econstor Make Your Publications Visible.

Transcription

1 econsor Make Your Publicaions Visible. A Service of Wirschaf Cenre zbwleibniz-informaionszenrum Economics Choi, Hankyeung; Leaham, David J.; Sukcharoen, Kunlapah Aricle Oil price forecasing using crack spread fuures and oil exchange raded funds Conemporary Economics Provided in Cooperaion wih: Universiy of Finance and Managemen, Warsaw Suggesed Ciaion: Choi, Hankyeung; Leaham, David J.; Sukcharoen, Kunlapah (205) : Oil price forecasing using crack spread fuures and oil exchange raded funds, Conemporary Economics, ISSN , Vizja Press & IT, Warsaw, Vol. 9, Iss., pp , hp://dx.doi.org/0.5709/ce This Version is available a: hp://hdl.handle.ne/049/4896 Sandard-Nuzungsbedingungen: Die Dokumene auf EconSor dürfen zu eigenen wissenschaflichen Zwecken und zum Privagebrauch gespeicher und kopier werden. Sie dürfen die Dokumene nich für öffenliche oder kommerzielle Zwecke vervielfäligen, öffenlich aussellen, öffenlich zugänglich machen, verreiben oder anderweiig nuzen. Sofern die Verfasser die Dokumene uner Open-Conen-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gesell haben sollen, gelen abweichend von diesen Nuzungsbedingungen die in der dor genannen Lizenz gewähren Nuzungsreche. Terms of use: Documens in EconSor may be saved and copied for your personal and scholarly purposes. You are no o copy documens for public or commercial purposes, o exhibi he documens publicly, o make hem publicly available on he inerne, or o disribue or oherwise use he documens in public. If he documens have been made available under an Open Conen Licence (especially Creaive Commons Licences), you may exercise furher usage righs as specified in he indicaed licence.

2 29 Primary submission: Final accepance: Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds Hankyeung Choi, David J. Leaham 2, Kunlapah Sukcharoen 2 ABSTRACT KEY WORDS: JEL Classificaion: Given he emerging consensus from previous sudies ha crude oil and refined produc (as well as crack spread) prices are coinegraed, his sudy examines he link beween he crude oil spo and crack spread derivaives markes. Specifically, he usefulness of he wo crack spread derivaives producs (namely, crack spread fuures and he ETF crack spread) for modeling and forecasing daily OPEC crude oil spo prices is evaluaed. Based on he resuls of a srucural break es, he sample is divided ino pre-crisis, crisis, and pos-crisis periods. We find a unidirecional relaionship from he wo crack spread derivaives markes o he crude oil spo marke during he pos-crisis period. In erms of forecasing performance, he forecasing models based on crack spread fuures and he ETF crack spread ouperform he Random Walk Model (RWM), boh in-sample and ou-of-sample. In addiion, on average, he resuls sugges ha informaion from he ETF crack spread marke conribues more o he forecasing models han informaion from he crack spread fuures marke. oil price forecasing, crack spread fuures, oil-relaed exchange raded funds, mulivariae GARCH model C8, C58, G7, Q47 Minisry of Sraegy and Finance, Sejong Special Self-Governing Ciy, Republic of Korea; 2 Deparmen of Agriculural Economics, Texas A&M Universiy, USA. Inroducion Crude oil is an imporan energy commodiy ha is vial for mos economic aciviies. Consequenly, numerous sudies have been devoed o modeling and forecasing crude oil prices. In recen years, he link beween he crude oil and refined produc markes has been addressed ofen in he energy economics lieraure. In paricular, several sudies have examined he exisence of a long-run equilibrium relaionship beween he prices of crude oil and refined producs (Asche, Gjolberg, & Völker, 2003; Gjolberg & Johnsen, Correspondence concerning his aricle should be addressed o: Kunlapah Sukcharoen, Deparmen of Agriculural Economics, Texas A&M Universiy, College Saion, TX 77843, USA, Phone: + (34) kunlapah@amu.edu 999; Haigh & Hol, 2002; Lanza, Manera, & Giovannini, 2005; Serleis, 994). The emerging consensus from hese sudies is ha crude oil and refined produc prices are coinegraed. As a resul, we should be able o forecas fuure oil price movemens based on informaion from he refined produc markes. However, only a few empirical works have invesigaed he abiliy of refined produc prices o forecas he price of crude oil (see, for example, Gjolberg & Johnsen, 999; Lanza e al., 2005; Mura & Toka, 2009). A deeper undersanding of he relaionship beween he crude oil and refined produc markes and of he predicive power of he refined producs are indeed worhy of invesigaion and may carry imporan implicaions for energy consumers, producers, invesors and policymakers. Vizja Press&IT

