0.34]
TIC 0.57288 0.56654 [0.56] 0.56163 [0.85] 0.56403 [0.75] 0.56482 [0.69] 0.57189 [0.10]
SAR/EUR
MAE 0.32772 0.32608 [0.76] 0.32621 [0.68] 0.32612 [0.70] 0.32606 [0.72] 0.32602 [0.73]
MSE 0.15922 0.15833 [0.73] 0.15851 [0.57] 0.15840 [0.64] 0.15839 [0.64] 0.15831 [0.68]
TIC 0.81339 0.81935 [0.71] 0.82018 [0.81] 0.82077 [0.85] 0.82402 [1.21] 0.82474 [1.24]
SAR/GBP
MAE 0.42714 0.42673 [0.45] 0.42668 [0.56] 0.42744 [0.20] 0.42665 [0.70] 0.42721 [0.06]
MSE 0.27097 0.26988 [1.06] 0.26996 [1.12] 0.27056 [0.26] 0.27024 [1.02] 0.27076 [0.18]
TIC 0.80716 0.80695 [0.06] 0.80733 [0.05] 0.80861 [0.27] 0.80816 [0.39] 0.80988 [0.69]
SAR/JPY
MAE 0.40681 0.40568 [0.85] 0.40534 [1.07] 0.40668 [0.11] 0.40614 [0.59] 0.40641 [0.33]
MSE 0.25908 0.25843 [0.51] 0.25838 [0.53] 0.25948 [0.33] 0.25901 [0.06] 0.25945 [0.32]
TIC 0.81626 0.81165 [0.80] 0.81061 [0.94] 0.81966 [0.69] 0.81667 [0.08] 0.81804 [0.35]

Notes: The values in bold refer to the lowest for forecasting error statistics. The numbers in brackets are the DM test statistic to evaluate the null hypothesis that
the forecasting accuracy of the ARMAGARCH(1,1) model is the same as the ARFIMAFIGARCH model with or without dummies. *** (**, *) indicates that the null
hypothesis of the DM test is rejected at the 1% (5%, 10%) significance level.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 117

Table 10
Evaluation of forecasting accuracy for volatility.

ARMAGARCH(1,1) ARFIMAFIGARCH(1,1) Scheduled news Unscheduled news Scheduled news Unscheduled news
model model announcements announcements announcements and announcements and
sudden changes sudden changes

SAR/USD
MAE 0.02392 0.02417 [0.57] 0.02393 [0.02] 0.02422 [0.72] 0.02460 [1.63]* 0.02490 [2.44]***
MSE 0.00081 0.00091 [4.30]*** 0.00092 [4.02]*** 0.00091 [4.35]*** 0.00090 [4.05]*** 0.00091 [4.33]***
TIC 0.56716 0.58944 [2.78]*** 0.58609 [2.02]** 0.59123 [3.10]*** 0.59725 [3.71]*** 0.60233 [4.43]***
LL 2.96686 2.77515 [2.24]*** 2.69870 [2.59]*** 2.78471 [2.19]** 2.87042 [1.35]* 2.93425 [0.50]
SAR/EUR
MAE 0.17530 0.15561 [0.25] 0.17534 [0.03] 0.18687 [4.60]*** 0.14634 [4.95]*** 0.14639 [4.70]***
MSE 0.05231 0.05165 [1.18] 0.05166 [1.13] 0.05679 [3.29]*** 0.04642 [1.75]** 0.04630 [1.77]**
TIC 0.51544 0.51695 [0.70] 0.51715 [0.71] 0.52260 [1.79]** 0.49339 [2.09]** 0.49390 [1.79]**
LL 4.01468 4.02037 [0.32] 4.00755 [0.37] 4.25183 [6.46]*** 3.23384 [6.34]*** 3.22174 [6.05]***
SAR/GBP
MAE 0.26323 0.26838 [2.28]** 0.26787 [1.96]** 0.28565 [4.38]*** 0.25722 [3.57]*** 0.25958 [0.82]
MSE 0.14841 0.14633 [0.72] 0.14567 [0.90] 0.15355 [1.09] 0.14703 [0.99] 0.14720 [0.32]
TIC 0.49500 0.49722 [1.01] 0.49673 [0.73] 0.50603 [2.17]** 0.49150 [1.90]** 0.49243 [0.48]
LL 3.78865 3.89128 [4.16]*** 3.89105 [3.84]*** 4.16767 [5.93]*** 3.72145 [4.07]*** 3.75401 [0.64]
SAR/JPY
MAE 0.27967 0.28579 [1.99]** 0.28619 [2.16]** 0.30536 [4.28]*** 0.28409 [1.47]* 0.27505 [1.03]
MSE 0.17486 0.17278 [0.65] 0.17331 [0.49] 0.17950 [0.74] 0.17200 [0.92] 0.17483 [0.01]
TIC 0.53164 0.53745 [1.83]** 0.53681 [1.61]* 0.54885 [2.87]** 0.53858 [2.10]** 0.53160 [0.01]
LL 4.79569 4.86461 [1.51]* 4.87659 [1.81]** 5.11239 [4.01]*** 4.82607 [0.67] 4.59028 [2.89]***

