Price Dynamics and Speculators in Crude Oil Futures Market

Transcription

1 Available online a Sysems Engineering Procedia (011) Price Dynamics and Speculaors in Crude Oil Fuures Marke Hui Bu * Beihang Universiy, Beiing , China Absrac This paper examines he behaviour of crude oil fuures price and volailiy, analyzes he relaionship beween speculaive raders posiions and reurns, and invesigaes wheher speculaive raders posiion changes have a significan effec on crude oil price. I also sudies how speculaion facor influence crude oil reurns and volailiy, wheher reurns are relaed o risks, and wheher financial crises increase volailiy in crude oil fuures markes. The empirical resuls from Granger causaliy reveal ha reurn lead speculaive posiion, which indicaes ha non-commercial or managed money raders are a class of posiive feedback raders or rend followers; and also reveal ha he posiion changes held by speculaive raders will cause crude oil price movemen. Based on he esimaion resuls of GARCH(1,1) model we verify posiion changes of non-commercial or managed money raders can impac crude oil fuures reurns significanly, and indicae reurns are no relaed o condiional variance. Moreover, during he financial crisis, crude oil fuures reurn shows an exreme large volailiy. These findings can help us beer undersand price discovery process in crude oil fuures marke, and is useful in risk managemen and financial engineering. 011 Published by Elsevier B.V. 011 Published by Elsevier Ld. Selecion and peer-review under responsibiliy of Desheng Dash Wu Keyword: crude oil fuures; price dynamics; speculaion; noncommercial posiions; managed money posiions; decision engineering 1. Inroducion In recen years, he imporance of commodiies, especially oil, as common invesmen alernaives o radiional markes has increased in recen years. This leads o more speculaion in crude oil markes han before, and may make he mechanism of price deerminaion a lile differen. We ofen noice aricles in newspapers ha discuss he effecs of speculaion aciviies, bu few conduc quaniaive analyses, or model he effecs. For example, many commens in newspapers imply ha speculaive aciviies of funds in he energy marke have recenly pushed up he price of oil, making i deviae from levels deermined by fundamenals, and have increased volailiy. Exan lieraure has discussed he relaionship beween raders posiions and marke prices, bu hese sudies mainly focus on forecasing abiliy of raders and use Granger causaliy ess for analysis (Harzmark, 1991; Leuhold e al., 1994; Buchanan e al., 001; Wang, 001, 00; Sanders e al., 004) [1-6]. Alhough researchers have examined he level and adequacy of speculaion, flows of funds, and forecasing abiliy of raders, in fuures markes, few have measured he magniude of effecs of raders especially funds rading aciviies, on crude oil fuures markes and price volailiy. This paper ries o invesigae his very ineresing quesion: wha s he relaionship of rading aciviies and price movemens, and wha are he effecs of speculaion on price volailiy. In his paper, we discuss speculaive raders posiions more horoughly, and incorporae his facor ino a model o measure he exen of is impac on crude oil fuures reurns and volailiy. This sudy can reveal he speculaive aciviies of funds in some degree. Differen heories explain differen mechanisms of price deerminaion. According o he classical economic heory, marke fundamenals, especially supply and demand, should be he maor facors ha deermine he price and drive is volailiy. The efficien marke hypohesis (EMH) says if financial markes are informaion efficien, hen he price of raded asses reflecs all known informaion and, herefore, is unbiased, in he sense ha i reflecs he collecive beliefs of all invesors abou fuure prospecs. The behavioural finance heory assers ha asse prices can change wihou changes in fundamenals. For example, volailiy can be induced by anomalies and mass psychology (Deaon and Laroque, 199, 1996 [7, 8] ; Chambers, 1996 [9] ; Shiller, * Corresponding auhor. Tel.: address: buhui@buaa.edu.cn Published by Elsevier B.V. doi: /.sepro