3 30 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen In his sudy, we explore he link beween he crude oil spo and refined produc derivaives markes and examine he abiliy of he refined produc derivaives prices o predic movemens in he crude oil price. In oil and energy markes, he profis of oil refiners, he major paricipans in he markes, depend largely on he crack spread (he difference beween he price of crude oil and he prices of refined producs ypically gasoline and heaing oil). Because he demand from oil refiners, whose producion decisions are ied direcly o he crack spread, largely affec he price of crude oil (Verleger, 982; Verleger, 20), i is possible ha here is a long-run equilibrium relaionship beween crude oil and crack spread derivaives prices. This sudy herefore examines he exisence of long-run equilibrium price relaionships beween he crude oil and crack spread derivaives markes. Specifically, we explore he equilibrium price relaionships beween (i) he OPEC crude oil spo and crack spread fuures and (ii) he OPEC crude oil spo and Exchange Traded Fund (ETF) crack spread. The Granger causaliy is hen used o analyze he lead-lag relaionship and deermine wheher he crack spread derivaives prices are useful for forecasing he movemens of crude oil spo prices. Finally, we compare he forecasing abiliy of he wo crack spread derivaives wih ha of a convenional Random Walk Model (RWM). Our paper conribues o he exising lieraure by enriching he undersanding of he dynamic relaionships beween crude oil spo and refined produc derivaives prices in he following ways. Firs, unlike many previous sudies, we focus on he prices of crude oil spo and crack spread derivaives. Wih he noable excepion of Mura and Toka (2009), our sudy is among he firs o examine he link beween he crude oil spo and crack spread derivaives markes and o invesigae oil price forecasing models based on informaion from he crack spread derivaives markes. Unlike Mura and Toka (2009), we analyze no only he fuures marke bu also he ETF marke. Thus, he findings from his sudy also have implicaions regarding which marke conribues more o he forecasing models. In addiion, while Mura and Toka (2009) ignore he presence of heeroskedasiciy in he residuals, we correc for he ime-varying naure of he variance and covariance of commodiy prices and reurns by uilizing a Mulivariae GARCH (MGARCH) model. Second, whereas mos of he previous lieraure, including Mura and Toka (2009), focuses on eiher Bren or Wes Texas Inermediae (WTI) crude oil prices, we invesigae he dynamics of he new OPEC Reference Baske (ORB) price. The undersanding of he OPEC crude oil price is imporan given he share of OPEC s crude oil producion and expors. The U.S. Energy Informaion Adminisraion (204) repors, OPEC member counries produce approximaely 40 percen of he world s crude oil and expors approximaely 60 percen of he oal peroleum raded inernaionally. In paricular, he sudy allows us o answer wheher he crack spread price daa from he U.S. derivaive markes can significanly explain OPEC crude oil price movemens. In summary, given he limied empirical invesigaions of he link beween he OPEC crude oil spo and refined produc derivaives markes, he resuls from his sudy should provide useful informaion for boh oil refiners and energy invesors regarding porfolio invesmen and risk managemen. The remainder of his paper is organized as follows. Secion 2 conains a brief discussion of he heoreical background on predicing oil price movemens using crack spread derivaives. Secion 3 describes he daa. Secion 4 presens he mehodology used in forming he forecasing models. Secion 5 discusses he empirical resuls, and finally, Secion 6 concludes. 2. Predicing oil price movemens using crack spread derivaives The idea of forecasing oil price movemens using informaion from he crack spread fuures marke is based on wo differen argumens. The firs argumen relies on he proposiion ha he price of crude oil largely depends on he demand from oil refiners (Verleger, 982; Verleger, 20). The raionale behind his proposiion is ha oil refiners are mos concerned wih he crack spread, and herefore, hey cu heir levels of producion when he price of crude oil is oo high compared wih he prices of heir refined producs (i.e., when he crack spread is oo low). A decrease in producion of refined producs will, in urn, lower he price of crude oil hrough he lower demand for inpu (Verleger, 20). This relaionship suggess ha we should be able o forecas fuure oil price movemens based on informaion from he crack spread markes. Assuming ha he efficien marke hypohesis holds, CONTEMPORARY ECONOMICS DOI: /ce

4 Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds paź-05 3-paź-06 3-paź-07 3-paź-08 3-paź-09 3-paź-0 3-paź- -20 ORB price Crack spread fuures price ETF crack spread price Figure. Daily ORB, crack spread fuures, and ETF crack spread prices hen he prices of fuures conracs based on crack spreads are he opimal forecass of he crack spreads (e.g., Chinn & Coibion, 204; Lean, McAleer & Wong, 200; Ma, 989). Accordingly, informaion conained in crack spread fuures should a leas parially explain fuure oil price movemens. The second argumen relies on he proposiion ha here is a posiive relaionship beween convenience yield (he benefi from physically holding a commodiy raher han holding a fuures conrac for ha commodiy) and marginal producion coss (Heinkel, Howe, & Hughes, 990). The reasoning behind his proposiion is ha, o maximize heir profis, oil refiners respond o increased demand hrough immediae producion when he marginal producion coss are low and hrough he sock kep in invenory when he marginal coss of producion are high. This sraegy implies ha when he marginal producion coss are relaively inexpensive (expensive), he convenience yields are low (high). Because low marginal coss imply high profi margins or crack spreads, he proposiion herefore suggess ha here is a negaive relaionship beween he convenience yields and crack spreads. This negaive relaionship is empirically verified by Edwards and Ma (992) and Kocagil (2004). Given he Theory of Sorage (Kaldor, 939), commodiy spo prices and commodiy fuures prices are relaed hrough he convenience yield. Thus, assuming he efficien marke hypohesis, one should expec he exisence of a long-run equilibrium relaionship beween crude oil spo and crack spread fuures markes and ha variaions in crack spread fuures can help explain crude oil price movemens (see, for example, Asche e al., 2003; Gjolberg & Johnsen, 999; Haigh & Hol, 2002; Lanza e al., 2005; Mura & Toka, 2009; Seleis, 994). Insead of relying solely on crack spread fuures, he recen inroducion of derivaive-based ETFs allows oil refiners and invesors o capure crack spread changes by purchasing equal unis of he ProShares UlraShor DJ-UBS Crude Oil ETF (SCO), he Unied Saes Gasoline Fund (UGA) and he Unied Saes Heaing Oil Vizja Press&IT

