(IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, Forecasing Model for Crude Oil Price Using Arificial Neural Neworks and Commodiy Fuures Prices Siddhivinayak Kulkarni Graduae School of formaion Technology and Mahemaical Sciences Universiy of Ballara Ballara, Ausralia S.Kulkarni@ballara.edu.au Imad Haidar Graduae School of formaion Technology and Mahemaical Sciences Universiy of Ballara Ballara, Ausralia I.Haidar@sudens.ballara.edu.au Absrac This paper presens a model based on mulilayer feedforward neural nework o forecas crude oil spo price direcion in he shor-erm, up o hree days ahead. A grea deal of aenion was paid on finding he opimal ANN model srucure. addiion, several mehods of daa pre-processing were esed. Our approach is o creae a benchmark based on lagged value of pre-processed spo price, hen add pre-processed fuures prices for 1, 2, 3,and four monhs o mauriy, one by one and also alogeher. The resuls on he benchmark sugges ha a dynamic model of 13 lags is he opimal o forecas spo price direcion for he shor-erm. Furher, he forecas accuracy of he direcion of he marke was 78%, 66%, and 53% for one, wo, and hree days in fuure conclusively. For all he experimens, ha include fuures daa as an inpu, he resuls show ha on he shor-erm, fuures prices do hold new informaion on he spo price direcion. The resuls obained will generae comprehensive undersanding of he crude oil dynamic which help invesors and individuals for risk managemens. Keywords-Crude Oil; Fuure Price; ANN; Predicion Models I. INTRODUCTION Crude oil is a key commodiy for global economy. As a maer of fac, i is a vial componen for he economic developmen and growh for indusrialized and developing counries in a likely manner. Moreover, poliical evens, exreme weaher, speculaion in financial marke, amongs ohers are major characerisics of crude oil marke which increase he level of price volailiy in he oil markes. The effec of oil price flucuaion exends o reach large number of goods and services which have direc impac on he economy as well as he communiies. Therefore, o reduce he negaive impac of he price flucuaions, i is very imporan o forecas he price direcion. Unforunaely, fundamenal variables such as oil supply, demand invenory, GDP 1 are no available on daily frequency which adds addiional difficully o he predicion. Alhough i could be argued ha he era of oil is abou o be over. However, some sudies have indicaed ha he global demand will coninue o rise for he long-erm despie he fac ha oil demands from OECD counries have decreased, However, he overall demand for oil has increased and his o a large exen due o he increasing demands of non OECD counries, especially China [1]. Furhermore, he fac ha significan amoun of oil come from he unsable Middle Eas means more price flucuaions are expeced. Therefore, forecasing oil price direcion is very useful for marke raders and for individuals. his paper we presen a ANN model for crude oil price predicion for he shor-erm. addiion we es wheher crude oil fuure prices 2 conain newer informaion abou spo price direcion on he shor-erm. This paper proceeds as follows, Secion 2 represens shor lieraure review, Secion 3 presens daa descripion, and preprocessing along wih our mehodology. Secion 4 deails he resuls and discussion, and finally he paper is concluded in Secion 5. II. LITERATURE REVIEW There has been an avalanche of sudies since he oil shock of 1973-74 on forecasing crude oil price. his secion we presen a very brief review of he relaed and recen sudies. Moshiri and Forouan [3] examined he chaos and nonlineariy in crude oil fuures prices. Performing several saisical and economerical ess led hem o conclude ha fuures prices ime series is sochasic, and nonlinear. Moreover, he auhors compared linear and nonlinear models for forecasing crude oil fuures prices, namely, hey compared ARMA and GARCH, o ANN, and found ha ANN is superior and produces a saisically significan forecas. However, in our opinion wo poins can be made regarding his sudy. Firs, he auhors used raw daa as inpu o ANN. Second, hey rained he nework wih raher old informaion (1983 o 2000). Laer in his paper we explain he shor-fall of hese wo poins. Wang e al [4] presen a hybrid mehodology o forecas crude oil monhly prices. The model consiss of combinaion of hree separae componens, Web mining from which he auhors exrac rule based sysem, in addiion ANN, and ARIMA models. These hree componens work disjoinedly, and hen inergraded ogeher o ge he final resuls. They 2 Fuure conracs are defined as: A firm commimen o make or accep delivery of specified quaniy and qualiy of commodiy during a specific monh in he fuure a price agreed upon a he ime he commimen is made' [2] p. 6.