2 Hui Bu / Sysems Engineering Procedia (011) [10] ; ec). Our sudy can provide a supplemen o hese heories and make us undersand price discovery process more horoughly. To measure he effec of rading aciviies on price volailiy in fuures markes is difficul since i is very difficul o rack he aciviies of differen raders. Forunaely, he Commodiy Fuures Trading Commission (CFTC) collecs daa on composiion of open ineres for all fuures conracs, and releases he Commimens of Traders (COT) repor o he public. We can analyze his repor o ge some informaion abou aciviies of differen raders. In he COT repor, open ineres is divided ino reporing and non-reporing raders, wherein raders holding posiions in excess of CFTC prescribed levels repor heir posiions. Reporing raders are furher caegorized as commercials or non-commercials. Commercials are hose associaed wih an underlying cashrelaed business, and are commonly considered o be hedgers. Non-commercials are no involved in an underlying cash business; hey are referred o as speculaors. Furhermore, he repored level of non-commercial aciviy is generally considered o be speculaive aciviies of managed fuures or commodiy funds. We noice ha here are some limiaions in he COT daa. For example, we know nohing abou he moives of non-reporing raders; hese raders may be hedgers, speculaors, or marke makers. Furhermore, as Sanders e al. (004) [6] has poined ou ha he disaggregaing of reporing raders ino commercial and non-commercial marke paricipans has poenial sources of error. In paricular, commercial raders may no always be hedgers, and hedgers may no always be hedging. True hedging posiions are some subses of commercial raders posiions. Toal commercial posiions are likely o reflec very diverse moives. This conclusion is consisen wih findings of Ederingon and Lee (00) [11], who examined commercial raders in he heaing oil marke. To ge over his problem in COT repors, some sudies use non-public daa o break down marke paricipans ino more groups, however hese posiions daa is no available o he public. To obain a rue picure of speculaors posiions is very difficul. On he oher hand, here are no obvious incenives for a rader o classify iself as a speculaor, and i would seem paricularly difficul for a CTA o describe iself as a commercial rader. Thus, repored non-commercial posiions in he COT repor mos likely represen a relaively pure subse of oal speculaive posiions, especially hose held by managed funds. Therefore, we can sill use non-commercial posiions in he COT repors o ge some informaion abou speculaion by managed funds indirecly. To increases ransparency, he CFTC began publishing a Disaggregaed Commimens of Traders (DCOT) repor on Sepember 4, 009, hisorical daa for which are available back o June 13, 006. The DCOT repor separaes reporable raders ino four caegories of raders: producer/merchan/processor/user, swap dealers, managed money, and oher reporables. The CFTC removes swap dealers from commercial caegory and creaes new swap dealers classificaion for reporing purposes. Managed money for he purpose of his repor, is a regisered commodiy rading advisor (CTA), a regisered commodiy pool operaor (CPO), or an unregisered fund idenified by CFTC. These raders are engaged in managing and conducing organized fuures rading on behalf of cliens. Every oher reporable rader ha is no placed ino one of he oher hree caegories is placed ino he oher reporables caegory. The DCOT ses ou open ineres by long, shor, and spreading for he hree caegories of raders swap dealers, managed money, and oher reporables. For he producer/merchan/processor/user caegory, open ineres is repored only by long or shor posiions. This paper makes use of boh COT repor and DCOT repor o analyze he speculaion aciviies.. Daa and saisics We use WTI (NYMEX) fuures prices, raher han spo prices, as he sudy sample. Daily closing prices of he neares conrac of WTI fuures (RCLC1 Cushing, Oklahoma, Crude Oil Fuure Conrac 1) are from EIA. This price series is used as he crude oil fuures price. Based on his price series, reurns are calculaed as he change of he logarihm of he daily closing price of crude oil fuures, i.e. R ln( P / P 1). The descripive saisics of price and reurns are shown in Table 1. Also we repor he auocorrelaion and parial correlaion of reurn and squared reurn in Table 1. The reurn series is a saionary series based on he augmened Dickey-Fuller (ADF) uni roo es. This paper makes use of boh COT repor and DCOT repor o analyze he speculaion aciviies. We use non-commercial posiions in he COT repor and managed money posiions in he DCOT repor o consruc indicaors o reflec speculaive rading aciviies. The ne long posiions (NL) is defines as he long minus shor posiion. And also we use he percen ne long o capure he ne long posiions of speculaive raders. The percen ne long (PNL) posiion is calculaed as long posiions minus shor posiions, divided by he sum of all posiions; he PNL for non-commercial posiions, i.e. NPNL is: NCL NCS NPNL NCL NCS NCSP (1) where NCL, NCS, and NCSP are non-commercial long, shor, and spread posiions, respecively. De Roon e al. (000) [1] calculaed he PNL for repored commercials and referred o i as hedging pressure, and Sanders e al. (004) [6] used PNL of non-commercial posiions in energy fuures markes o deermine if any relaionships exised beween rader posiions and marke prices. Here, we follow he lieraure and make use of PNL of non-commercial posiions and PNL of managed money as he indicaor o describe speculaive rading aciviies.