Notes: See the notes of Table 9.

should note that a smaller forecasting error statistic reflects the superior forecasting ability of a given model. Looking at Table 9,
the overall evaluation indicates that the model with both the unscheduled (the scheduled) news announcements and dummies of
sudden changes provides relatively better forecasts of the SAR/EUR (SAR/USD and SAR/JPY) exchange rate returns, whereas the
model without the news announcements and sudden changes dummies seems to be fit alternative for SAR/GBP rates. The DM test
statistics do not reject the null hypothesis that the forecasting accuracy of the ARMAGARCH(1,1) model is the same as that of the
ARFIMAFIGARCH model with or without the dummies.
Table 10 presents the evaluation of forecasting accuracy for conditional volatility. It shows that the ARMAGARCH(1,1) model
is superior to forecasting the volatility for the SAR/USD rates, while the scheduled and unscheduled news and structural changes
models are suitable for forecasting the SAR/GBP and SARJPY, respectively. Finally, for the SAR/EUR exchange rates, the model with
news announcements and sudden changes dummies seem to be fit model better. Generally speaking, the DM test rejects the null
hypothesis of no difference in forecast accuracy between the conditional models of the SAR exchange rates. This result is also
supported by the evaluation of the residual test and model selection criteria.
All in all, the four forecast error statistics have no support for any one specific model, and thus no ideal volatility models under
consideration is appropriate for analyzing the LM dynamics in the Saudi Arabia foreign exchange market. Thus, the investors
should be vigilant when examining the risk in such market.

6. Conclusions

The existing literature contains conflicting evidence regarding the long memory process and the accuracy of volatility forecasts
in financial time series. We address in this research the question of whether the exchange rates with respect to major currencies
of the world's largest oil exporter follow a dual LM process and discern the impacts of news announcements and structural breaks
on conditional variance. Finally, we apply both the out-of-sample forecast analysis and the equal accuracy test of Diebold and
Mariano (1995) to assess the forecast performance models for the daily conditional mean and conditional volatility of the SAR's
bilateral exchange rates.
Evidence of a dual LM in the SAR markets is limited to both the SAR/USD and SAR/EUR rates. We find strong evidence for LM in
variance for all rates. The results also reveal that the scheduled news announcements have no significant impacts to reduce or
increase the persistence of conditional variance for these markets, while the unscheduled news announcements contribute to
increase the variance in most the SAR markets. We also detect at least four structural breaks for these foreign exchange rates over
the period 20032009. The incorporation of these variables appears strongly significant on the ARFIMAFIGARCH model. More
specifically, the level of volatility persistence decreases significantly when adjusting for the structural breaks for all SAR exchange
rates. These empirical findings underpin the relevance of incorporating heterogeneous (unexpected) news events and structural
breaks to explain the different volatility patterns of exchange rates. As a result, they provide evidence of the predictive effect of
macroeconomic releases on LM in the exchange rate markets of this oil-rich exporter. These results refute the efficient market
hypothesis in its weak form. Our conclusions have several implications for financial risk managers and policy-makers. We also
118 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119

show that the estimates values of DM statistics for out-of-sample performance are mixed and that none of the volatility models is
appropriate for analyzing the LM dynamics in the Saudi Arabia foreign exchange market.
It will be intriguing in the future to extend this work by three main avenues. (1) The joint ARFIMAFIGARCH model should be
improved in order to consider the asymmetric effects of financial time-series. (2) It would be interesting to consider the
assumption of non-normality for capturing the asymmetry and fait tails of residual distributions (see the significance of the
Pearson goodness-of-fit statistic) based on the Student's t-distribution and the Skewed Student's t-distribution. (3) To enhance
the accuracy of the forecasting ability of models, it would be worthwhile to apply the asymmetric loss functions to penalize
under/over prediction.

