← Back

Synthesis, structure, electrochemistry and cytotoxicity studies of Ru(II) and Pt(II)–N-heterocyclic carbene complexes of CNC-pincer ligand

Copernican Journal of Finance & Accounting e-ISSN 2300-3065 p-ISSN 2300-1240 2015, volume 4, issue 2 Al-Hajieh H., AlNemer H., Rodgers T., & Niklewski J. (2015). Forecasting the Jordanian stock index: modelling asymmetric volatility and distribution effects within a GARCH framework. Copernican Journal of Finance & Accounting, 4(2), 9–26. http://dx.doi.org/10.12775/CJFA.2015.013 Heitham Al-Hajieh* Department of Finance, King Abdulaziz University, Saudi Arabia Hashem AlNemer** Department of Finance and Insurance, University of Jeddah, Saudi Arabia Timothy Rodgers*** School of Economics, Finance and Accounting, Coventry University, UK Jacek Niklewski**** School of Economics, Finance and Accounting, Coventry University, UK forecasting the jordanian stock index: modelling asymmetric volatility and distribution effects within a garch framework Keywords: GARCH, asymmetry, distributions. J E L Classification: C01, C58, G15. Date of submission: May 16, 2015; date of acceptance: October 26, 2015. Contact information: Haawadh@kau.edu.sa, Department of Finance, King Abdulaziz University, Abdullah Sulayman, Jeddah 21589, Saudi Arabia, phone: +966 2 695 2000. ** Contact information: Halnemer@kau.edu.sa, Department of Finance and Insurance, University of Jeddah, Saudi Arabia. *** Contact information: T.Rodgers@coventry.ac.uk, School of Economics, Finance and Accounting, Coventry University, UK. **** Contact information: J.Niklewski@coventry.ac.uk, School of Economics, Finance and Accounting, Coventry University, UK. * 10 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski Abstract: The modelling of market returns can be especially problematical in emerging and frontier financial markets given the propensity of their returns to exhibit significant non-normality and volatility asymmetries. This paper attempts to identify which representations within the GARCH family of models can most efficiently deal with these issues. A number of different distributions (normal, Student t, GED and skewed Student) and different volatility of returns asymmetry representations (EGARCH and GJR-GARCH) are examined. Our data set consists of daily Jordanian stock market returns over the period January 2000 – November 2014. Using both the Superior Predicative Ability (SPA) and Model Confidence Set (MCS) testing frameworks it is found that using GJR-GARCH with a skewed Student distribution most accurately and efficiently forecasts Jordanian market movements. Our findings are consistent with similar research undertaken in respect to developed markets. Introduction The global financial crisis of 2007-09 and subsequent shocks in the Euro-area and beyond has led researchers to examine again the ways in which they model stock market returns. It has become increasingly apparent that the ‘standard assumptions’ made in respect to the ways in which statistical series are distributed are not applicable in financial markets. As the global economy becomes more integrated it is also becoming increasingly important to understand how emerging and frontier markets react in periods of high volatility. In this paper we explore which elements of the GARCH family of models can be used to efficiently and effectively model markets in Jordan from January 2000 to November 2014. The paper begins with a brief review of the literature in the second Section. This is followed in the subsequent section by a description of the data and methodology. The most efficient model is then identified using the Superior Predictive Ability (SPA) and Model Confidence Set (MCS) prediction frameworks before, finally, some brief conclusions are drawn. Literature: volatility modelling in the GARCH framework The GARCH model was first introduced by Bollerslev (1986). Much of the subsequent research in this area has focused on developing the model to better reflect the data found in real-world settings, such as, for example, financial markets. The finance-related literature focuses principally on modelling (i) the structure of the volatility (ii) the nature of the distribution of the returns. Forecasting the Jordanian stock index… 11 Much of the work on modelling the structure of volatility relates to the asymmetries found in stock market returns. For example, Engle and Ng (1993) found evidence supporting the Quadratic-GARCH model. Others, such as Brailsford and Faff (1996), found evidence to support GJR-GARCH and Heynen and Kat (1994) argued that EGARCH has a superior predictive ability. Although the literature does not show one individual asymmetry specification as being clearly superior to others, Awartani and Corradi (2005) argue that they generally outperform non-asymmetric specifications in financial market prediction. However, it can be noted that the evidence is not unequivocal; McMillan, Speight, and Apgwilym (2000) found GARCH, moving average and exponential smoothing models to provide marginally superior daily volatility forecasts. Their work also strongly suggested that EGARCH does not necessarily outperform simple GARCH model in forecasting market volatility. For the purposes of our paper methodologies with the greatest out-of-sample forecasting accuracy are the most desirable. Balaban (2004) tested a series of both symmetric and asymmetric models (included ARCH, GARCH, GJRGARCH and EGARCH). Their results suggest that all models are biased and generally over-predict volatility. Model performance in these latter respects was best for GARCH and worst for GJR-GARCH. However, they also noted that if avoidance of under-prediction was the key decision criteria, ARCH was the preferred model. A further issue that is important to consider is that the nature of volatility in emerging and frontier markets differs considerably from that found in developed markets (Andrikopoulos, Niklewski and Rodgers forthcoming). Given that the