3 116 Hui Bu / Sysems Engineering Procedia (011) Because he hisorical daa of DCOT repor are back o June 13, 006, herefore sudy sample of his paper is from June 13, 006 o Dec. 8, 010. We sudy he relaionship beween he raders posiion and he price or reurns of crude oil fuures, wih weekly daa. The COT and DCOT daa reflec raders posiions as of Tuesday s close; hus, a maching se of fuures reurns calculaed by Tuesday-o-Tuesday closing prices is used. Fig. 1 gives a descripion of crude oil price and raders posiions. Firs, we sudy he correlaion of price and posiions, wih he resuls repored in Table. We can see ha non-commercial posiion and managed money posiion have a posiive correlaion wih price or reurn. And non-commercial ne long posiion has a very srong posiive correlaion wih managed money ne long posiion, wih a correlaion coefficien This indicaes ha noncommercial posiion is a good proxy for managed fund posiion. Table 1. The descripive saisics of price and reurn, and he auocorrelaion parial correlaion of reurn and squared reurn reurn squared reurn Saisics price Reurn lags AC PAC Q-Sa Prob AC PAC Q-Sa Prob Mean E Median Maximum Minimum Sd. Dev Skewness Kurosis Jarque-Bera Probabiliy Observaions Fig. 1. Crude oil fuures price and he posiions of wo kinds of raders Table. The correlaion of price and posiions NL_noncomm price reurn NL_noncomm PNL_noncomm NL_Money PNL_noncomm NL_Money PNL_Money Based on he augmened Dickey-Fuller (ADF) uni roo es, crude oil price series, ne long of managed money (recorded as MNL), and percen ne long of managed money (recorded as MPNL) are no saionary even a 10% significan level, while ne long of non-commercial (recorded as NNL) and percen ne long of non-commercial (NPNL) are saionary a 5% significan level, bu no significan a 1% level. The Johansen coinegraion es of price and MNL based on no inercep or rend in equaions or es and lag inervals assumpions indicaes ha 1 coinegraing equaion a he 5% level. 3. Granger causaliy es Nex, we sudy he relaionship beween speculaive posiions and reurns of crude oil fuures using Granger causaliy es (Granger, 1969, 1980) [13,14] wih weekly daa. The crude oil fuures weekly reurns and percen ne long of non-commercial and managed money (NPNL, MPNL) series are all saionary series based on he augmened Dickey-Fuller (ADF) uni roo es. We use saionary series, i.e. reurns and percen ne long, o se up he model and make he es. We firsly se up he model shown as

4 Hui Bu / Sysems Engineering Procedia (011) equaion () and (3) and hen process insananeous Granger causaliy es shown as equaion (4) and (5). The lag srucure (m,n) for each OLS regression is deermined o minimize Akaike informaion crierion (AIC). To make sure he equaion is specified correcly, we use Lagrange Muiplier (LM) es o examine he serial correlaion of he residuals. If here exiss serial correlaion, we will add lags of independen variables. We also es he homoscedasiciy of residuals, and we use Whie s heeroscedasic consisen covariance esimaor of coefficien if here exis heeroscedasiciy. We use boh he percen ne long of noncommercial posiions and ha of managed money o make he es. For equaion (), he null hypohesis is H 0 : 0 for each. If his null hypohesis is reeced, ha means reurns lead posiions. For equaion (4), he null hypohesis is H 0 : 0 0 and 0 for each. For equaion (3), he null hypohesis is H : 0 0 for each. For equaion (5), he null hypohesis is H0 : 0 0and 0 for each. If his null hypohesis is reeced, ha means posiions lead reurns. We use Wald coefficien es o es he hypohesis. The resuls of Granger causaliy es are repored in Table 3. m PNL PNL R i i i1 1 m R R PNL i i i1 1 n m PNL PNL R R i i 0 i1 1 n n () (3) (4) m n i i 0 (5) i1 1 R R PNL PNL Table 3. Granger causaliy es beween reurns and percen ne long posiions Granger causaliy es insananeous Granger causaliy es Null hypohesis Reurns don lead NPNL NPNL don lead reurns Reurns don lead NPNL NPNL don lead reurns (m,n) a 1,3 1,1 1, 1,1 F-saisic b.76(0.049) (0.9505) (0.0000).1353(0.0000) χ -saisic 8.866(0.0404) (0.9505) (0.0000) (0000) Impac c (+)* (+)* (+)* Null hypohesis Reurns don lead NPNL NPNL don lead reurns Reurns don lead NPNL NPNL don lead reurns (m,n) 1,1 1,1 1,1 1,1 F-saisic.8191(0.0945) (0.5099) 3.177(0.0000) (0.0000) χ -saisic.8191(0.0931) (0.5099) (0.0000) (0.0000) a. b. c. Impac (+)* (+)* (+)* The lag srucure (m,n) for each OLS regression. The p-value from he Wald F-es and Chi-squared es of he null is in he parenheses. The cumulaive impac of lagged values of he esed variable. For example, for Eq. (a-1) of noe (a), (+) or (-) is he sign of, and an aserisk (*) denoes a reecion of he null a he 10% level (Wald Chi-squared es). From he empirical resuls, we find ha here exiss a unidirecional Granger causaliy from reurns o percen ne long posiions held by speculaive raders, while a bi-direcional insananeous Granger causaliy exiss beween reurns of crude oil and percen ne long posiions. These resuls are similar o he findings of Sanders e al. (004) [6], i.e. non-commercial posiions do no conain any predicive informaion abou reurns, while posiive fuures reurns resul in non-commercial ne long posiions increasing in he following week. When we analyze hese resuls carefully, we can find ou ha he difference beween Granger causaliy es and insananeous Granger causaliy es ells us ha he posiion changes cause he crude oil price movemen. Thus, aking he managed money posiion as an example, we can express our resuls as follows: MPNL MPNL R (6) 1 [0.0035] [0.066] [0.0386] adused R =0.8657

5 118 Hui Bu / Sysems Engineering Procedia (011) R R D( MPNL ) (7) [0.0033] [0.0589] [0.108] adused R = where D X X X, and he number in square brackes below he esimaed coefficien is he sandard error. From he resuls we can clearly see ha reurns lead he posiions, while he changes of posiion held by speculaive funds will cause he price movemen. The effec of posiion changes on crude oil price movemen is srong, wih a coefficien Because reurns lead non-commercial and managed money posiions, his could be indicaive of a class of posiive feedback raders, described as rend followers by De Long e al. (1990) [15]. Because speculaive raders ne long posiions influence crude oil reurns, when we modelling he price dynamics of crude oil fuures price in shor run, we should idenify changes of speculaive posiions as a deerminan facor, i.e. an explanaory variable. 4. GARCH model of crude oil fuures reurns and is implicaions The es resuls of saisics of crude oil fuures reurns menioned in Table 1 have shown he exisence of condiional heeroscedasiciy of reurn series. Though he auocorrelaion (AC) of reurns suggess he series has no significan serial correlaions, he AC of squared reurns clearly suggess ha he reurns are no independen. Combining hese paerns, i seems ha he reurns are indeed serially uncorrelaed, bu dependen. The Lung-Box saisics (Q-Saisics) for auocorrelaion also show he exisence of condiional heeroscedasiciy. Therefore, GARCH model can fi he reurn series. Sadorsky(006) [16] has shown ha he GARCH model fis well for crude oil price and volailiy, and GARCH model is a popular way o modeling he volailiy of crude oil. Thus, we ake use of GARCH model as he ool o do analysis. Bollerslev(1986) [17] proposed he generalized auoregressive condiional heeroscedasic (GARCH) model o capure he properies of volailiy clusers of ime series. The GARCH (1,1) model has been widely used in modeling uncerainy in financial asses reurns. I provides a measure of condiional volailiy in he presence of ime-varying second-order momens, expressed as:, where ~ 1 0, N R, (8) where R is he reurn a ime, is a consan, is serially uncorrelaed errors (innovaions) of reurns wih mean zero, while is he condiional variance of. Coefficiens 1 and 1 reflec he dependence of he curren volailiy on is pas levels, indicaes he degree of volailiy persisence. and he sum 1 1 In his paper, we wan o invesigae some ineresing issues besides capuring he volailiy clusers of ime series of crude oil fuures reurns. From he resuls above, non-commercial raders or managed money raders are posiive feedback raders or rend followers. When he reurn of crude oil fuures is good, posiive feedback raders will increase he posiions of crude oil fuures. And because posiion changes of speculaive raders can affec conemporary reurn, we can infer ha reurn should be auocorrelaed wih a leas 1 lag. However, from he auocorrelaion of reurn series in Table 1, we know ha reurn series has no significan serial correlaions. Then we can infer