Acknowledgment

The authors wish to thank Editor Hamid Beladi and an anonymous reviewer for their helpful comments during the reviewing
process.

References

Aggarwal, R., Inclan, C., & Leal, R. (1999). Volatility in emerging stock markets. Journal of Financial and Quantitative Analysis, 34, 3355.
Al-Khazali, O. M., Pyun, C. S., & Kim, D. (2012). Are exchange rate movements predictable in Asia-Pacific markets? Evidence of random walk and martingale
difference processes. International Review of Economics and Finance, 21, 221231.
Aloy, M., Boutahar, M., Gente, K., & Pguin-Feissolle, A. (2012). Purchasing power parity and the long memory properties of real exchange rates: Does one size fit
all? Economic Modelling, 28, 12791290.
Arouri, M. H., Hammoudeh, S., Lahiani, A., & Nguyen, D. K. (2012). Long memory and structural breaks in modeling the return and volatility dynamics of precious
metals. The Quarterly Review of Economics and Finance, 52, 207218.
Baillie, R. T., Bollerslev, T., & Mikkelsen, H. O. (1996). Fractionally integrated generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 74,
330.
Baillie, R. T., Cecen, A. A., Erkal, C., & Han, Y. (2004). Measuring non-linearity, long memory and self-similarity in high-frequency European exchange rates. Journal
of International Financial Markets Institutions and Money, 14, 401418.
Bekiros, S., & Marcellino, M. (2013). The multiscale causal dynamics of foreign exchange markets. Journal of International Money and Finance, 33, 282305.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31, 307327.
Brailsford, T. J., & Faff, R. W. (1996). An evaluation of volatility forecasting techniques. Journal of Banking and Finance, 20, 419438.
Brooks, C., & Persand, G. (2003). Volatility forecasting for risk management. Journal of Forecasting, 22, 122.
Cairns, R. D., & Calfucura, E. (2012). OPEC: Market failure or power failure? Energy Policy, 50, 570580.
Cajueiro, D., & Tabak, B. (2008). Testing for long-range dependence in world stock markets. Chaos, Solutions and Fractals, 37, 918927.
Caporale, G. M., & Gil-Alana, L. A. (2013). Long memory and fractional integration in high frequency data on the US dollar/British pound spot exchange rate.
International Review of Financial Analysis, 29, 19.
Cheung, Y. W. (1993). Long memory in foreign exchange rates. Journal of Business and Economic Statistics ll, 93101.
Chkili, W., Aloui, C., & Nguyen, D. K. (2012). Asymmetric effects and long memory in dynamic volatility relationships between stock returns and exchange rates.
Journal of International Financial Markets, Institutions & Money, 22, 738757.
Chortareas, G., Jiang, Y., & Nankervis, J. C. (2011). Forecasting exchange rate volatility using high-frequency data: Is the euro different? International Journal of
Forecasting, 27, 10891107.
Degiannakis, D. (2004). Volatility forecasting: evidence from a fractional integrated asymmetric power ARCH skewed-t model. Applied Financial Economics, 14,
13331342.
Diebold, F. X., & Mariano, R. S. (1995). Comparing predictive accuracy. Journal of Business & Economic Statistics, 13, 253263.
Diebold, F. X., & Rudebusch, G. D. (1989). Long memory and persistence in aggregate output. Journal of Monetary Economics, 24, 189209.
Dornbusch, R. (1980). Exchange rate economics: Where do we stand? Brookings Papers on Economic Activity, 10, 143185.
Dufrnot, G., Lardic, S., Mathieu, L., Mignon, V., & Pguin-Feissolle, A. (2008). Explaining the European exchange rates deviations: Long memory or non-linear
adjustment? Journal of International Financial Markets, Institutions & Money, 18, 207215.
Engle, R. F. (1982). Autoregressive conditional heteroskedasticity with estimates of the variance of United Kingdom inflation. Econometrica, 50, 9871007.
Ewing, B. T., & Malik, F. (2013). Volatility transmission between gold and oil futures under structural breaks. International Review of Economics and Finance, 25,
113121.