focus of this paper is Jordan, it is important to consider how different GARCH specifications perform in these market-types. Gokcan (2000) examined seven emerging market (Argentina, Brazil, Colombia, Malaysia, Mexico, Philippines, Taiwan) and found that GARCH(1,1) outperformed EGARCH everywhere with the exception of Brazil. The most compelling conclusion we draw from the literature in respect to modelling volatility structure is that it is difficult to identify one single model that is clearly superior to others. This indicates to us that it may be necessary to test a number of volatility specifications. A standard feature of most financial markets is that their returns are nonnormal with distributions exhibiting ‘fat-tailed’ characteristics (Mittnik, Paolella, and Rachev 2000). This appears to be particularly an issue in emerging markets. Brooks (2007) studied a set of such markets (including MENA re- 12 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski gion countries) using the Asymmetric Power ARCH model. He found that unlike developed markets, where non-normal conditional error distributions appear to fit the data well, there were a set of emerging markets where estimation problems arise using a conditional t distribution. It was also found that the degree of volatility asymmetry appears to vary across markets, with the Middle Eastern and African markets having very different volatility asymmetry characteristics to Latin American markets. Brooks (2007) found that a fat-tailed t-distribution was needed to model the distribution of returns in most MENA markets. However, there were differences. For example, Turkey, Egypt and Morocco display much larger kurtosis and exhibit fatter tails than Jordan. Likelihood ratio tests were found to clearly favour the APARCH with t-distribution rather than a normal distribution. It is possible that such differences may reflect the Islamic nature of these markets. For example, Al-Hajieh, Redhead, and Rodgers (2011) found that the month of Ramadan (Islamic holy month) shows high level of volatility and the overall impact of Ramadan on returns is statistically significant in most Middle East countries. We conclude the literature review by identifying that we are aware of no studies of volatility forecasting in emerging markets that have examined the combined issues of the distribution of returns and the GARCH model specification. We have also identified from the literature that (i) no single GARCH model specifications clearly outperforms other forms in all circumstances and (ii) evidence to suggest that the distribution of returns in emerging markets (like Jordan) can follow a number of possible forms and that these can interact with volatility models in different ways. In this paper we therefore test a number of possible distribution-types (Normal, Student-t, GED, and Skewed Student) and GARCH model types (GARCH, GJR-GARCH and EGARCH) in order to identify the most efficient volatility forecasting model for Jordan. Data description The empirical investigation is undertaken in respect to daily closing price data for the Jordanian Amman Stock Exchange Index (ASE) covering the period 2nd January 2000 to 27th November 2014. The data source is the Thomson-Reuters Eikon database and the dataset comprises of a total of 3655 trading days. Daily returns computed as the log-difference of the daily closing prices: Amman Stock Exchange Index (ASE) covering the period 2nd January 2000 to mber 2014. The data source is the Thomson-Reuters Eikon database and the mprises of a total of 3655 trading days. Daily returns computed as the logForecasting the Jordanian stock index… of the daily closing prices: Rt  ln Pt  ln Pt 1 13 [1] [1] ASE Index closing prices are presented in Figure 1 and the daily returns in dex closing prices are presented in Figure 1 and the daily returns in Figure 2. A Figure 2. A number of volatility clusters can be observed in the returns data; for volatility clusters can be observed in the returns data; for example, a cluster example, a cluster corresponding to the 2007-09 global financial crisis. ing to the 2007-09 global financial crisis. Figure 1. Daily Jordanian Price Index January 2000–November 2014 Figure 1. Daily Jordanian Price Index January 2000–November 2014 S o u r c e : created by the authors using OxMetricsTM 7 software and data from Thomson Reuters EikonTM. A preliminary statistical analysis of the daily returns is presented in Table 1. It can be noted that average daily returns are small relative to the standard deviation. The series also displays negative skewness and strong positive kurtosis; these are indicative of a heavy tailed non-Gaussian distribution. Table 1. Descriptive Statistics of Daily Returns January 2000–November 2014 Mean Std. Dev Min Max Skewness Excess Kurtosis J-B Test ARCH Test1 L-B Test1 ADF Test 0.025 0.946 -6.428 6.198 -0.363 6.167 5872.2** 107.9** 2571** -32.8** ** Significant at 1%. Both the Ljung-Box and the ARCH tests use 10 lags. 1 S o u r c e : estimated by the authors using OxMetricsTM 7. 14 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski Figure 2. Jordanian Daily Percentage Returns January 2000–November 2014 S o u r c e : created by the authors using OxMetricsTM 7. This is confirmed by the Jarque-Bera test which rejects unconditional normality and further confirmatory evidence is provided in the related histogram (Figure 3). Figure 3. Histogram of Jordanian Daily Returns January 2000–November 2014 S o u r c e : created by the authors using OxMetricsTM 7. Forecasting the Jordanian stock index… 15 The Ljung-Box Q-test and the ARCH tests suggest autocorrelation and hetroskedasticity within the data and the ADF (Augmented Dickey-Fuller) unit root test rejects the null hypothesis of data non-stationary. The research methodology A total of 12 GARCH-model-specification/distribution pairs are tested with the results being presented in Tables 3–7. The models tested are: GARCHbased specifications (GARCH-N, GARCH-T, GARCH-GED and GARCH-ST)1; and two sets of asymmetry-type specifications: (i) EGARCH (EGARCH-N, EGARCH-T, EGARCH-GED and EGARCH-ST) and (ii) GJR-GARCH (GJR-GARCH-N, GJR- GARCH- T, GJR- GARCH-GED and GJR- GARCH-ST). The alternative models are subsequently evaluated by: (i) an evaluation of model parameters and (ii) an evaluation of model forecasting performance. For robustness, the latter undertakes a series of tests using both the Superior Predictive Ability (SPA) test Hansen (2005) and the Model Confidence Set (MCS) test (Hansen, Lunde, and Nason 2011). Both are available in the OxMetricsTM 72 software package used in this paper. The forecast-based tests use a ‘loss-function’ to identify the most efficient model. The loss function can be estimated using Mean Squared Error (MSE) and Mean Absolute Deviation (MAD) statistics. SPA identifies the ‘best’ model in terms of predictive ability and MCS identifies the ‘best’ model set. The model specifications and the distributions tested are identified below. They consist of three different conditional volatility specifications and four different statistical distributions. (i) GARCH The GARCH model, as introduced by Bollerslev (1986), is a generalisation of the ARCH specification of Engle (1982). The model specifies that the conditional variance is a function of the lagged squared residuals as well as of its past conditional variances. Although the equation may be specified with a number 1 N stands for the Normal distribution, T the Student t distribution, GED the Generalised Error Distribution and ST the Standardised Skewed Student distribution. 2 The MULCOM 3.0 package running SPA and MCS was developed by Hansen and Lunde (2014). ons. (i) GARCH GARCH ) GARCH GARCH model, as introduced (1986),ofisthe a generalisation of the The GARCH model, The as introduced by Bollerslev (1986),byis Bollerslev a generalisation The GARCH model, as introduced by Bollerslev (1986), is a generalisation of the of Engle (1982).that Thethe model specifies that theisconditional variance is a CH specification Engle specification (1982). The model specifies conditional variance a 16ofARCH H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski RCH specificationfunction of Engleof(1982). The model specifies that the conditional variance is a the lagged squared well as ofvariances. its past conditional variances. Altction of the lagged squared residuals as well as ofresiduals its past as conditional Altnction of the of lagged squared residuals as well asinofeach its past variances. Although the equation belag specified aconditional number lags in a single lag in lags in each term a may single is adequate in financial gh the equation may be specified with a number of lagswith in usually each term aofsingle lageach in term market ough the equation may specified with ainnumber ofmarket lags indata. eachWe term a single conversion lag in data. Weisbe follow this conversion in this paper. each usually adequate follow in this paper. h is usually adequate in financial market data. financial We follow this conversion in this this paper. ach is usually adequate in financial market data. We follow this conversion in this paper. q 2 p 2 p t t  i2 j  1 j i t i q p 2 2 2  u  i[ 2t ]  i    j t  j       i u     t  j2    i 1  u    ji 1t  j [ 2 ] j 1 2 t 2 t q i 1 [2] j 1 [2] (ii) EGARCH EGARCH (ii) EGARCH i) EGARCH It isvolatilities often observed that volatilities associated with downward movements in financial t is often observed that associated with downward movements in financial It is often observed that volatilities associated with downward movements in financial It is often observed that volatilities associated with downward movements markets are greaterobserved than theby volatilities observed byofupward movements of the same magkets are greater than the volatilities upward movements the same magin financial markets are greater than the volatilities observed by upward movearkets are greater than the volatilities observed by upward movements of the same magnitude. Inthesuch circumstances imposed on thestructure conditional variance structure de. In such circumstances symmetry imposedthe onsymmetry the conditional variance ments of the same magnitude. In such circumstances the symmetry imposed tude. In such circumstances the symmetry imposed on the conditional variance structure inmay the not GARCH model mayTonot be appropriate. To address this proissue, Nelson (1991) prohe GARCH model be appropriate. address this issue, Nelson (1991) on the conditional variance structure in the GARCH model may not be approthe GARCH model may not be appropriate. To address this issue, Nelson (1991) pro- for the conditional poses exponential GARCH (EGARCH) specification es the exponential GARCH (EGARCH) model.Nelson The specification for The the the conditional priate. Tothe address this issue, (1991)model. proposes exponential GARCH oses the exponential GARCH (EGARCH) model. The specification for the conditional variance is: (EGARCH) model. The specification for the conditional variance is: ance is: ariance is: �� [� � [� � ����] � ����]������ � [3] ��� �� ���[��������]��� ��� ��� � � � [� � ����] ��� � [3] � �� ��� �� � � � [� � ����] [� � ����]������ � [3] [3] � |�� �|� | � �|� �� [|� where �� �|]� � �� ������ ���� � |] re ���� � � � � �� ���� � ������� ��� �� � �� � [|� � ��������� ������ ���� ������ |� � � �� [|� �|� here ���� � � � �� ������ ���� ��������� ������ ���� ������ � �� � |] ���� ������ ��������� ������ The log specification implies that the asymmetric effect is exponential, rathThe log specification implies effect that theis asymmetric exponential, rather than The log specification that and the that asymmetric exponential, effect ratheristhan er thanimplies quadratic, forecasts of the conditional variance that are generThe log specification implies that the asymmetric effect is exponential, rather thangenerated are nonquadratic, and that forecasts of the conditional variance that