ha maybe he informaion of price movemen is conveyed by oher variables raher han direcly by iself, for example speculaion posiions. Based on his analysis and he resuls from above secion, we know ha changes of speculaive posiions may be a deerminan facor of price movemen. Thus, we bring his facor ino he GARCH model o modelling he effecs of speculaion. Based on asse pricing model in financial economic heory, if here exiss he raional raders in he marke, reurn should be deermined by he risk. We can use variance as he risk measure, and bring he condiional variance in o he GARCH model, i.e. se up a GARCH-M model, o es wheher he relaionship beween reurn and risk exiss. This resul can also ell wheher here exis he raional raders in crude oil fuures marke. To accomplish he above analysis, we use he GARCH model ha allows exogenous variables o affec he condiional mean. The exogenous variables are non-commercial or managed money percen ne long posiion changes, and condiional variance. For differen issue ha we ry o es, we will choose differen exogenous variables ino he model. By comparing he empirical resuls of each model ha conains differen exogenous variables, we can deermine how he facors influence crude oil fuures prices and volailiy. From he Jarque-Bera saisic of reurns series in Table 1, he probabiliy value shows we can reec he null hypohesis of a normal disribuion. Therefore, in his paper we will use a GARCH model wih suden- disribuion innovaion o capure of fla-ail propery of innovaions. All of he GARCH models are esimaed using he mehod of maximum likelihood. The number of ARCH erm and GARCH erm orders in he model is chosen o minimize he Akaike informaion crierion. Afer he esimaion of GARCH model, we also need o es of residuals, coefficiens, and goodness of fi of he model. We provide he esimaion resuls of GARCH models in Table 4. Table 4. The esimaion resuls of GARCH models wih Suden- innovaion Mean Equaion (1) () (3) (4) (5) (0.73) (1.65)* (0.65) (1.33) (1.0)

6 Hui Bu / Sysems Engineering Procedia (011) (0.13) (0.54) (-0.15) D(NPNL) 1.535(8.35)*** 1.838(8.35)*** D(MPNL) (10.81)*** (10.8)*** Variance Equaion (1.58) (1.59) (1.39) (1.71)* (1.58) (.6)*** (.1)** (.19)** (.41)** (.36)** (8.5)*** (8.75)*** (6.37)*** (8.75)*** (6.16)*** 1 T-DIST. DOF Log likelihood Adused R-squared F-saisic *** *** *** *** ARCH es: F-saisic 1.31 [0.533] 0.301[0.5831] [0.6939] [0.893] [0.9164] volailiy persisence Uncondiional variance a The resuls are repored in equaions of mean and variance. Each column conains he coefficien of he variable, he z-saisics are in parenheses, and he aserisk (*) denoes significance. *** (**, *) denoe significance a 1% (5%, 10%) level. b We also provide ARCH LM diagnosic ess of residuals of he model. Resuls of ARCH LM es are obained wih 1 lag. The p-values of he diagnosic saisics are presened in square brackes. From Table 4, we can see ha he esimaes mee he general requiremen of a GARCH(1,1) model. The general GARCH(1,1) model wih suden- disribuion innovaion or he GARCH(1,1)-M model wihou exogenous variable canno fi he weekly reurn of crude oil fuures very well. If we bring he posiion changes of speculaive raders ino he mean equaion shown in column ()-(5) of Table 4, we find ou ha his facor has some explanaory power for weekly reurn, and is significan in he mean equaion, no maer wha proxy we use, non-commercial or managed money posiions. To invesigae wheher reurns are relaed o risks, we inroduce condiional variance ino he mean equaion, as he model shown in column (1), (3) and (5). All of he resuls show ha condiional variance is no significan in he model, which means reurn is no deermined by risk. The esimaes of all models mee he general requiremen of a GARCH (1, 1) model, he F-saisics of ARCH LM es of hese models indicaing no ARCH effec any more. When we compare hese models, model in column (4) is bes fied he daa, wih maximum Log likelihood. From his resuls, we find ou managed money posiions is beer han non-commercial posiions as a proxy of speculaion o explain he reurn series. We use model in column (4) as an example, we can express our sudy model as follows: R D MPNL, Where 1 ~Suden- (n), (9) The sum ˆ ˆ 1 1 indicaes he degree of volailiy persisence, which is around 0.93 o 0.96 for