5 32 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen Fund (UHN) for he 2:: crack spread (which refers o an approximaion of he profi margin ha oil refiners earn by urning wo barrels of crude oil o one barrel of gasoline and one barrel of heaing oil) because he price of he SCO fund represens he cos of purchasing wo unis of oil, where he values of he UGA and UHN funds are he proceeds derived from selling one uni of he respecive disillae. Because hese funds basically aemp o rack he movemens of he nearby fuures, forwards, opions and swap conracs, a posiive relaionship beween crack spread fuures and he ETF spread can be expeced. Specifically, during he period from January 2009 o December 20, he correlaion coefficien beween he daily prices of crack spread fuures and he ETF crack spread is Figure clearly shows ha he price of he ETF crack spread racks he price of he crack spread fuures fairly well (hough hey are no exacly he same). Given he posiive relaionship beween he wo crack spread derivaives, he ETF crack spread should also conain useful informaion abou fuure oil price movemens. Because he enry barrier in he ETF marke is no as sric as in he fuures markes, more diverse ypes of invesors (i.e., no only oil refiners and insiuional invesors) can ener he ETF marke, which raises he quesion of wheher he ETF crack spread is beer a explaining spo oil price movemens han he crack spread fuures. Given he recen increased invesor ineres in oil and refined produc ETFs and he convenien rading sysem, we expec ha he ETF marke should conribue more o he forecasing models. To he bes of our knowledge, no empirical research has ye direcly addressed he role of he ETF crack spread in predicing he movemens of he crude oil spo price. Therefore, his paper examines he forecasing power of boh crack spread fuures and crack spread ETFs for he firs ime. 3. Daa The analysis uses daily closing spo price daa for he OPEC Reference Baske (ORB), daily closing fuures price daa for he fron-monh crude oil, RBOB regular gasoline, and No. 2 heaing oil fuures conracs, and daily closing ETF price daa for he ProShares Ulra- Shor DJ-UBS Crude Oil ETF (SCO), he Unied Saes Gasoline Fund (UGA) and he Unied Saes Heaing Oil Fund (UHN). Given ha he new ORB price was inroduced on June 6, 2005, he daa for spo and fuures prices used in his sudy cover he period from Ocober 3, 2005 o December 30, 20. The daa for he oil ETFs span he period from January 2, 2009 o December 30, 20. Daa from earlier periods were no used because rading did no begin on he SCO unil November 25, The spo price, fuures price and ETF price daa are obained from OPEC, New York Mercanile Exchange (NYMEX), and New York Sock Exchange (NYSE) daa mines, respecively. The closing prices for he hree series are available a differen imes of he day. The closing spo price on a paricular calendar dae is obained a 4:00 AM (EST) on he nex calendar dae. For insance, he Monday closing spo price is available in he early morning on Tuesday. On each rading day, he NYMEX fuures closing prices are deermined a 2:30 PM (EST), whereas he NYSE ETF closing prices are deermined a 4:00 PM (EST). Given he discrepancies in real ime, cauion mus be exercised in modeling and inerpreing marke relaionships on he same calendar dae. In paricular, on a given calendar dae, he fuures price and he ETF price daa are available before he OPEC crude oil price daa. Thus, i would be inappropriae o allow he daily closing ORB price o influence he fuures and ETF prices on he same calendar dae. This dispariy suggess ha he model describing he relaionship beween he crude oil spo and crack spread derivaives should be recursive in naure. The mos common raio of he crack spread for ligh oil is 3:2: (hree crude oil, wo gasoline, and one heaing oil). However, OPEC crude oil is considered a represenaive of heavier oil compared wih ligher oil such as WTI and Bren. Consequenly, he 2:: crack spread is a beer descripion for he case of ORB prices because heavier crude oil usually yields less gasoline han he ligher oil such as WTI. Given ha crude oil is quoed in dollars per barrel and refined producs are quoed in cens per gallon, gasoline and heaing oil prices are convered o dollars per barrel by muliplying he cens-per-gallon price by 42. Accordingly, he 2:: crack spread fuures are calculaed as = crackspread fuures 42 RB 42 HO CL () where RB, HO, and CL refer o gasoline fuures price per gallon, heaing oil fuures price per gallon, and CONTEMPORARY ECONOMICS DOI: /ce

6 Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds 33 crude oil fuures price per barrel, respecively. Finally, he 2:: ETF crack spread can be calculaed as ETF spread = SCO + UGA + UHN (2) where SCO, UGA, and UHN are he prices of he Pro- Shares UlraShor DJ-UBS Crude Oil ETF (SCO), he Unied Saes Gasoline Fund (UGA) and he Unied Saes Heaing Oil Fund (UHN), respecively. 4. Mehodology The aim of he analysis is o examine he conribuion of wo differen derivaives producs, crack spread fuures and he derivaive-based ETF crack spread, in explaining and forecasing crude oil price movemens. Because he sample period includes he financial crisis in 2008, we firs es for srucural breaks in he daa. We adop a Zivo and Andrews (992) es (hereafer a ZA es) o deermine he breakpoin endogenously from he daa. The ZA es is a simple modificaion of an Augmened Dickey-Fuller (ADF) es for uni roo in which a dummy variable for a mean shif and/or a dummy variable for a rend shif occurring a each possible break dae are added o he ADF es equaion. In his sudy, we use he following ZA model ha allows for boh a mean shif and a rend shif a each possible break dae ( T b ) : µ α β γ ( ) θ ( ) δ ε (3) y = + y + + DU T + DT T + y + b b j j j= where DU, DT = 0 if > T b oherwise The es assumes no srucural break(s) under he null hypohesis of uni roo (i.e., α = 0). The model is esimaed for every possible break dae. The candidae break dae is hen chosen where he one-sided -saisic of α = 0 (agains he alernaive hypohesis ha α < 0) in equaion (3) is a a minimum (mos negaive). If he null hypohesis of no break(s) on he seleced candidae break dae is rejeced, we conclude ha here is a srucural break on ha paricular candidae break dae. Afer idenifying he firs breakpoin, he es is reapplied o each subsample unil he es fails o deec evidence of an addiional break. If srucural breaks over he sample period exis, he analysis is also conduced for k each sub-sample period o examine possible changes in he relaionship beween he variables of ineres. Similar o Mura and Toka (2009), we explore he long-run equilibrium relaionships beween he crude oil spo and each crack spread derivaive using he Error Correcion Model (ECM). Given he iming consideraion discussed above, he ECM framework is appropriae because he model is specified such ha he curren daily reurn in a paricular marke depends on is own pas reurns and pas reurns of he oher marke. Le s denoe he log of he crude oil price a ime and cs denoe he log of he crude oil fuures price a ime. According o he Engle and Granger (987) represenaion heorem, if boh s and cs are inegraed of order one, I(), bu he sochasic error erms are saionary (inegraed of order zero, I(0)), he wo variables are said o be coinegraed. Coinegraion beween he wo variables could hen be esablished hrough he error correcion represenaion as n α α () Δ α2 () Δ τ ε, (4) Δs = + l s + l cs + ECT + n s l l s s l= l= n α α2 () α22 () τ ε, (5) Δcs = + l s + l cs + ECT + n cs l l cs cs l= l= where ε is he saionary error erm, and ECT is he error correcion erm. τ s and τ cs are he adjusmen coefficiens. The ECM for he OPEC crude oil spo and he ETF crack spread follows a similar represenaion as equaions (4) and (5) bu wih ETF in place of cs. We also conduc Granger causaliy ess o examine he lead-lag beween markes and o deermine wheher he crack spread derivaives prices are useful for forecasing he movemens of crude oil spo prices. In he forecasing exercise, we also use he MGARCH model, which was inroduced by Bollerslev, Engle, and Wooldridge (998), o accoun for he presence of heeroskedasiciy in he residuals of he esimaed ECM. In his sudy, he consan condiional correlaion (CCC) MGARCH model and he dynamic condiional correlaion (DCC) MGARCH model are employed. Under he same bivariae error correcion model (equaions (4) and (5)), following Kroner and Sulan (993), he condiional variance of he CCC MGARCH model can be specified as T ε = εs, εcs, I ~ N( 0, H) (6) Vizja Press&IT

7 34 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen where ε s, and ε cs, are error erms following he GARCH (,) model wih a zero mean and a condiional covariance marix H wih a consan correlaion ρ. H h h h h 0 ρ h 0 2 s, scs, s, s, = 2 = DRD 0 h h,, cs, 0 h scs h = cs ρ cs, = β + β ε + β h s, 0s s s, 2 s s, h = β + β ε + β h cs, 0cs cs cs, 2 cs cs, (7) (8) 2 2 where h and h are he condiional variances for he s, cs, crude oil spo and crack spread fuures reurns, respecively. The ε is a sandardized residual vecor wih a mean zero and a variance of one, which, in his sudy, is a 2 vecor of normal, independen, and idenically disribued (iid) innovaion. Because he assumpion of consan correlaion may be oo resricive, we also employ he DCC MGARCH model. In conras o he CCC MGARCH model, he DCC MGARCH model, proposed by Engle (2002), allows a ime-varying correlaion ρ : 2 h,,, 0, 0 s hscs hs ρ hs H = 2 = 0 h h,, cs, 0 h scs h cs ρ (9) cs, ρ = ( λ λ ) ρ+ λρ + λε ε 2 2 s, cs, (0) where ε is he sandardized disurbance vecor, ρ is he uncondiional correlaion of he sandardized residual ( ε ), and λ and λ 2 are parameers ha capure he dynamics of a condiional quasi-correlaion. λ and λ 2 are nonnegaive and saisfy 0 λ+ λ2 <. The MGARCH model for he OPEC crude oil spo and he ETF crack spread follows a similar represenaion as equaions (6) o (0) bu wih ETF in place of cs. Finally, we consruc forecass using a recursive window scheme. One-sep-ahead forecass of crude oil prices are generaed dynamically from he insample parameer esimaes. The forecasing performance of he wo crack spread derivaives is hen evaluaed. A benchmark for evaluaing differen forecasing models is he random walk model (RWM) wihou drif. The accuracy of he forecass obained from he error correcion models is assessed on he basis of he mean-absolue error (MAE) and he roo mean-squared error (RMSE): ( ) n = MAE I = y y / n RMSE ( I) = y y / n n = 2 () where ŷ is he one-sep-ahead forecas of he dependen variable for =,, n, and y is he acual value. For he MGARCH models, he forecasing performance is evaluaed by he following modified MAE and RMSE: n y y MAE( II) = variance ( ) RMSE II = = n = / n y y variance 2 / n (2) where variance is a ime-varying condiional variance in he MGARCH model. This mehod is used for scaling he residuals, which can be hough of as an elemen of variaion ha is no explained by he fied model. The inuiion behind his modificaion is ha he MGARCH model could reduce he unexplained variance (or residual) in he ECM; hus, his benefi should also be aken ino accoun. In addiion, he Diebold-Mariano (Diebold & Mariano, 995) es is also conduced o assess wheher he differences beween he wo rival forecass are saisically significan. 5. Empirical resuls Descripive saisics are presened in Table. The presence of skewness, lepokurosis, and non-normaliy (implied by he significan Jarque-Bera saisics) in all of he series sugges ha he uncondiional disribuions of he crude oil spo, crack spread fuures, and ETF crack spread prices and reurns are asymmeric, fa ailed-ailed, and non-normal. Plos of he enire daa series for he daily prices and daily log reurns are presened in Figures and 2, respecively. As shown in he figures, here are considerable co-movemens among he prices and reurns. However, he crack CONTEMPORARY ECONOMICS DOI: /ce