are dratic, and that forecasts of the conditional varianceofthat are generated areis nonated are non-negative. The presence asymmetry effects tested in terms of uadratic, and the thatnegative. forecasts of presence the conditional variance that aretested generated areof nonThe of asymmetry effects is in terms the sign and magnitude sign and magnitude effects identified above. ative. The presence of asymmetry effects is tested in terms of the sign and magnitude egative. The presence of asymmetry effects is tested in terms of the sign and magnitude effects identified above. cts identified above. (iii) GJR-GARCH fects identified above. GJR-GARCH (iii) GJR-GARCH ii) GJR-GARCH This variation the GARCH model was proposed by Glosten, Jagannathan, This variation on variation the on GARCH model was proposed by proposed Glosten, Jagannathan, and This on the GARCH model by Glosten, Jagannathan, This variationononthe theGARCH GARCH model waswas proposed byGlosten, Glosten, Jagannathan, andand This variation model was proposed by Jagannathan, and Runkle (1993) as an alternative way of dealing with asymmetric shocks in and Runkle (1993)Runkle as an alternative way of dealing with asymmetric shocks in financial series. (1993) asalternative an alternative way of dealing with asymmetric shocks in financial series. Runkle (1993) as an alternative way of dealing with asymmetric shocks financial series. Runkle (1993) asIts angeneralized way of dealing asymmetric shocks inin financial series. financial series. version is: with Its generalizedItsversion is: generalized version generalized version ItsIts generalized version is:is: is: q q p p 2q 2p 2 2   t2   2   i S(2t2i t22 i )  [4] 2 2 (2i  2  t j  t  i j)  S    (     S  t  i t i i t i t i     t it   (     S  )   jjt2 t2jjj t [ 4j []4 ][ 4 ]  i it  it  i i it  it j it 1 it  i ) 1 q i 1 i  1i  1 p j 1 j  1j  1 [4] � variable that take the value 1 when ���� < 0 , and 0 when a dummy Where ��� is Where a dummy variable the value 1 when ���� , 0and 0 when Where������is ais dummy variable thatthat taketake thevalue value when ����< < ,, 0 and when is�is�aa dummy variable that take the value when and 00when Where Where dummy variable that take the 1 11when ���� 0 ,0<and ���� ≥when 0. A feature of A the GJR model isGJR that the null hypothesis ofhypothesis no leverage (asymfeature of the model isthat the null ofleverage no lever���� ≥ A feature of the GJR model isthat that the null hypothesis no leverage (asym����≥ ≥ Afeature feature theGJR GJR model thenull null hypothesis no leverage (asym���� 0.0. A 0. ofofthe model is isthat the hypothesis ofofnoof (asymmetry) effect metry) is simple to test. Indeed, = γq =γ10 =implies that impact of the a shock is of a shock is 1 = … 0the implies that impact effect is simple toγIndeed, test. Indeed, = 0=implies that the impact a shock metry) effect simple test. Indeed, …… =… γqγ==q = 0γqimplies that the impact ofof a shock is is metry) effect is is simple toto test. γ1γ=1 = symmetric, i.e., past positive shocks have the same impact on today’s volatility as past symmetric, i.e., positive shocks have the same impact on today’s volatility as symmetric, i.e.,past pastpast positive shocks have same impact today’s volatility pastpast symmetric, i.e., positive shocks have thethesame impact onontoday’s volatility asaspast negative shocks. negative shocks. negative shocks. negative shocks. This variation on the GARCH model was proposed by Glosten, Jagannathan, and This variation the GARCH was proposed by Glosten, Jagannathan, This variation on theon GARCH model model was proposed by Glosten, Jagannathan, and and unkle (1993) as an alternative way of dealing with asymmetric shocks in financial series. as an alternative of dealing with asymmetric in financial RunkleRunkle (1993) (1993) as an alternative way ofway dealing with asymmetric shocks shocks in financial series. series. the Jordanian stock index… s generalized versionForecasting is: Its generalized is: Its generalized versionversion is: q 17 p 2 2 p Indeed, γ = … = γ = 0 implies that the to q( i  t q i  effect p test.  t2  age  (asymmetry)  i S2t iist2simple [ 4 ]2 1 q  i)  j t  j 2 2 2 2 2  2   ti   (a shock (     S     [ 4 ] 1of j i.e., t 1 i ) is symmetric, past positive shocks have the same impact  t impact     S  )   [ 4 ]   i t i i t i j t j     j t j i t i i t i t i i 1 j 1 i 1 j 1 on today’s volatility as past negative shocks. This variable variation that on the GARCH model1was proposed and Where ��� is a dummy take the value when ���� <by0 ,Glosten, and 0 Jagannathan, when Four statistical functions are tested. use the Gauss� is adifferent dummy variable thatdensity take the 1asymmetric < The 0 0in , and Where Where ��� is a�� dummy variable that take the value 1value when �when < �0��� ,shocks and when0 when ��� Runkle (1993) as an alternative way of dealing with financial series. of normal the GJR distribution model is thatisthe null hypothesis of no leverage (asym�� ≥ 0. A feature ian or potentially appropriate as the conditional distri0. A Its feature theversion GJR is: model null hypothesis of no leverage ���� ≥ �0. ofgeneralized the of GJR model is that isthethat nullthehypothesis of no leverage (asym-(asym���A≥feature butiontooftest. theIndeed, residuals forthat some inisreturns. The … = γqaccount = 0 implies thenon-normality impact of a shock etry) effect is simple γ1 = would = … = γ = 0 implies that the impact of a is shockeffects; is is simple to test. Indeed, γ 1 q q p = … = γ = 0 implies that the impact of a shock metry) metry) effect iseffect simple to test. Indeed, γ 1 q and third density functions can more fully account for ‘fat-tail’ mmetric, i.e., second past positive shocks have the same impact 2 2 2 on today’s 2volatility as past      (     S  )    [ 4 ]   t i t i i t i t i j t j     symmetric, i.e.,positive past positive shocks have the impact same on today’s volatility past they do have a drawback in thatimpact they are symmetric. symmetric, i.e.,however, past