claimed ha nonlinear inegraion of hese hree models has ouperformed any single one. However, here are several issues in his sysem. For example, he rule base sysem of he ex mining model 3 depends on he knowledge base which developed by human expers. This process is no only conroversial, bu also unreliable, because expers opinions vary on he same problem. Moreover, neiher he rules nor he knowledge base was made available o he public. Wen e al [6] proposed SVM model for monhly crude oil prices, he auhors claim ha SVM is ouperforming MLP and ARIMA for ou of. Neverheless, boh sudies used monhly price which limi he daa significanly. addiion, he includes raher old daa from 1970. The relaion beween fuures prices and spo price has been he cenre of aenion for a large number of sudies, and he lieraure is rich wih several sudies covering a range of aspecs wih respec o his relaionship. Lead-lag, efficiency, predicion amongs oher facors, are he mos sudied areas in fuures-spo lieraure. The idea of using commodiy fuures price o predic spo price is based on he assumpion ha he fuures price reacs faser o he new informaion enering he marke han spo price. According o Silvapulle, and Moosa [7] rading in he fuures marke has many advanages, such as low ransacion cos, high liquidiy, and low cash in up-fron, among ohers. This makes i much more aracive for invesors o reac for new informaion han aking posiion in he spo marke. This argumen applies for mos of he commodiy lised in he financial markes; however, i is more relevan o he energy markes. The reason for his is, when new informaion relaed o he oil marke is inroduced, invesors have wo opions, eiher o ake a posiion (buy or sell) in spo or in he fuures marke. mos of he cases, aking a posiion in spo marke is no he bes way for reacing o he new informaion, because i requires high ransacion coss, sorage coss, and delivery coss ec. This is especially rue, if invesors are no ineresed in he commodiy iself raher hey are hedging for anoher commodiy, or simply jus invesing in he marke in hope of arbirage opporuniy i.e. speculaion. his conex, fuures marke is much more aracive place for an invesor o reac o new informaion for he reasons discussed above [7]. However, Brooks, e al [8] examined he lead-lag relaion on high frequency daa 10 min, for FTSE index. The auhors claim he lead-lag 4 relaion could only hold for no more han half hour, and heir resuls confirm ha changes in fuures prices help predicing he changes in spo price. On he oher hand, an early sudy by Bopp and Sizer [9] esed wheher fuures prices are good predicors for cash price in he fuure for he heaing oil marke. aemp o answer if fuures prices has he capaciy o improve forecasing abiliy of economerical models. The resuls showed ha only fuures conrac 1 and 2 monhs o mauriy are saisically significan for cash price forecas. oher words conain new informaion. Similar resuls were obained by Chan [10] who sudied he 3 For a survey of ex mining for financial predicion see [5] 4 Lead-lag refers o a ype of relaion beween fuures prices and spo price for a given asse or commodiies; in which he fuures prices lead spo price, while he spo price lag behind he fuures. (IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, lad-lag relaionship for S&P500. He concluded ha fuures marke is he main source of informaion on marke wide level, while cash marke is he main source of firm specific informaion. Moreover, Silvapulle, and Moosa [7] resuls on he lead-lag of he oil marke suggesed ha he paern of lead lags is no consan and changeable over ime. a relaed sudy by Coppola [11] on crude oil marke; he found some evidence ha fuures conracs are able o reflec he informaion abou he spo price fuure. Abosedra & Baghesani [12] esed monhly fuure prices for long erm-forecas. The resuls showed ha only fuure 1-and 12 monhs ahead produced significan forecas and could be useful for policy making purposes. Alhough he body of oil lieraure is subsanial, here is sill a grea deal of inconsisency in he findings. This is paricularly he case in he relaion beween spo prices and fuures price. While mos sudies agree on he imporance of fuures prices for financial markes, only a few sudies, if any, agree on how, and why