8 Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds 35 ORB reurn 0,5 0, 0,05 0-0,05-0, 3-paź-05 3-paź-06 3-paź-07 3-paź-08 3-paź-09 3-paź-0 3-paź- Crack spread fuures reurn 0,8 0,6 0,4 0,2 0-0,2-0,4-0,6-0,8-3-paź-05 3-paź-06 3-paź-07 3-paź-08 3-paź-09 3-paź-0 3-paź- Crack spread ETF reurn 0,06 0,04 0,02 0-0,02-0,04-0,06-0,08 3-paź-05 3-paź-06 3-paź-07 3-paź-08 3-paź-09 3-paź-0 3-paź- Figure 2. Daily log reurns of ORB, crack spread fuures, and ETF crack spread Vizja Press&IT

9 36 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen Table. Descripive saisics Saisic Crude oil spo Log price Crack spread fuures ETF crack spread Firs difference of log price (reurn) Crude oil spo Crack spread fuures ETF crack spread Mean Median Sd. Dev Skewness Kurosis JB es (p-value) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) Noe: The JB (Jarque-Bera) es is a es of wheher he series are derived from a normal disribuion. The p-value of zero indicaes ha he null hypohesis he disribuion is normal is rejeced. Table 2. Uni roo ess Variable (X) and is firs difference (ΔX) ADF KPSS log (crude oil spo) X ** ΔX ** log (crack spread fuures) X * ** ΔX ** log (ETF crack spread) X ** ΔX ** Noe: The null hypohesis of he ADF ess is he non-saionariy of he series considered, while he null hypohesis of he KPSS ess is he saionariy of he series considered. * and ** imply rejecion of he null hypohesis a he 5% and % levels, respecively. Table 3. Johansen Tes for Coinegraion Maximum rank Trace saisic 5% criical value crude oil spo and crack spread fuures crude oil spo and ETF crack spread Noe: The null hypohesis of he Johansen es is ha here is no lesser coinegraion equaion han he maximum rank level. The coinegraing vecor of oil spo and crack spread is log(s).073 log(cs).65 = 0, and he coinegraing vecor of oil and ETF is log(oil).756(etf) = 0. CONTEMPORARY ECONOMICS DOI: /ce

10 Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds 37 Table 4. Wald Tes-Granger Causaliy Null hypohesis Whole sample Period Period 2 Period 3 Δs does no Granger cause Δcs 4.469* ** * Δcs does no Granger cause Δs ** Δs does no Granger cause ΔETF ΔETF does no Granger cause Δs 78.9** Noe: The able repors Wald es saisics. * and ** denoe rejecion of he null hypohesis a he 5% and % levels, respecively. spread fuures prices and reurns appear o be relaively more volaile han he ORB and ETF crack spread prices and reurns over he period analyzed. The uni roo behavior of he log price and log reurns series is hen invesigaed using he Augmened Dickey-Fuller (ADF) and he Kwiaknowski-Phillips- Schmid-Shin (KPSS) ess. Table 2 shows he uni roo es resuls. The ADF ess show ha all of he log prices excep he log price of crack spread fuures have a uni roo, bu firs-differencing leads o saionariy. Given he observed dynamics of he price and reurn daa, he ADF regressions are run wih a rend and a consan for he log price series and wih a consan for he reurn series. The KPSS ess sugges ha he log price series are non-saionary, whereas he reurn series are saionary. In summary, alhough here is a dispue regarding he saionariy of he crack spread fuures, he logarihms of he crude oil and ETF crack spread prices are inegraed of order one, I(). We herefore employ he Johansen s coinegraion echnique (Johansen, 988; Johansen, 99) o es wheher he crude oil spo price and is possible predicors (he prices of crack spread fuures and he ETF crack spread) are coinegraed. The resuls are shown in Table 3. The lag lengh selecion is based on he Akaike Informaion Crierion (AIC), Schwarz Bayesian Informaion Crierion (SIC) and Hannan-Quinn Informaion Crierion (HQC). For each pair, he null hypohesis of no coinegraed relaionship is rejeced, whereas he null hypohesis of having a mos one coinegraed relaionship canno be rejeced. This resul implies ha one coinegraing vecor exiss for each pair; hus, he longrun relaionship should be considered when modeling he reurn dynamics. Therefore, an error correcion model (ECM) would be appropriae o model and forecas he dynamic of crude oil prices. 5. Srucural breaks and causaliy es From he resuls of he ZA es, we idenify wo srucural breaks: one on Sepember 2, 2008 and he oher on April 29, Those daes are consisen wih he period of he 2008 financial crisis. In Sepember 2008, Lehman Brohers submied a bankrupcy peiion, and Merrill Lynch was sold o he Bank of America. Addiionally, during ha monh, he global sock markes, including he Dow Jones Indusrial Average, he FTSE of England, he CAC40 of France, he Dax30 of Germany, and he Hang Seng of Hong Kong, dropped precipiously. Likewise, he price of oil decreased significanly o dollars per barrel, is lowes level since he beginning of he 2008 financial crisis. Afer April 2009, he global oil marke recovered from he price collapse resuling from he financial crisis. Because no furher srucural breaks are deeced, he enire sample period was divided ino hree sub-periods: sub-period (Ocober 2005 o Sepember 2008), sub-period 2 (Ocober 2008 o April 2009), and sub-period 3 (May 2009 o December 20). These divisions are denoed as he pre-crisis, crisis, and pos-crisis periods. Based on he deeced srucural breaks, we conduc a Granger-causaliy/ block exogeneiy Wald es for he OPEC crude oil spo marke and he crack spread fuures marke on he enire sample period and he hree sub-sample periods. Because he ETF crack spread daa could only be obained beginning in January Vizja Press&IT