shocks have the same on volatility as Our pastaspreferred i 1 j today’s 1 gative shocks. option negative shocks.is therefore the skewed-Student density proposed by Fernández and negative shocks. they do have a drawback are symmetric. therefore is athey dummy variable that Our take preferred the value option 1 whenis���� < 0 , and 0 when Wherein���that Steel (1998). a drawback thatare they are symmetric. Our preferred is therefore they dothey havedoa have drawback in that in they symmetric. Our preferred option option is therefore A featurebyofFernández the GJR model is that(1998). the null hypothesis of no leverage (asym���� ≥ 0. e skewed-Student density proposed and Steel the skewed-Student density proposed byIndeed, Fernández andγSteel (1998). he skewed-Student density proposed bytoFernández and … = (1998). metry) effect is simple test. γ1 =Steel (iv) Gaussian (normal) distribution q = 0 implies that the impact of a shock is symmetric, i.e., past positive shocks have the same impact on today’s volatility as past v) Gaussian (normal) distribution Where the log-likelihood (iv) Gaussian (normal) distribution function of the distribution is: negative shocks. (iv) Gaussian (normal) distribution here the log-likelihood function of the distribution is: theyfunction do have a drawback in distribution that they Our preferred option is therefore the log-likelihood function of the is: Where Where the log-likelihood of the distribution is: are symmetric. the skewed-Student and Steel (1998). 1 T density proposed by Fernández Lnorm   1T[log( 2T  )  log(  t2 )  z t2 ]2 [5 ] 2 1 2  )  log(  2 ) [5z]t [5] 1 [log( Lnorm Lnorm 2 t  2[log(  ) 2log(   t )  z tt 2 2 t  1distribution (iv) Gaussiant (normal) 1 [5] ) Student-t distribution Where the log-likelihood function of the distribution is: (v) Student-t distribution (v) Student-t distribution (v) Student-t distribution 1 T is: here the log-likelihood function of the distribution L ofthe function [log(of 2the )  distribution log(  t2 )  z t2 is:[5 ]  Wherefunction the log-likelihood the log-likelihood function is: Where Where the log-likelihood of thenorm distribution is: 2distribution t 1    1    1 Student-t 1(  2 )    Lstud  T(v)    distribution   log      1log  log 1     1  2   log  log 2log  2 log  T   log   log Lstud LTstud ( 2() 2 )  Where   2of thedistribution 22 2 2 is:   2  2 function T the log-likelihood   zt   2 1 2 T  1T  log( 1  t )2  (1 2  ) log  1  v z2t2   z t[ 6 ]  1 1) log  1  1 6] [(6 ] 2 )  (1t )  log( 1   )(1log 2 t  1log( t)T log  logv2 v 2[log  2 t 1 2 Lt stud  2     2  2     i) Generalized Error distribution (GED)  z t2   1 T  (GED) 2 Generalized Error distribution   log(  ) ( 1  ) log 1  (GED)    (vi) (vi) Generalized Error distribution   t  v  2   2distribution t  1 distribution here the log-likelihood function of the is: (GED)  (vi) Generalized Error the log-likelihood function of the distribution is: Where Where the log-likelihood function of the distribution is: [6] [6]  (vi) Generalized Error distribution (GED)    Where  thez tlog-likelihood is: 2  1  1 function of thedistribution  T    0 .5     [ 72]        log ( 1  ) log( 2 ) log 0 . 5 log( T     ED ofthe  Where    thelog-likelihood 1   1   0.5t )log( z t 0 .5 z t function 1 distributionis: 1 2       log ( 1  ) log( 2 ) log L   [t 7)] [ 7 ] 1   t GED                log 0 . 5 ( 1  ) log( 2 ) log 0 . 5 log(      GED t )   1               t       t 1 T         zt 1 1 2 ii) (vii) T  (1   ) log( 2 )  log     0 .5 log(  t )  LGED    log    0 .5     t 1      Standardized (zero mean and unit variance) skewed-Student distribution (vii) Standardized (zeroand mean and unit variance) skewed-Student distribution Standardized (zero mean unit variance) skewed-Student distribution (vii) Standardized (zero mean and unit variance) skewed-Student distribution [[7] 7] the log-likelihood function the distribution Where Where the log-likelihood function of theofdistribution is: is: 18 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski Results: model evalutation (vii) Standardized (zero mean and unit variance) skewed-Student distribution Results: model evalutation We evaluate the alternative models an assessment of parameters the parameters associated We evaluate the alternative models by (i)by an(i) assessment of the associated Where themodel log-likelihood function of the distribution is: performance. withmodel each an evaluation of each model forecasting with each set (ii)set an(ii) evaluation of each model set’sset’s forecasting performance.               1 2      2   log( s )    T log  1  L          0 .5 log  ( v  2 )   log  log          log( ) L SkSt  SkSt T  log   v s  log 0 . 5 log ( 2 ) log   1     2   2   2   2     1                  2 T     ( sz  m )  t2  (1   )log (1szt  mt ) 2  2 It  2I   [8] T 0.5   log 2 [8]   2  ( 1 ) log 1  0 .5   log    t 1     [8 ]   t  2 t 1    t (i) Parameter Based Evaluation Parameter Based Evaluation Tables 2, 3 and 4, present the parameter and associated significance tests for the Results: modelvalues evalutation Tables 2, 3 and 4, present the parameter values and associated significance tests for the GARCH, EGARCH and GJR-GARCH specified model sets. For the first and third model We evaluate the alternative models by (i)model an assessment the parameters asGARCH, andinGJR-GARCH specified sets. parameters For the of first third and model setsEGARCH the constants the mean equations and the variance areand positive statis- (i) sociated with in each model set (ii)and an the evaluationparameters of each model set’sand forecasting sets thetically constants the mean statissignificant for allequations distributions. Forvariance the EGARCH modelare setpositive the constants for the performance. tically mean significant for are all not distributions. the EGARCH model set the constants for the equations statisticallyFor significant. mean equations are not statistically significant. The alpha coefficient for all models and distributions is statistically significant