all hese models. This indicaes he srong volailiy clusers in crude oil fuures weekly reurns. From he model resuls in column (4) of Table 4, we see ha coefficien of D MPNL is 0.76, he coefficiens of ARCH (1) erm and GARCH (1) erm are 0. and 0.74, respecively. This indicaes ha if he percen ne long of managed money increased by 1%, he reurn will increase 0.76%. The speculaion effec in no small. And for he condiional variance, large 1 and 1 give rise o a large. This means ha a large 1 ends o be followed by anoher large, generaing, again, he well-known phenomenon of volailiy clusering. For a GARCH (1, 1) model, he mulisep ahead volailiy forecass converge o he uncondiional variance of, as he forecas horizon increases o infiniy, provided ha Var ( ) exiss. The uncondiional variance of in he GARCH (1, 1) model can be calculaed by (1 ) 1 1. The uncondiional variance calculaed by he GARCH (1, 1) model shown as equaion (9) is Furhermore, we calculae he variance series esimaed by his GARCH model and depic he variance series in Fig.. We can see ha here is an exreme large condiional variance during Sep. 008 o Apr. 009, wih he maximum variance reached a Jan. 0, 009. When we combine he resuls shown in Fig. and Fig. 1, we find ou ha during he period of very large condiional variance, he percen ne long of managed money (MPNL) flucuae largely, even from he ne long posiion o ne shor posiion. This make crude oil fuures price drop quickly, which follows by a large volailiy clusering.

7 10 Hui Bu / Sysems Engineering Procedia (011) Fig.. Condiional variance series esimaed by he GARCH model 5. Conclusion Price dynamic and volailiy has some inheren characerisics ha are commonly seen in asses reurns. Undersanding he behaviour of volailiy is imporan, because i is useful in derivaives valuaion, hedging decisions, and decisions o inves in physical capial ied o producion or consumpion. Furhermore, volailiy is imporan in risk managemen. Volailiy modelling provides a simple approach o calculaing value a risk of a financial posiion. Finally, modelling volailiy of a ime series can improve he efficiency of parameer esimaion and he accuracy of inerval forecass (Tsay, 00) [18]. Thus, his sudy can illusrae he price discovery process in crude oil fuures marke, and is helpful for risk managemen and financial engineering. This paper discusses wheher here are some oher new facors ha influence price volailiy, in addiion o marke fundamenals, as saed by heories of sorable commodiies prices, when he imporance of commodiies as common invesmen alernaives o radiional markes has increased. Using WTI crude oil fuures prices as he sudy obec, his paper concenraes on invesigaing wheher rading aciviies of speculaive raders have a significan effec on crude oil price and is volailiy. We also reveal he characers of rading aciviies of speculaive raders, and sudy how his facor influences crude oil fuures reurns and volailiy. From he empirical resuls, we find ha here exiss a unidirecional Granger causaliy from reurns o percen ne long posiions held by speculaive raders, which indicaes non-commercial raders or managed money are a class of posiive feedback raders or rend followers. Because a bi-direcional insananeous Granger causaliy exiss beween reurns of crude oil and percen ne long posiions, he changes of posiion held by speculaive funds will cause he price movemen. Therefore, when we modelling he price dynamics of crude oil fuures price in shor run, we can idenify changes of speculaive posiions as a deerminan facor ino he model. Based on he saisics of reurn series, we se up he GARCH (1, 1) model ha allows exogenous variables o affec he condiional mean wih Suden- disribuion innovaion. From he esimaion resuls, we verify posiion changes of noncommercial or managed money raders can impac crude oil fuures reurns significanly. When speculaive raders increase he percen ne long posiion, he price will rise; oherwise, crude oil will fall. We also find ha he condiional variance is no a deerminan of crude oil fuures weekly reurn. This indicaes ha raional raders who make invesmen decision by compare he reurn and risk is no enough in crude oil fuures marke, mos of he raders are speculaive raders, who are posiive feedback raders or rend followers. Because