i is imporan. Furhermore, he vas majoriy of he lieraure is based on analyical models. A major shorfall of economerical model is making srong assumpion abou he problem. This means if he assumpions are no correc; he model could generae misleading resuls. Furhermore, a recen survey by Labone [13] concludes ha one of he mos common caviies of economerical models is omied variable and in some cases srucural misspecificaion. III. RESEARCH METHODOLGY By viewing he marke as a model ha akes hisorical and curren informaion as an inpu hen marke paricipans reac o his informaion based on heir undersanding, posiions, speculaions, analysis ec. he aggregaion of marke paricipans' aciviies will finally ranslae ino oupu or closing price. order o imiae he marke, a model needs o ake a subse of he informaion available, and ry map i o he desirable arge, hen greaer a forecas (figure 3.1), of course wih a cerain degree of accuracy i.e. wih an error [14]. Figure 1 model selecion o mimic he marke [14] p. 223 his conex ANN is seleced as a mapping model, and viewed as nonparameric, nonlinear, assumpion free model [15]. This means i does no make a priori assumpion abou he problem; raher i les he daa speak for iself [14]. Furhermore, i has been proven ha feedforward nework wih nonlinear funcion is able o approximae any coninuous
(IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, funcion [16], and has been around for a while now and successfully used in several sudies for variey of problems including crude oil price forecas. The mehodology of his sudy is based on hree layers feedforward 5 nework wih backpropagaion algorihm. The goal is o forecas crude oil prices for he shor-erm, and es wheher oil fuures prices conain newer informaion abou he direcion of spo price in he near fuures, hree days ahead. addiion, wheher informaion in fuures price inegraed wih spo price will lead o beer forecas accuracy, he overall sraegy is, creaing a benchmark based on he curren and pas informaion embedded in crude oil spo price solely, using hree layer feedforward ANN. Once his benchmark is creaed fuures prices are added and performance is measured. order o do so effecively, aenion was paid o finding opimal ANN model. A. Design Consideraions he broades sense, here are hree main requiremens for any successful ANN model [14]: Convergence, or in- accuracy. Generalizaion, he abiliy of he model o perform wih new daa. Saiabiliy, consisency of he nework oupu. To insure he above poins are successfully me, a large number of consideraions need o be aken ino accoun, he size and frequency of he daa, nework archiec, he number of hidden neurons, acivaion funcion opimizaion mehods, amongs oher. Alhough design guide lines, and some rules of hump do exiss. However, here is no evidence ha any of hese rules should work for a given problem. Therefore designing neural neworks could be challenging ask. Figure 3.2 shows he main seps o design neural nework for predicion ask. We chose o flow a sysemaic manual approach for finding he opimal nework design for his problem. 5 Recurren neworks were also esed, bu hey made no significan difference in addiion o his hey were compuing exensive. Figure 3.2: A flow char of he main seps for developing ANN models [17] p. 48 (modified) B. Choice of Performance Merics The firs sep in nework design is o choose how o measure he performance of he sysem. For his sudy he ulimae goal is o provide a risk managemen ool. Therefore, our emphasis is on successfully predicing he direcion of he price, raher han he magniude, since i is very difficul, if no impossible, o correcly predic he magniude of he price for financial daa. Besides, our aim is no profiabiliy raher risk managemen, hence predicing he direcion is sufficien o fulfil his goal. The success raio for direcion predicion (or he hi rae) was considered [18].