at the (i) Parameter Based Evaluation The99% alpha coefficient for allThis models andthe distributions statistically significant the level of confidence. implies existence ofisthe ARCH process in the at residuals 99% Tables level confidence. This implies existence the ARCH process in the residuals term.ofThe time-varying volatilityofvalues clustering; this indicates that periods of 2, 3returns and 4,exhibit present thetheparameter and associated significance term. time-varying clustering; this indicates that periods volatility areexhibit followed by periods ofvolatility relative calm. testsThe for returns the GARCH, EGARCH and GJR-GARCH specified model sets. For theoffirst The beta coefficients of the threein models in all equations distribution are and third sets constants the mean andalso thestatistically variance signifiparamvolatility aremodel followed bythe periods of relative calm. eters are significant for all thesignifiEGARCH atcoefficients the 99%and level of three confidence. This indicates thatdistributions. the is For dependent on its Thecant betapositive ofstatistically the models in all distribution arevariance also statistically model set99% the constants for theofmean equations are statistically significant. average. a subset the models the that sum of not alpha andisbeta is close to unity, cant atmoving the level ofIn confidence. This indicates the variance dependent on its The alpha coefficient for all models and distributions is statistically signifiimplies these cases volatility quiteand persistent suggests that a movingwhich average. In ainsubset of thethat models the shocks sum ofare alpha beta is and close to unity, cant at thepositive 99% level of confidence. This implies the existence of the ARCH prolarge (or negative) return will lead future forecasts of the variance to be high which implies in these cases that volatility shocks are quite persistent and suggests that a for cess in the residuals term. The returns exhibit time-varying volatility clustering; an extended period. return will lead future forecasts of the variance to be high for large positive (or negative) this indicates that volatility are followed byterm periods ofinrelative Theperiod. GARCH periods coefficientof(beta) is larger than the ARCH (alpha) all three calm. model an extended The beta coefficients of the three models in all distribution are also statistisets. This is a further indication that the conditional variance will exhibit long persistence The GARCH coefficient (beta) is larger than the ARCH term (alpha) in all three model cally significant at the 99% level of confidence. This indicates that the variance of volatility. sets. This is a further indication the conditional variance willmodels exhibit long is dependent on its movingthat average. In a subset of the the persistence sum of alpha of volatility. and beta is close to unity, which implies in these cases that volatility shocks are quite persistent and suggests that a large positive (or negative) return will lead future forecasts of the variance to be high for an extended period. The GARCH coefficient (beta) is larger than the ARCH term (alpha) in all three model sets. This is a further indication that the conditional variance will exhibit long persistence of volatility. 19 Forecasting the Jordanian stock index… The GJR models show no evidence of asymmetry effects being statistically significant. EGARCH models however, indicate significance in the magnitude effect but not in the sign effect. Const. (M)α Distribution Normal Student GED Skewed Student α Table 2. The GARCH model set Const. (V)β ARCH (Alpha) GARCH (Beta) Student (DF)µ GED (DF)µ Asymm. Tail Coefficient 0.02 0.01 0.11 0.89 NA NA NA NA p-value 0.05 0.02 0 0 NA NA NA NA Coefficient 0.02 0.01 0.14 0.86 6.23 NA NA NA p-value 0.02 0.02 0 0 0 NA NA NA Coefficient 0.02 0.01 0.12 0.88 NA 1.36 NA NA p-value 0.02 0.02 0 0 NA 0 NA NA Coefficient 0.02 0.01 0.14 0.86 NA NA -0.01 6.22 p-value 0.05 0.02 0 0 NA NA 0.69 0 Mean equation, β Variance equation, µ Degrees of freedom. S o u r c e : estimated by the authors using OxMetricsTM 7. Table 3. The EGARCH model set Const. Const. ARCH GARCH (M)α (V)β (Alpha) (Beta) Student (DF)µ GED (DF)µ Coefficient 0.01 0.12 -0.53 0.99 NA NA NA NA 0 0.39 p-value 0.16 0.74 0 0 NA NA NA NA 0.8 0 Coefficient 0.02 -0.59 -0.52 0.99 6.47 NA NA NA -0.01 0.43 p-value 0.17 0.02 0 0 0 NA NA NA 0.55 0 Coefficient 0.02 -0.68 -0.53 0.99 NA 1.38 NA NA 0 0.41 p-value 0.11 0 0 0 NA 0 NA NA 0.93 0 Coefficient 0.01 -1.11 -0.53 0.99 NA NA -0.02 6.44 -0.01 0.43 p-value 0.15 0.15 0 0 NA NA 0.48 0 0.65 0 Distribution Normal Student GED Skewed Student α Mean equation, β Variance equation, µ Degrees of freedom. S o u r c e : estimated by the authors using OxMetricsTM 7. Asymm. Tail EGARCH EGARCH (Theta1) (Theta2) 20 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski Table 4. The GJR-GARCH model set Const. (M)α Const. (V)β ARCH (Alpha) GARCH Student (Beta) (DF)µ GED (DF)µ Asymm. Tail GJR (Gamma) Coefficient 0.02 0.01 0.11 0.89 NA NA NA NA -0.01 p-value 0.03 0.03 0 0 NA NA NA NA 0.53 Coefficient 0.02 0.01 0.14 0.86 6.22 NA NA NA 0.01 p-value 0.03 0.02 0 0 0 NA NA NA 0.6 Coefficient 0.02 0.01 0.12 0.88 NA 1.36 NA NA 0 p-value 0.02 0.03 0 0 NA 0 NA NA 0.92 Coefficient 0.02 0.01 0.14 0.86 NA NA -0.01 6.21 0.01 p-value 0.05 0.03 0 0 NA NA 0.75 0 0.62 Distribution Normal Student GED Skewed Student α Mean equation, β Variance equation, µ Degrees of freedom. S o u r c e : estimated by the authors using OxMetricsTM 7. We turn now to the issue of identifying the most efficient model(s) from the groups that have been tested. The diagnostic tests of the standardized residuals (Table 5) give us little help in this respect. Both ARCH(10) and Q2(10) statistics indicate that hetroskedasticity has not been fully accounted for by the models which means the estimated volatility equations have to be treated with some degree of caution. Furthermore, Log likelihood and Akaike information criteria based tests all have approximately similar results; this suggests they provide minimal help in distinguishing between model sets on the basis of model fit. We can conclude from this that alternative forecasting-based testing procedures will be required. It can be noted as a caveat to the above conclusion however, that for all model sets, Akaike indicates that the normal distribution produces the worst performance. From this it can possibly be concluded that this distribution can probably be discounted at the outset. 