11 38 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen 2009, he causaliy es beween he oil spo and ETF crack spread markes is only applied for he pos-crisis period. Table 4 repors he Wald es saisics. For he whole sample, we find ha he crack spread fuures marke has no significan impac on he crude oil marke, bu an impac of he oil spo marke on he crack spread fuures marke is observed. Similar resuls are obained for he pre-crisis and crisis periods. For he pos-crisis period, however, he dynamic beween he wo markes is reversed; insead, a unidirecional relaionship from he crack spread fuures marke o he crude oil spo marke is deeced. This resul is consisen wih ha of Mura and Toka (2009), who sudy he relaionship beween WTI crude oil and he 3:2: crack spread fuures markes. For he relaionship beween crude oil spo prices and ETF crack spread prices, he block exogeneiy es indicaes a srong unidirecional relaionship from he ETF crack spread marke o he OPEC crude oil marke. In summary, in he pos-crisis period, he price changes in he crack spread fuures and ETF crack spread markes led o fuure OPEC crude oil spo price changes. This resul implies ha, for he pos-crisis period, he crack spread fuures and he ETF crack spread are useful for forecasing he movemens of he OPEC crude oil spo prices. 5.2 Esimaion of he ECM and MGARCH models The esimaion resuls from he four alernaive error correcion models are provided in Table 5. The firs hree models correspond o he relaionship beween he crude oil spo and crack spread fuures markes during (i) he enire sample period, (ii) he pre-crisis and crisis periods (given he Granger-causaliy resuls, hese wo periods are combined), and (iii) he pos-crisis period. The las model corresponds o he relaionship beween he crude oil spo and ETF crack spread markes during he pos-crisis period. The esimaed coefficiens on he error correcion erms (i.e., ECT ) of he crack spread fuures and ETF crack spread reurns are all significan. This resul confirms ha all of he markes adjus o heir long-run equilibrium. The lagged reurns of he crack spread fuures and ETF crack spread are significan only in he pos-crisis period. This resul implies ha, afer he crisis, he curren OPEC crude oil reurn did respond o he reurns of hose variables in he previous period. The esimaion resuls are consisen wih he Granger-causaliy es resuls discussed in he previous secion and hence confirm ha he informaion provided by he wo crack spread derivaives are useful in explaining OPEC crude oil price movemens. To compare and discuss he forecasing performance of he crack spread fuures and ETF crack spread, we only focus on he pos-crisis daa. Taking ino accoun he possibiliy of heeroskedasiciy in he residuals of he esimaed error correcion models, he Breusch-Pagan (BP) and Whie ess are used o es for he presence of heeroskedasic disurbances. The resuls of boh ess confirm ha he null hypohesis of homoscedasiciy is rejeced a he 5% significance level for all residuals from he esimaed error correcion models. Therefore, he mulivariae GARCH-ype models are applied o accoun for he ime-varying variance characerisic in he daa. The esimaion resuls from he CCC and DCC MGARCH models are provided in Table 6. The sum of he coefficiens of arch() and garch() are close o, implying ha shocks cause a high persisence in he volailiy. For boh he ECM-MGARCH (CCC and DCC models for oil spo and crack spread fuures) and ECM-MGARCH 2 (CCC and DCC models for oil spo and ETF crack spread), he sum in he crack spread equaion is higher han in he crude oil spo equaion. This resul suggess ha he shock effec of boh he crack spread fuures and he ETF crack spread is more persisen han he shock effec of he crude oil spo on hose wo crack spread derivaives. Moreover, he λ+ λ2 esimaes of he DCC ECM-MGARCH 2 model are close o (bu less han), meaning ha a shock can move he correlaion away from is long-run average for a considerable amoun of ime. Therefore, he DCC MGARCH model may capure he variaion in he correlaion beween he crude oil spo and ETF crack spread markes more effecively han he CCC MGARCH model. Figure 3 presens he condiional correlaion of he CCC MGARCH models (he sraigh line) and he condiional correlaion of he DCC MGARCH models (he ime-varying line). The correlaion beween he oil spo and ETF crack spread markes in he CCC model (0.5990) is relaively higher han ha beween he oil spo and crack spread fuures markes (0.05). CONTEMPORARY ECONOMICS DOI: /ce

12 Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds 39 Table 5. Error correcion models ECM (Whole sample) OPEC crude oil and crack spread fuures ECM 2 (2005:0-2009:04) ECM 3 (2009:05-20:2) OPEC crude oil and ETF crack spread ECM 4 (2009:05-20:2) ETF s cs s cs s cs s ECT ** ** -0.04* ** ** * s ** * 0.240** 0.380* ** cs ** ETF ** consan Noe: * and ** denoe rejecion of he null hypohesis ha he coefficien is no significan a he 5% and % levels, respecively. Table 6. Mulivariae GARCH models ECM MGARCH (oil spo and crack spread fuures) (2009:05-20:2) ECM MGARCH 2 (oil spo and ETF crack spread) (2009:05-20:2) s s c s ETF CCC arch() 0.759** ** ** ** garch() ** ** 0.895** 0.968** consan * 0.000* * ρ 0.05** ** DCC arch() 0.800** 0.085** ** 0.032** garch() ** ** ** ** consan * 0.000* * λ * ** 0.928** Noe: * and ** denoe rejecion of he null hypohesis ha he coefficien is no significan a he 5% and % levels, respecively. In erms of he DCC model, he condiional correlaion beween he oil spo and crack spread fuures markes is low and someimes negaive, while ha beween he oil spo and ETF crack spread markes is more han 0.5 mos of he ime. Because he dynamic correlaions change srongly over ime in boh he ECM-MGARCH and ECM-MGARCH 2 models, he DCC model is more appropriae han he CCC Vizja Press&IT