n 1 h = z n n = 1 z = 1 if x + 1. o + 1 > 0, and 0 oherwise. where: n is he size x, + 1 o + 1 are he value of he arge and he oupu a ime + 1 consecuively. addiion o his he roo mean squared error was also used. The is by far he mos used meric for ANN performance regardless of he nework goal. Furhermore, he correlaion coefficien R and R 2 was also used; as a measure of he linear correlaion beween he forecased value and he acual one [14]. Mean squared error, means absolue error, and sum squared error was also calculaed. Finally, he informaion coefficien given by equaion 3.2 was used. (IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, daa was divided ino raining and esing ses, we use 90% of (3.1) he daa for raining and 10% for ou-of- esing (approximaely one financial year). D. Pre-processing and Normalizaion Pre-processing he raw daa is a very sensiive issue. Alhough in one hand i is desirable o use raw daa, since preprocessing he daa could desroy he srucure inbuil only in he original ime series [15, 21]. However, on he oher hand in order o fi a saisically sound model, ime series need o be saionary. A weakly saionary process should have a consan mean, variance, and auocovariance in is firs and second momenum. Consan auocovariance means ha he covariance of any sequenial values is he same for saionary series [22]. Ic = = 1 n n = 1 ( y x ) ( x x 2 ) 2 1 (3.2) where: y is he prediced value, and x is he acual value. This raio provides an indicaion of he predicion compared o he rivial predicor based on he random walk [14]. Where Ic 1 indicae poor predicion, and Ic < 1 means he predicion is beer predicion han he random walk. C. Daa Collecion Imporan poin for neworks design is deermining he daa frequency and daa size. This is mainly depending on he final goal. For shor-erm forecas, high frequency daa is preferred i.e. inraday or daily daa. However, his daa is no always available and in mos of he case is very cosly o purchase. On he oher hand, weekly, and monhly daa are preferred for oher forecasing horizon because i is less noisy. Equally imporan consideraion is he lengh of he daa. Generally, when dealing wih ANN he more daa poins, he beer he nework generalizaion. Noneheless his is no necessarily he case when dealing wih financial or economical ime series. As economic condiions change over ime, hence non-curren (old informaion) could affec predicion resuls negaively. Because raining he nework wih irrelevan informaion o he curren condiions could resul in a poor model generalizaion [18, 19]. his sudy five ime series are used, Wes Texas ermediae (WTI) ligh swee crude oil spo price and fuures conracs raded a NYMEX. Fuures daa include four conracs 1, 2, 3, 4 monhs o mauriy. The daa frequency is daily closing price; from Sep 1996 o Aug 2007, i includes 2705 daa poins for each ime series. All daa ses rerieved from US Deparmen of energy: Energy formaion Adminisraion web sie: hp://www.eia.doe.gov/ [20]. The For saisical poin of view i is essenial o have a weak saionary ime when modelling for several reasons. Firs and foremos, he behaviour of saionary ime series differs from non-saionary one. As for saionary series a shock will have less influence over ime while for non-saionary he influence of shock remains for longer ime seps [22]. Furhermore, using non-saionary daa could lead o misleading resuls or so called 'spurious regressions' [14, 22]. For ANN model using non-saionary daa makes i easier for he nework o approximae he general characerisics of he daa raher han he acual relaionship [14]. Unforunaely, mos of economical and financial series are non-saionary, which make ransformaion ino saionary form essenial. Several mehods of ransformaion do exis, logarihmic difference, logarihmic reurn, amongs oher. Neuneier & Zimmermann [23] and Gorhmann [24] suggesed he use of he combinaion of firs order relaive change (equaion 3.3) o represen he change in direcion momenum, and he second order relaive change force (equaion 3.4) o represen urning poin of he ie series as nework inpu. x x = x n n y (3.3) x 2x n + x = x n 2n y (3.4) where: x is he original ime series and y is he ransformed of x and n is he forecas horizon. As he auhors did no explain he pre-processing of he arge we chose o es combinaion of hese mehods in addiion o 3 day moving average filer on he raw daa hen ransform i ino relaive change. Anoher imporan issue is nework normalizaion o ransfer he daa o fi wihin he limi of ransfer funcion. A
(IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, linear normalizaion mehod 6 is given by equaion 3.5 o our ineres on modelling wih feedforward nework. The ransfer he daa o fi beween [-1, 1] [18]. Choice of acivaion funcion, learning rae, was deermined by experimens. The opimisaion algorihms 8 also was xz, min( xz ) compared and Leveberg-Maquard was chosen as i y z, = 2. 1 (3.5) approximaes he second order, which leads o fas max( xz ) min( xz ) convergence, and higher hi rae compared o firs order algorihms like gradien decen. E. Nework Archiecure Two ypes of nework archiecs were considered, a recurren nework (Elman) and mulilayer feedforward nework. Anoher imporan design consideraion is he choice of he acivaion funcion 7 because i serves as he nonlinear par of he model. The mos widely used acivaion funcions for financial applicaion in he hidden layer are he sigmoid funcions and he hyperbolic angen, also known as he symmerical sigmoid funcion. [25]. According o McNelis 2005 [18] sigmoid funcions are mosly used in financial applicaions as ransfer funcions because of is hreshold behaviour which describes mos of financial and economical series. However, he recen rend in financial applicaion is oward he hyperbolic angen. The mos vial consideraion, however, is finding he opimal number of hidden neurons since his conrols he number of free parameers in he model. This is especially imporan because, oo many neurons resul in over fiing while oo few could lead o under fiing. Therefore, we used manual sysemaic approach; for each nework 1 o 10 neurons are considered. Each nework was run for minimum of hree imes afer being re-iniiaed wih differen se of weighs. Finally, oher parameers such as learning rae, raining ime, opimizaion algorihm were seleced base on experimens. IV. A. ANN Model EXPERIMENTAL RESULTS AND ANALYSIS Two neworks opology was compared a fully conneced feedforward neworks wih one, and wo, hidden layers o recurren nework (Elman) wih one hidden layer. heory recurren neworks have an advanage over feedforward by modelling dynamic relaionship since he oupu of he neuron is funcion o he previous inpu as well as o he curren one, noneheless, by adding number of lagged value of he inpus feedforward can model such dynamic relaionship as well [17]. Boh ypes has produced comparable resuls, however, recurren neworks ook very long ime o converge (around 8 hours for 100 ieraions on sand alone machine). addiion o his since Elman neworks keep differen numbers of previous seps in he memory layer he resuls vary each ime we re-rain which violae he condiion of consancy. Moreover, he recurren neworks need more hidden neurons o converge because of he memory layer. Hence, we focus TABLE I. HIT RATE AND FOR IN-SAMPLE AND OUT-OF-SAMPLE of of 1 65.89 64.44 0.0313 0.0637 2 67.79 71.11 0.0292 0.0209 3 69.97 69.77 0.0276 0.0204 4 70.09 70.22 0.0273 0.0201 5 70.98 71.55 0.0266 0.0199 6 71.10 70.66 0.0264 0.0203 7 71.30 70.66 0.0260 0.0202 8 72.47 71.55 0.0256 0.0203 9 72.27 73.77 0.0254 0.0199 10 72.80 72.88 0.0248 0.0206 12 72.92 72.88 0.0244 0.0206 14 73.80 73.77 0.0233 0.0208 16* 73.89 72 0.0233 0.0204 18* 75.42 72.44 0.0229 0.0202 20* 74.89 72 0.0242 0.0210 *= model is no sable anymore general, as can be seen from Table 4.3 he resuls were ranging around 50% which is no unusual for noisy daa. For noisy daa he expecaion of he oupu o be correcly prediced is ½ + ε of he ime [14]. This is mainly for wo reasons, he firs is he noise in he daa 9, and he seconds is he limiaion of he daa. Solving noise issue can be done hrough noise filer such as moving average, while solving he daa limiaion or lack of informaion is a more a difficul ask, alhough, in general more feaure need o be added o subsidise for he missing informaion, alernaively if here is any a priori knowledge abou he funcion of he inpu/ oupu his can be modelled and as hin o improve he learning process [14]. Unforunaely, his is no he case of oil ime sires, herefore we have o rely on filering he noise. On he oher hand, daa ransformaion using eq. 3.4 has generaed beer resuls for in and ou of, as eq. 3.4 conain wo seps differencing. Noneheless, combinaions of hese ransformaion mehods were esed as well using one lag each. The bes combinaion which ouperformed all oher opions in is momenum eq. 3.3 and force eq. 3.4 as inpu, and force solely as arge. Following his furher, only his combinaion was esed for muliple lags (up o 12 lag form each equaion). The hi rae performance was improved abou 8% compared o ransformaion by eq 3.4 as inpu and arge 6 A nonlinear normalizaion mehods ([18] p. 64,65) were also esed bu did no improve he resuls. 7 See Duch and Jankowski [26] for a survey of differen acivaion funcions. 8 The convergence ime was also esed when comparing among opimisaion algorihms, as some algorihms need longer ime o converge. 9 For more informaion on he effec of noisy daa on neural nework performance, see Refenes 1995 p.60 and p. 223,224.