Frenkel, J. A. (1981). Flexible exchange rates, prices and the role of news: Lessons from the 1970's. Journal of Political Economy, 89, 665705.
Geweke, J., & Porter-Hudak, S. (1983). The estimation and application of long memory time series models. Journal of Time Series Analysis, 4, 221238.
Gil-Alana, L. A. (2006). Seasonal and non-seasonal long memory effects in the Japanese real effective exchange rate. Journal of the Japanese and International
Economies, 20, 8798.
Granger, C. W. J. (1980). Long memory relationships and the aggregation of dynamic models. Journal of Econometrics, 14, 227238.
Grossman, A., & Orlov, A. G. (2012). Exchange rate misalignments in frequency domain. International Review of Economics and Finance, 24, 185199.
Hammoudeh, S., & Li, H. (2008). Sudden changes in volatility in emerging markets: The case of Gulf Arab stock markets. International Review of Financial Analysis,
17, 4763.
Hurst, E. (1951). Long term storage capacity of reservoirs. Transactions on American Society of Civil Engineering, 116, 770799.
Hwang, Y. (2001). Asymmetric long memory GARCH in exchange return. Economics Letters, 73, 15.
Incln, C., & Tiao, G. C. (1994). Use of cumulative sums of squares for retrospective detection of changes of variance. Journal of the American Statistical Association,
89, 913923.
Jiang, G. J., Konstantinidi, E., & Skiadopoulos, G. (2012). Volatility spillovers and the effect of news announcements. Journal of Banking and Finance, 36, 22602273.
Kang, S. H., Cheong, C., & Yoon, S. -M. (2011). Structural changes and volatility transmission in crude oil markets. Physica A, 390, 43174324.
Kang, S. H., & Yoon, S. -M. (2013). Modeling and forecasting the volatility of petroleum futures prices. Energy Economics, 36, 354362.
Kodongo, O., & Ojah, K. (2012). The dynamic relation between foreign exchange rates and international portfolio flows: Evidence from Africa's capital markets.
International Review of Economics and Finance, 24, 7187.
Kumar, D., & Maheswaran, S. (2013). Detecting sudden changes in volatility estimated from high, low and closing prices. Economic Modelling, 31, 484491.
Laakkonen, H. (2007). The impact of macroeconomic news on exchange rate volatility. Finnish Economic Papers, 20, 2340.
Lai, C., & Fang, C. (2012). Is the honeymoon effect valid in the presence of both exchange rate and output expectations? A graphical analysis. International Review
of Economics and Finance, 21, 140146.
Lastrapes, W. D. (1989). Exchange rate volatility and U.S. monetary policy: An ARCH application. Journal of Money Credit and Banking, 21, 6677.
Lee, C., & Chou, P. (2013). The behavior of real exchange rate: Nonlinearity and breaks. International Review of Economics and Finance, 27, 125133.
Lin, C. (2012). The comovement between exchange rates and stock prices in the Asian emerging markets. International Review of Economics and Finance, 22,
161172.
 W. Mensi et al. / International Review of Economics and Finance 30 (2014) 101119 119

Lo, A. W. (1991). Long-term memory in stock market prices. Econometrica, 5, 12791313.

Marshall, A., Musayev, T., Pinto, H., & Tang, L. (2012). Impact of news announcements on the foreign exchange implied volatility. Journal of International Financial
Markets, Institutions & Money, 22, 719737.
Nikkinen, J., & Vhmaa, S. (2009). Central bank interventions and implied exchange rate correlations. Journal of Empirical Finance, 16, 862873.
Omrane, W., & Heinen, A. (2009). Is there any common knowledge news in the Euro/Dollar market? International Review of Economics and Finance, 18, 656670.
Robinson, P. M., & Henry, M. (1999). Long and short memory conditional heteroscedasticity in estimating the memory parameter of levels. Econometric Theory, 15,
299336.
Verheyen, F. (2012). Bilateral exports from euro zone countries to the US Does exchange rate variability play a role? International Review of Economics and
Finance, 24, 97108.
Vivian, A., & Wohar, M. E. (2012). Commodity volatility breaks. Journal of International Financial Markets, Institutions & Money, 22, 395422.