13 40 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen Esimaed condiional correlaion of oil spo and crack spread fuures (ECM MGARCH ) Esimaed condiional correlaion of oil spo and ETF crack spread (ECM MGARCH 2) Correlaion_CCC Correlaion_DCC Correlaion_CCC Correlaion_DCC Figure 3. Comparison of uncondiional correlaions and esimaed condiional correlaions model, which assumes consan correlaion. Accordingly, only he DCC MGARCH model is used in he forecasing exercise. 5.3 Forecasing performance The forecasing resuls are repored in Table 7. For he ou-of-sample forecass, he daa are divided ino wo periods: he firs period is from May 2009 o Sepember 20, and he second period is from Ocober 20 o December 20. Based on he ECM and ECM MGARCH models, he forecasing performance of he crack spread fuures and he crack spread ETF is evaluaed. The RWM is used as a benchmark for evaluaing he differen forecasing models. We firs discuss he forecasing resuls from he ECMs and hen hose from he ECM MGARCH models. Based on he MAE ( I ) and RMSE ( I ), he ECMs for crack spread fuures and he ETF crack spread ouperform he RWM boh in-sample and ou-of-sample. This resul is consisen wih ha of Mura and Toka (2009), who show ha he ECM for crack spread fuures ouperforms he RWM in predicing WTI crude oil price movemens. Moreover, we find ha he ETF crack spread is a beer predicor of OPEC crude oil price movemens han he crack spread fuures boh in-sample and ou-of-sample. The resul of he DM es indicaes ha hese resuls are significan a he 5% level. Based on he MAE ( II ) and RMSE ( II ), he ECM- MGARCH model for he ETF crack spread shows beer in-sample forecasing performance han he model for crack spread fuures and he RWM. However, for ou-of-sample, i is difficul o derive consisen resuls beween he wo insrumens. The MAE ( II ) suggess ha he ETF crack spread is a beer predicor han he crack spread fuures, whereas he RMSE ( II ) suggess oherwise. Neverheless, boh ECM-MGARCH models ouperform he RWM boh in-sample and ou-ofsample. Overall, he forecasing performance of boh he ECM and ECM-MGARCH models confirms ha informaion in boh he crack spread fuures and he ETF crack spread markes can be used o predic oil price movemens. In addiion, on average, he ETF crack spread seems o be a beer predicor han he CONTEMPORARY ECONOMICS DOI: /ce

14 Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds 4 Table 7. Forecasing performance In-sample Ou-of-sample cs ETF RWM cs ETF RWM ECM MAE(I) RMSE(I) ECM MGARCH MAE(II) RMSE(II) Noe: The MAE ( II ) and RMSE ( II ) of he RWM are derived by dividing he residual by he ime-invarian variance. crack spread fuures. This resul is as expeced given ha he correlaion beween he oil spo and ETF crack spread markes is relaively higher han ha beween he oil spo and crack spread fuures markes (see Figure 3). A possible explanaion for his resul may be he insiuional differences (such as he magniude of ransacion coss and rading sysems) beween he fuures and ETF markes. Specifically, given ha he ETF rading sysem is accessible o anyone, he ETF marke aracs a greaer variey of invesors (no only oil refiners) and is hus more likely o incorporae new informaion regarding crude oil spo prices faser han he fuures markes, in which mos ransacions are compleed by refinery companies or insiuional invesors. 6. Conclusion A number of sudies have invesigaed he dynamics of crude oil spo prices. However, here is lack of research on modeling and forecasing he movemens of OPEC Reference Baske (ORB) prices. Accordingly, his sudy focuses on he dynamics of he OPEC crude oil prices. In response o he emerging consensus ha crude oil and refined produc (as well as crack spread) prices are coinegraed, we examine he usefulness of he wo crack spread derivaives producs (namely, crack spread fuures and he ETF crack spread) for modeling and forecasing he daily OPEC crude oil spo prices. Specifically, using he Error Correcion Model (ECM) and he Mulivariae GARCH (MGARCH) model, we explore he long-run relaionship beween he OPEC crude oil spo marke and he wo crack spread derivaives markes: he crack spread fuures and ETF crack spread markes. We also evaluae he forecasing performance of he wo crack spread derivaives wih ha of he convenional Random Walk Model (RWM). Based on he wo deeced srucural breaks, we apply a Granger-causaliy es for he OPEC crude oil spo marke and he crack spread fuures marke on he enire sample period and he hree sub-sample periods: he pre-crisis, crisis, and pos-crisis periods. A change in he lead-lag relaionship beween he oil spo and crack spread fuures markes is observed over he sub-sample periods. In paricular, a unidirecional relaionship from he crude oil spo marke o crack spread fuures is deeced for he pre-crisis and crisis periods; however, he relaionship beween he wo markes is reversed in he pos-crisis period. The change in his causal relaionship may be explained by he increasing need of he oil-relaed financial marke for oil price hedging and invesmens. Regarding he relaionship beween he crude oil spo and ETF crack spread markes, he Granger-causaliy es indicaes a srong unidirecional relaionship from he ETF crack spread marke o he OPEC crude oil marke. The resuls herefore sugges ha, in he pos-crisis period, boh crack spread derivaives may be good predicors of OPEC oil price movemens. In erms of he forecasing performance, we find ha he forecasing models (he ECM and ECM- MGARCH models) based on crack spread fuures and he ETF crack spread ouperform he RWM boh insample and ou-of-sample. Thus, he resuls confirm ha here is valuable informaion in boh he crack spread fuures and he ETF crack spread markes. In Vizja Press&IT