alone, while comparing o eq 3.3 he improvemen was subsanial. Finally, he bes performance which is considered as candidae for he benchmark is a combinaion of 7 lags of momenum and 7 lags of force (eq 3.3, 3.4) in he inpu, and 1 lag of force as arge. Furhermore, he nework archiecure is hree layers feedforward wih 8 neurons in he hidden layer. The nework was rained for 1000 ieraions or unil one of he sopping crieria is me. The learning rae is 0.01 and raining algorihm is Levenberg-Marquard. Table 4.5 summarizes he benchmark performance averaged over 5 rials. TABLE II. SUMMARY OF PERFORMANCE OF CANDIDATE BENCHMARK Meric Hi MSE MAE Rae -Sample 74.93 0.02312 0.00052 0.01724 -of- Sample 76 0.01922 0.00038 0.01502 addiion, 3 days moving average was applied on he raw daa, hen daa ransformed ino relaive change. Table 4.6 presens he resuls a differen lags afer applying he moving average filer. TABLE III. HIT RATE AND INPUT 3 DAY FOR IN-SAMPLE AND OUT- OF-SAMPLE of of 1 72.77 73.75 0.0108 0.0079 2 72.88 74.01 0.0107 0.0080 3 73.74 74.54 0.0104 0.0077 4 75.42 76.37 0.0099 0.0073 5 76.02 76.90 0.0096 0.0074 6 76.16 77.16 0.0095 0.0072 7 77.25 77.11 0.0085 0.0077 8 78.01 75.59 0.0091 0.0070 9 78.23 76.77 0.0088 0.0070 10 77.78 78.08 0.0089 0.0069 11 78.03 76.37 0.0086 0.0071 12 77.97 77.95 0.0087 0.0070 13 79.45 79.79 0.0083 0.0068 14 79.39 77.42 0.0083 0.0072 15 79.75 79.11 0.0078 0.0073 16 79.77 79 0.0081 0.0071 17 79.45 78.34 0.0080 0.0069 18 80.40 78.87 0.0076 0.0133 19 80.95 77.16 0.0076 0.0075 20 81.38 77.55 0.0074 0.0074 (IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, Clearly, 3 days simple moving average has improved he resuls significanly, for in and ou of. The resuls above sugges ha 13 lags are opimal, i produced relaively high hi rae and he hi rae was he same for in- as well as for ou-of-. addiion, he was also low. Beside, he model was sable as we esed i for 5 differen rails wih differen ses of weighs. The number of lags is exremely imporan as i explains he dynamics of he sysem. Saic model which includes one lag of he dependen variable ess only he simulaneous relaionship beween he variable [22]. This is a very imporan poin in financial model; as marke paricipan eiher under reac or over reac o new news, hence afer new news inroduced he price migh go up (down) sharply and seles down afer a ime (lags) herefore, including one lag is usually no enough and could generae misleading resuls [22]. B. Fuures Prices order for fuures conracs o predic spo price direcions, informaion embedded in fuures daa should improve he overall resuls. Oherwise, he assumpion ha fuures prices conain newer informaion abou of spo price direcion canno be acceped. We sar measuring how much informaion could be exraced from each fuures conrac abou he direcion of spo price nex day. The moving average and relaive change ransformaion was applied o all fuures conracs and each conrac was presened o he nework as inpu while he spo price was in he same ransformaion. TABLE IV. FUTURES 1 PERFORMANCE AT DIFFERENT LAGS of of 1 71.72 73.55 0.0112 0.0079 2 71.92 73.80 0.0109 0.0081 3 71.42 71.96 0.0108 0.0079 4 73.65 75.15 0.0104 0.0075 5 74.37 74.91 0.0094 0.0085 6 74 76.51 0.0102 0.0074 7 75.23 77.61 0.0099 0.0073 8 74.99 75.26 0.0090 0.0081 9 75.40 75.52 0.0096 0.0085 10 75.64 77.49 0.0095 0.0072 11 76 76.51 0.0095 0.0073 12 75.82 77.24 0.0095 0.0070 13 76.35 76.01 0.0093 0.0073 14 76.91 75.89 0.0092 0.0074 15 76.78 75.52 0.0092 0.0074 16 76.84 78.11 0.0091 0.0072 17 77.94 75.52 0.0088 0.0076 18 78.09 76.63 0.0087 0.0079 19 77.66 77.24 0.0087 0.0078 20 78.87 76.14 0.0086 0.0078