speculaive raders are rend followers, if he reurn of crude oil fuures is good, he speculaor is willing o increase heir ne long posiions, and his rading aciviy can push up he crude oil fuures price higher. Then, we can see he crude oil fuures price become higher and higher. Speculaion conribues very much o his price movemen. During he period he financial crisis, here is an exreme large condiional variance. Observing he daa carefully, we find ou ha during his period, he percen ne long of managed money (MPNL) flucuae largely, even from he ne long posiion o ne shor posiion. This make crude oil fuures price drop quickly, which follows by a large volailiy clusering. These resuls have profound implicaions for crude oil fuures marke supervision and risk managemen. Firs, regulaory auhoriy and invesors, especially hedgers, should pay close aenion o rading aciviies of speculaive raders. They are indeed an imporan deerminan of crude oil fuures price level. By changing posiions, non-commercial raders or managed money raders can push up and down crude oil fuures price. Second, if regulaory auhoriy can provide more deailed daa abou differen kinds of raders o he public, analyss and researchers can beer undersand he relaionship beween speculaive aciviies and crude oil price, and hereby can give more accurae forecas of crude oil price. 6. Copyrigh All auhors mus sign he Transfer of Copyrigh agreemen before he aricle can be published. This ransfer agreemen enables Elsevier o proec he copyrighed maerial for he auhors, bu does no relinquish he auhors' proprieary righs. The copyrigh ransfer covers he exclusive righs o reproduce and disribue he aricle, including reprins, phoographic reproducions, microfilm or any oher reproducions of similar naure and ranslaions. Auhors are responsible for obaining from he copyrigh holder permission o reproduce any figures for which copyrigh exiss.

8 Hui Bu / Sysems Engineering Procedia (011) Acknowledgemens This paper is suppored by Naional Naural Science Foundaion of China (NSFC No ). References 1. M. Harzmark. Luck and forecas abiliy: deerminans of rader performance in fuures markes. Journal of Business. 64 (1991) R.M. Leuhold, P. Garcia, and R. Lu, The reurns and forecasing abiliy of large raders in he frozen pork bellies fuures marke. Journal of Business. 67 (1994) W.K. Buchanan, P. Hodges, and J. Theis, Which way he naural gas price: an aemp o predic he direcion of naural gas spo price movemens using rader posiions. Energy Economics. 3 (3) (001) C. Wang, Invesor senimen and reurn predicaabiliy in agriculural fuures markes. Journal of Fuures Markes. 1 (001) C. Wang, The effec of ne posiions by ype of rader on volailiy in foreign currency fuures markes. Journal of Fuures Markes. (00) D.R. Sanders, K. Boris, and M. Manfredo, Hedgers, funds, and small speculaors in he energy fuures markes: An analysis of he CFTC s Commimens of Traders repors. Energy Economics. 6 (004) A. Deaon and G. Laroque, On he behavior of commodiy prices. Review of Economic Sudies. 59 (199) A. Deaon and G. Laroque, Compeiive sorage and commodiy price dynamics. Journal of Poliical Economy. 104 (5) (1996) M. Chambers and R. Bailey, A heory of commodiy price flucuaions. Journal of Poliical Economy. 104 (5) (1996) R. Shiller, From efficien marke heory o behavioral finance. Journal of Economic Perspecives. 7 (1) (003) L. Ederingon and J.H. Lee, Who rades fuures and how: Evidence from he heaing oil marke. Journal of Business. 75 (00) F.A. De Roon, T.E. Niman, and C. Veld, Hedging pressure effecs in fuures markes. Journal of Finance. 55 (000) C.W.J. Granger, Invesigaing causal relaions by economeric models and cross-specral mehods. Economerica. 37 (1969) C.W.J. Granger, Tesing for causaliy: A personal view. Journal of Economic Dynamics and Conrol, (1980) J.B. De Long, A. Shleifer, L.H. Summers, and R.J. Waldmann, Posiive feedback invesmen sraegies and desabilizing raional speculaion. Journal of Finance. 45 (1990) P. Sadorsky, Modeling and forecasing peroleum fuures volailiy. Energy Economics. 8 (006) T. Bollerslev, Generalized auoregressive condiional heeroscedasiciy. Journal of Economerics. 31 (1986) R.S. Tsay (eds.), Analysis of Financial Time Series: Financial Economerics, John Wiley & Sons, New York, 00.