15 42 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen addiion, on average, he ETF crack spread is a beer predicor of OPEC crude oil price movemens han he crack spread fuures boh in-sample and ou-ofsample. This resul suggess ha he ETF crack spread marke conribues more o he forecasing models han he crack spread fuures marke. Our findings provide he following pracical implicaions for policymakers and invesors. Firs, he resuls sugges ha shocks in refined produc fuures and ETF markes could easily spread o he crude oil spo marke, which could impac oil producion decisions. Policymakers should herefore design policies o preven exreme flucuaions in he prices of derivaive producs caused by speculaors. Second, our resuls sugges ha invesors could (parly) predic crude oil spo price movemens using informaion flows from he crack spread fuures and ETF crack spread markes. In addiion, he ETF crack spread price is a beer predicor in ha he ETF marke incorporaes new informaion regarding crude oil spo prices faser han he fuures marke. Hence, our findings are useful for insiuional and individual invesors who are ineresed in undersanding and forecasing OPEC crude oil dynamics. However, our resuls are no wihou limiaions. Firs, in comparing he forecasing performance of crack spread fuures and he ETF crack spread, only he pos-crisis daa (he daa from May 2009 o December 20) are used. However, he daa used in he forecasing models correspond o he launch of oil and refined produc ETFs in approximaely 2008 and he change in he lead-lag relaionship beween crude oil and crack spread fuures prices in April Furher sudy using longer periods of ETF daa would be helpful in undersanding he relaionship beween he crude oil spo and crack spread ETF markes. Second, only he 2:: crack spread is evaluaed in his sudy. Oher muli-produc crack spread varians may also be useful in explaining oil price movemens. Finally, our analysis only focuses on forecasing OPEC crude oil prices. The approach can be easily adaped o forecasing oher benchmarks of crude oil prices, including WTI (which is considered by Mura and Toka (2009)), Bren and Dubai. Furher research on his subjec could lead o improved price forecass and risk managemen in oil and refined produc markes. References Asche, F., Gjolberg, O., & Völker, T. (2003). Price relaionships in he peroleum marke: an analysis of crude oil and refined produc prices. Energy Economics, 25 (3), Bollerslev, T., Engle, R. F., & Wooldridge, J. M. (988). A capial asse pricing model wih ime-varying covariances. Journal of Poliical Economy, 96 (), 6-3. Chinn, M. D., & Coibion, O. (204). The predicive conen of commodiy fuures. Journal of Fuures Markes, 34 (7), Diebold, F. X., & Mariano, R. S. (995). Comparing predicive accuracy. Journal of Business and Economic Saisics, 3 (3), Edwards, F. R, & Ma, C. (992). Fuures and opions. New York, NY: McGraw-Hill. Engle, R. F. (2002). Dynamic condiional correlaion: A Simple Class of Mulivariae Generalized Auoregressive Condiional Heeroskedasiciy Models. Journal of Business and Economic Saisics, 20 (3), Engle, R. F., & Granger, C. W. J. (987). Co-inegraion and error correcion: represenaion, esimaion, and esing. Economerica, 55 (2), Gjolberg, O., & Johnsen, T. (999). Risk managemen in he oil indusry: can informaion on long-run equilibrium prices be uilized? Energy Economics, 2 (6), Haigh, M. S., & Hol, M. T. (2002). Crack spread hedging: accouning for ime-varying volailiy spillovers in he energy fuures marke. Journal of Applied Economerics, 7 (3), Heinkel, R., Howe, M., & Hughes, J. S. (990). Commodiy convenience yields as an opion profi. Journal of Fuures Markes, 0 (5), Johansen, S. (988). Saisical analysis of coinegraion vecors. Journal of Economic Dynamics and Conrol, 2 (2-3), Johansen, S. (99). Esimaion and hypohesis esing of coinegraion vecors in Gaussian Vecor Auoregressive Models. Economerica, 59 (6), Kaldor, N. (939). Speculaion and economic sabiliy. Review of Economic Sudies, 7 (), -27. Kocagil, A. E. (2004). Opionaliy and daily dynamics of convenience yield behavior: an empirical CONTEMPORARY ECONOMICS DOI: /ce

16 Oil Price Forecasing Using Crack Spread Fuures and Oil Exchange Traded Funds 43 analysis. The Journal of Financial Research, 27 (), Kroner, K. F., & Sulan, J. (993). Time-varying disribuions and dynamic hedging wih foreign currency fuures. Journal of Financial and Quaniaive Analysis, 28 (4), Lanza, A., Manera, M., & Giovannini, M. (2005). Modeling and forecasing co-inegraed relaionships among heavy oil and produc prices. Energy Economics, 27 (6), Lean, H. H., McAleer, M., & Wong, W. K. (200). Marke efficiency of oil spo and fuures: a mean-variance and sochasic dominance approach. Energy Economics, 32 (5), Ma, C. (989). Forecasing efficiency of energy fuures prices. Journal of Fuures Markes, 9 (5), Mura, A., & Toka, E. (2009). Forecasing oil price movemens wih crack spread fuures. Energy Economics, 3 (), Serleis, A. (994). A co-inegraion analysis of peroleum fuures prices. Energy Economics, 6 (2), U.S. Energy Informaion Adminisraion. (204, May 6). Wha drives crude oil prices? Rerieved from hp:// cfm Verleger, P. K. (982). The deerminans of official OPEC crude oil prices. Review of Economics and Saisics, 64 (2), Verleger, P. K. (20). The margin, currency, and he price of oil. Business Economics, 46 (2), Zivo, E., & Andrews, D. W. K. (992). Furher evidence on he grea crash, he oil-price shock, and he uni-roo hypohesis. Journal of Business and Economic Saisics, 0 (3), Acknowledgemens We sincerely hank wo anonymous referees for consrucive commens and suggesions ha improved he conribuions and presenaion of he paper. Vizja Press&IT

17 44 Vol. 9 Issue Hankyeung Choi, David J. Leaham, Kunlapah Sukcharoen CONTEMPORARY ECONOMICS DOI: /ce