21 Forecasting the Jordanian stock index… Model GARCH EGARCH GJR-GARCH α Table 5. Diagnostic Tests Of The Standardised Residuals Distribution Log Likelihood Q 2(10)α ARCH(10)α Akaike Normal -4044.15 37.14 (0) 3.49 (0) 2.22 Student -3963.3 31.31 (0) 2.92 (0) 2.18 GED -3966.62 33.86 (0) 3.15 (0) 2.17 Skewed Student -3963.22 31.29 (0) 2.92 (0) 2.17 Normal -3980.23 21.62 (0) 2.14 (0.01) 2.18 Student -3893.84 21.01 (0) 2.11 (0.02) 2.13 GED -3900.81 21.28 (0) 2.12 (0.01) 2.13 Skewed Student -3890.96 21.04 (0) 2.11 (0.02) 2.13 Normal -4043.82 38.11 (0) 3.57 (0) 2.22 Student -3963.12 30.56 (0) 2.86 (0) 2.17 GED -3966.62 34.01 (0) 3.17 (0) 2.17 Skewed Student -3963.07 30.59 (0) 2.86 (0) 2.17 The p-values are shown in brackets. S o u r c e : estimated by the authors using OxMetricsTM 7. (ii) Forecasting Based Evaluation The Superior Predictive Ability test can be used for comparing the performances of two or more forecasting models. Forecasts are evaluated using a pre-specified loss function with the ‘best’ forecast model being the one producing the smallest expected loss. An important issue that researchers face is identifying what the loss function is estimated against. For the purposes of this paper losses are estimated relative to the observed returns. The two potential loss functions used by SPA are the Mean Squared Error (MSE) and Mean Absolute Deviation (MAD). The losses estimated from the cho- 22 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski sen function are then compared against a benchmark model. Identifying an appropriate benchmark to use is another important issue in respect to this methodology. In this paper we benchmark against a random walk. The result presented in Table 6 below identify that GJR-GARCH with Skewed distribution produces the smallest loss and is therefore the best fitting model. These findings are consistent with those of Marcucci (2005) and Awartani and Corradi (2005). The latter found GARCH-N to be outperformed by both EGARCH and GJR-GARCH models across different forecast horizons. Performance Table 6. Superior Predictive Ability Tests Model Sample Loss (MSE*103) t-statistic p-value Most Significant GJR-GARCH-Skewed Student 0.00039 16.75819 0 Best GJR-GARCH-Skewed Student 0.00039 16.75819 0 Model_25% GJR-GARCH-Student-t 0.00043 16.75729 0 Median GARCH-GED 0.00049 16.75627 0 Model_75% EGARCH-Normal 0.02761 16.63402 0 Worst EGARCH-Skewed Student 0.03208 16.59516 0 S o u r c e : estimated by the authors using OxMetricsTM 7 and MULCOM 3.0 package (Hansen, and Lunde 2014). Forecasts can also be tested using the model confidence set (MCS) procedure. This uses the same loss function as SPA but requires no benchmark. It identifies efficient model sets at different levels of confidence. Hansen and Lunde state “the set,, that consists of the ‘best’ model(s) from a collection of models,; where ‘best’ is defined in terms of a criterion that is user-specified. The MCS procedure yields a model confidence set,, that is a set of models constructed to contain the best models with a given level of confidence. The models in are evaluated using sample information about the relative performances of the models in ”. (Hansen and Lunde 2014, 16). The result of MCS presented in Table 7 identifies that the 90% confidence model set consists of a single model; namely, GJR-GARCH with Skewed Student distribution. 23 Forecasting the Jordanian stock index… Table 7. Model Confidence Set Tests Model * MSE*103 p-value GARCH-Normal 0.00041 0 GARCH-Student-t 0.00047 0 GARCH-GED 0.00049 0 GARCH-Skewed Student 0.00041 0 EGARCH-Normal 0.02761 0 EGARCH-Student-t 0.03117 0 EGARCH-GED 0.02926 0 EGARCH-Skewed Student 0.03208 0 GJR-GARCH-Normal 0.00047 0 GJR-GARCH-Student-t 0.00043 0 GJR-GARCH-GED 0.0005 0 GJR-GARCH-Skewed Student 0.00039 1.0000* 90% model confidence set. S o u r c e : estimated by the authors using OxMetricsTM 7 and MULCOM 3.0 package (Hansen, and Lunde 2014). Discussion of findings and conclusions Given the large differences between developed and emerging markets, it is possibly a little surprising that the result of our study, made in respect to Jordan, are consistent with previous studies made of developed markets. For example, Engle and Ng (1993), examining Japanese stock return also found strong support for the GJR-GARCH model. Similarly, Bentes, Menezes, and Ferreira (2013) examining NIKKEI 225, S&P 500 and STOXX 50 from 1987–2013 found all stock index returns tested exhibited asymmetry. Further similarities can be identified between our results and other studies. Liu & Hung (2010) investigated the performance of one-step-ahead forecasting using asymmetric GARCH models with different distribution assumptions. Their work, in respect to United States data, concluded that GJR-GARCH generated volatility forecasts were more accurate that those produced by their EGARCH counterparts. Furthermore, their results indicated that modelling the asymmetric component was much more important than specifying the correct 24 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski error distribution when it came to improving volatility forecasting. This was especially the case in the presence of fat-tails, leptokurtosis, skewness and the leverage effect. A number of other studies have examined volatility in MENA countries like Jordan. Assaf (2015), for example, examined the forecasting performance of the Value-at-Risk (VaR) models in Egypt, Jordan, Morocco, and Turkey. Their results