TABLE V. FUTURES 2 PERFORMANCE AT DIFFERENT LAGS of of 1 70.96 71.46 0.0115 0.0083 2 70.80 72.32 0.0112 0.0086 3 70.93 74.29 0.0111 0.0083 4 72.87 74.54 0.0107 0.0080 5 73.20 74.54 0.0106 0.0080 6 73.13 74.54 0.0105 0.0079 7 74.44 74.29 0.0103 0.0077 8 75.22 74.17 0.0100 0.0078 9 74.84 72.94 0.0101 0.0079 10 75.12 75.52 0.0098 0.1078 11 75.71 75.28 0.0097 0.0077 12 76.21 76.75 0.0097 0.0181 13 76.44 76.63 0.0094 0.0078 14 76.64 75.65 0.0096 0.0083 15 76.74 75.15 0.0094 0.0521 16 76.74 74.78 0.0091 0.0086 17 75.52 73.75 0.0109 0.0085 18 77.64 74.29 0.0091 0.0082 19 77.35 74.42 0.0089 0.0085 20 77.28 73.43 0.0092 0.0084 TABLE VI. FUTURES 3 PERFORMANCE AT DIFFERENT LAGS of of 1 70.98 72.45 0.0114 0.0082 2 71.51 73.31 0.0111 0.0084 3 71.53 74.05 0.0110 0.0081 4 73.17 73.43 0.0106 0.0077 5 73.24 73.58 0.0096 0.0088 6 73.79 73.80 0.0104 0.0077 7 75.10 75.03 0.0102 0.0075 8 75.43 72.69 0.0101 0.0077 9 74.73 74.29 0.0099 0.0076 10 75.27 75.28 0.0097 0.0083 11 75.87 76.51 0.0098 0.0075 12 76.72 73.80 0.0094 0.0079 13 76.51 75.15 0.0094 0.0096 14 76.96 75.40 0.0094 0.0077 15 77.22 74.05 0.0094 0.0079 16 77.54 73.68 0.0091 0.0103 17 77.96 74.42 0.0090 0.0077 18 77.08 74.17 0.0089 0.0080 19 77.68 75.77 0.0088 0.0081 20 77.67 72.82 0.0088 0.0085 (IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, TABLE VII. FUTURES 4 PERFORMANCE AT DIFFERENT LAGS of of 1 70.11 72.32 0.0116 0.0084 2 70.20 71.71 0.0113 0.0087 3 70.57 72.94 0.0112 0.0083 4 72.42 74.05 0.0108 0.0081 5 73.05 74.29 0.0107 0.0082 6 72.94 74.66 0.0106 0.0079 7 74.42 75.65 0.0104 0.0079 8 74.16 72.94 0.0103 0.0078 9 74.21 73.68 0.0102 0.0137 10 74.47 74.91 0.0100 0.0081 11 75.72 76.26 0.0099 0.0432 12 74.88 74.66 0.0100 0.0080 13 75.42 74.29 0.0097 0.0154 14 75.49 75.65 0.0098 0.0079 15 75.61 74.17 0.0097 0.0126 16 76.07 76.14 0.0097 0.0085 17 76.43 72.69 0.0095 0.0082 18 76.59 73.80 0.0093 0.0086 19 76.90 74.91 0.0090 0.0398 20 77.02 72.82 0.0092 0.0083 Tables 4.7 o 4.10 show ha none of he fuures conracs alone as inpu was able o ouperform he benchmark. For conrac 1, even wih 20 lags, he forecas was less accurae han wha we obained from spo price solely. Noneheless, i is fair o say ha he performance of fuures as inpu is no poor eiher. However, he real es is wheher any of hese conracs (or a combinaion of hem) will improve he resuls if added o he benchmark. TABLE VIII. FUTURES 1ADDED TO THE BENCHMARK Meric Hi rae MSE MAE 79.18 0.0084 0.0001 0.0063 of 80.44 0.0059 0.0000 0.0046 C. Muli-seps Forecass The final sep is o es wheher he model is able o forecas more han one sep ahead. The inpu included he lagged spo price, while he arge is spo price a imes +1, +2, +3. TABLE IX. HIT RATES FOR IN-SAMPLE AND OUT-OF-SAMPLE FOR 3 DAYS FORECAST OF SPOT PRICE Hi rae +1 +2 +3-78.60 67.45 54 -of- 78.72 66.66 50 Table 4.16 shows ha he hi rae for ou of was accepable for up of 2 days ahead while on average i was equal o flipping a coin for ou-of-. Moreover, able 4.17 illusraes he resuls of adding one lag of fuures prices 1, 2, 3, and 4 monhs o mauriy on op of he benchmark.
TABLE X. (IJCSIS) ernaional Journal of Compuer Science and formaion Securiy, HIT RATES FOR IN-SAMPLE FOR 3 DAYS FORECAST OF SPOT PRICE VI. REFERENCES ADDING 1 LAG OF FUTURES 1, 2, 3, 4 -Sample Hi rae +1 +2 +3 Fuures1 78.35 68.31 56 Fuures2 78.81 67.96 56.67 Fuures3 78.73 67.72 55.45 Fuures4 78.94 68.28 55.35 TABLE XI. HIT RATES FOR OUT-OF-SAMPLE FOR 3 DAYS FORECAST OF SPOT PRICE ADDING 1 LAG OF FUTURES 1, 2, 3, 4 -of-sample Hi rae +1 +2 +3 Fuures1 77.50 66 53 Fuures2 78.69 66.78 52.95 Fuures3 78.97 66.30 48.59 Fuures4 78.84 67.28 49.82 By comparing able 4.16 and 4.17 i is clear ha fuures conracs 1 and 2 improved he direcion forecas for ou-of for ime +3 while conracs 3, and 4 did no add any informaion o he spo price. V. CONCLUSION his paper we presened a model for forecasing crude oil prices on he shor-erm. addiion, we esed he relaion beween crude oil fuures prices and spo price, and if fuures are good predicors o he spo applying nonlinear ANN model. Namely, Daily spo price for WTI and fuures prices for 1, 2, 3, and 4, monhs o mauriy was considered. Daa was obained from Energy formaion Adminisraion covering he period from 1996 o 2007. Several ransformaion mehods