suggested that returns had a significantly fatter tails than the normal distribution and that Student APARCH model produced more accurate results than those generated using Normal APARCH models. The considerable variety of results found in these different studies suggests to us that it is difficult to conclude that there is a ‘one size fits all’ model that can be used to model asymmetry affects in stock market returns. We believe that our study contributes significantly to the literature by examining the relative forecasting performances of different distribution-type (Normal, Student-t, GED, and Skewed Student) and asymmetry-type (GJR- GARCH and EGARCH) GARCH models. Both our Superior Predictive Ability and Model Confidence Set results identify that GJR-GARCH with Skewed Student distribution is the best fitting model for Jordan. The finding in our research chimes with the findings of similar studies undertaken in different market contexts; such as Liu & Hung (2010) work in respect to the United States. References Al-Hajieh, H., Redhead, K., & Rodgers T. (2011). Investor sentiment and calendar anomaly effects: A case study of the impact of Ramadan on Islamic Middle Eastern markets. Research in International Business and Finance, 25 (3), 345-356. http://dx.doi. org/10.1016/j.ribaf.2011.03.004. Andrikopoulos, P., Niklewski, J., & Rodgers T. (forthcoming). The portfolio diversification benefits of frontier markets: an investigation into regional effects. Handbook of Frontier Markets. Ed. by Andrikopoulos P., Gregoriou G., and Kallinterakis V. Elsevier. Assaf, A. (2015). Value-at-Risk analysis in the MENA equity markets: Fat tails and conditional asymmetries in return distributions. Journal of Multinational Financial Management, 29, 30-45. http://dx.doi.org/10.1016/j.mulfin.2014.11.002. Awartani, B. M., & Corradi, V. (2005). Predicting the volatility of the S&P-500 stock index via GARCH models: the role of asymmetries. International Journal of Forecasting, 21 (1), 167-183. http://dx.doi.org/10.1016/j.ijforecast.2004.08.003. Balaban, E. (2004). Comparative forecasting performance of symmetric and asymmetric conditional volatility models of an exchange rate. Economics Letters, 83 (1), 99– –105. http://dx.doi.org/10.1016/j.econlet.2003.09.028. Forecasting the Jordanian stock index… 25 Bentes, S. R., Menezes, R., & Ferreira, N. B. (2013). On the asymmetric behaviour of stock market volatility: evidence from three countries. International Journal of Academic Research, 5 (4), 24-32. http://dx.doi.org/10.7813/2075-4124.2013/5-4/A.4. Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of econometrics, 31 (3), 307-327. http://dx.doi.org/10.1016/0304-4076(86)90063-1. Brailsford, T. J., & Faff, R. W. (1996). An evaluation of volatility forecasting techniques. Journal of Banking & Finance, 20 (3), 419-438. http://dx.doi.org/10.1016/03784266(95)00015-1. Brooks, R. (2007). Power arch modelling of the volatility of emerging equity markets. Emerging Markets Review, 8(2), 124-133. http://dx.doi.org/10.1016/j.ememar.2007.01.002. Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica: Journal of the Econometric Society, 50 (4), 987-1007. http://dx.doi.org/10.2307/1912773. Engle, R. F., & Ng, V. K. (1993). Measuring and testing the impact of news on volatility. The journal of finance, 48(5), 1749-1778. http://dx.doi.org/10.1111/j.1540-6261.1993. tb05127.x. Fernández, C., & Steel, M. F. (1998). On Bayesian modeling of fat tails and skewness. Journal of the American Statistical Association, 93(441), 359-371. http://dx.doi.org /10.1080/01621459.1998.10474117. Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993). On the relation between the expected value and the volatility of the nominal excess return on stocks. The journal of finance, 48 (5), 1779-1801. http://dx.doi.org/10.1111/j.1540-6261.1993.tb05128.x. Gokcan, S. (2000). Forecasting volatility of emerging stock markets: linear versus non‐linear GARCH models. Journal of Forecasting, 19 (6), 499-504. http://dx.doi. org/10.1002/1099-131X(200011)19:6%3C499::AID-FOR745%3E3.0.CO;2-P. Hansen, P. R. (2005). A test for superior predictive ability. Journal of Business & Economic Statistics, 23 (4), 365-380. http://dx.doi.org/10.1198/073500105000000063. Hansen, P. R., & Lunde, A. (2005). A forecast comparison of volatility models: does anything beat a GARCH (1, 1)?. Journal of applied econometrics, 20 (7), 873-889. http:// dx.doi.org/10.1002/jae.800. Hansen, P. R., & Lunde, A. (2014). MULCOM 3.00. Econometric toolkit for multiple comparisons. Unpublished working paper. Hansen, P. R., Lunde, A., & Nason, J. M. (2011). The model confidence set. Econometrica, 79 (2), 453-497. http://dx.doi.org/10.3982/ECTA5771. Heynen, R. C., & Kat, H. M. (1994). Volatility prediction: a comparison of the stochastic volatility, Garch (1, 1) and Egarch (1, 1) models. The Journal of Derivatives, 2 (2), 50– –65. http://dx.doi.org/10.3905/jod.1994.407912. Liu, H. C., & Hung, J. C. (2010). Forecasting S&P-100 stock index volatility: The role of volatility asymmetry and distributional assumption in GARCH models. Expert Systems with Applications, 37 (7), 4928-4934. http://dx.doi.org/10.1016/j.eswa.2009.12.022. Marcucci, J. (2005). Forecasting stock market volatility with regime-switching GARCH models. Studies in Nonlinear Dynamics & Econometrics, 9(4), 1-53. http://dx.doi. org/10.2202/1558-3708.1145. 26 H. Al-Hajieh, H. AlNemer, T. Rodgers, J. Niklewski McMillan, D., Speight, A., & Apgwilym, O. (2000). Forecasting UK stock market volatility. Applied Financial Economics, 10(4), 435-448. http://dx.doi.org/10.1080/ 09603100050031561. Mittnik, S., Paolella, M. S., & Rachev, S. T. (2000). Diagnosing and treating the fat tails in financial returns data. Journal of Empirical Finance, 7 (3), 389-416. http://dx.doi. org/10.1016/S0927-5398(00)00019-0. Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach. Econometrica: Journal of the Econometric Society, 59(2), 347-370. http:// dx.doi.org/10.2307/2938260.