was esed, we find ha applying 3 days simple moving average o he original daa hen ransform i ino relaive change is he bes mehods amongs he oher means esed. Moreover, aenions was paid for finding ANN model srucure, as well as discovering he opimal number of lags based on spo price solely as inpu, and use i for benchmark purposes. Then fuures price was added o he benchmark and he performance was compared. Evidence was found in suppor ha fuures prices of crude oil WTI conain new informaion abou oil spo price direcion. Fuures conracs 1, 2 have preformed beer han conracs 3, 4 bu he overall improvemen was insignifican., Alhough, i could be argued ha he relaion beween spo and fuures could be differen during he day. oher words, esing wih inraday daa could produce differen resuls. However, inra day daa for crude oil prices is no available. Finally, our fuure research coninue o invesigae oher variable which could lead o improving he shor-erm forecas, such as heaing oil prices, ineres rae, and gold prices. [1] The ernaional Energy Agency, 2004, Analysis of he impac of high oil prices on he global economy, Repor [2] NYMEX 2006, A guide o energy hedging. [3] S. Moshiri, and F. Forouan, Forecasing nonlinear crude oil fuures prices, The Energy Journal vol. 27, pp. 81-95, 2005. [4] S. Wang, L. Yu, and K. Lai. Crude oil price forecasing wih TEI@I mehodology. Journal of Sysems Science and Complexiy, vol. 18(2), 145-166, 2005. [5] M. Miermayer, and G. Knolmayer, Tex mining sysems for marke response o news: A survey, 2006. [6] W. Xie, L. Yu, L, S. Xu, and S. Wang, A new mehod for crude oil price forecasing based on suppor vecor machines, Lecure noes in compuer science, V.N. Alexanderov e al., Ed. Springer, Heidelberg, pp. 444-451, 2006. [7] P. Silvapulle, and A. Mossa, The relaion beween spo and fuure prices: Evidence from he crude oil marke, The Journal of Fuures Markes, vol. 19(2), pp. 175-193, 1999. [8] C. Brooks, G. Rew, S. Alisair and S. Rison, S, A rading sraegy based on he lead-lag relaionship beween he spo index and fuures conracs for he FTSE 100, ernaional Journal of Forecasing vol. 17, pp. 31-44, 2001. [9] E. Bopp, and S. Sizer, Are peroleum fuures prices good predicors of cash value? The Journal of Fuures Marke, pp. 705-719, 1987. [10] K. Chan, A furher analysis of he lead-lag relaionship beween he cash marke and sock index fuures marke The Review of Financial Sudies vol. 7(6), pp. 123-152, 1992. [11] A. Coppola, Forecasing oil price movemens: Exploiing he informaion in fuure marke. 34, hp://papers.ssrn.com/paper.af?absrac_id=967408, 2007. [12] S. Abosedra, and H. Baghesani, On he predicive accuracy of crude oil fuures prices, Energy Policy, pp. 1389-1393, 2004. [13] M. Labone, The effec of oil shocks on he economy: A review of he empirical evidence, RL31608, 2004. [14] A. Refenes, A., Edior, Neural neworks in he capial markes John Wiley & Sons, New York, 1995. [15] M. Azoff, Neural nework ime series forecasing of financial markes, John Wiley & Sons, Chicheser, 1994. [16] K. Hornik, M. Sinchcombe, and H. Whie, Mulilayer feedforwards are universal approximaors, Neural Neworks 2, pp. 359-366, 1989. [17] G. Bowden, K. James, Forecasing waer resources variables using arificial neural neworks, 2003. [18] D. McNeils, Neural neworks in finance gaining predicive edge in he marke, Elsevier Acadmic Press, Massachuses, 2005. [19] M. Smih, Neural neworks for saisical modeling, Van Nosrand Reinholo, New York, 1993. [20] Energy formaion Adminisraion, 2007, URL: hp://www.eia.doe.gov/, 2007 [21] B. Vansone, Trading in he Ausralian sock marke using arificial neural neworks, 2005. [22] C. Brooks, roducory economerics for finance, Cambridge Universiy press, Cambridge, 2002. [23] R. Neuneier, and H. Zimmermann, How o rain neural neworks, Neural neworks: Tricks of he rade, Genevieve, B. Orr, Klaus-Rober Müller, Ed. Spring, Berlin, pp. 273-423, 1998. [24] R. Grohmann, Muli-agen marke modeling based on neural neworks, 2005. [25] S. Haykin, Neural neworks: A comprehensive foundaion, Prenice Hall, New Jersey, 1999. [26] W. Duch, and N. Jankowski, Survey of neural ransfer funcions. Neural Compuing Surveys 2, pp. 163-212. hp://www.fizyka.umk.pl/publicaions/kmk/99ncs.pdf, 1999.