Predicting Direction of Stock Prices Index Movement Using Artificial Neural Networks: The Case of Libyan Financial Market

Transcription

1 British Joural of Ecoomics, Maagemet & Trade 4(4): , 2014 SCIENCEDOMAIN iteratioal Predictig Directio of Stock Prices Idex Movemet Usig Artificial Neural Networks: The Case of Libya Fiacial Market Najeb Masoud1* 1 Departmet of Bakig ad Fiace, College of Ecoomics ad Busiess, Al-zaytooah Uiversity of Jorda, P.O. Box 130, Amma 11733, Jorda. Author s cotributio This work was carried out by the author. The author read ad approved the fial mauscript. th Origial Research Article Received 25 Jue 2013 th Accepted 19 August 2013 th Published 9 Jauary 2014 ABSTRACT Aims: The aim of this paper is to preset techiques idicators of artificial eural etworks (ANNs) model usig for predictig the exact movemets of stock price i the daily Libya Stock Market (LSM) idex forecastig. Study Desig: Research paper. Place ad Duratio of Study: Libya stock market from Jauary 2, 2007 to March 28, Methodology: The data from a emergig market Libya Stock Market are applied as a case study. Twelve techical idicators were selected as iputs of the proposed models. The forecastig ability of the ANN model is accessed usig back-propagatio eural 2 etwork of errors such as MAE, RMSE, MAPE ad R. Two comprehesive parameter settig experimets for both the techical idicators ad the levels of the idex i the market were performed to improve their predictio performaces. Results: The experimetal statistical results show that the ANN model accurately predicted the directio of movemet with the average predictio rate 91% of data aalysis i its best case, which is a perfectly good outcome. The relatioship stregth betwee parameter combiatio ad forecast accuracy measures such as MAE, MAPE, ad 2 RMSE is strog (R 0.99). The statistical ad fiacial performace of this techique is evaluated ad empirical results revealed that artificial eural etworks ca be used as a better alterative techique for forecastig the daily stock market prices. Coclusio: This study proved the sigificace of usig twelve particular techical *Correspodig author: ajeb2000@gmail.com;

2 market idicators which gave also useful results i predictig the directio of stock price movemet. To improve ANN model capabilities, a mixture of techical ad fudametal factors as iputs over differet time period were used to be a effective tool i forecastig the market level ad directio. Keywords: Libya stock market; artificial eural etworks; predictio of stock price idex; techical idicators; back-propagatio errors. 1. INTRODUCTION Stock prices preset a challegig task for traders ad ivestors sice the stock market is essetially dyamic, oliear, complicated, oparametric, ad chaotic i ature [1]. Thus, ivestors ca hedge agaist potetial market risks ad speculators ad arbitrageurs have opportuities to make profit by tradig i stock price idex [ 2]. The early Efficiet Market Theory (EMT) claims that prices move i a radom way ad it is ot possible to develop a algorithm of ay kid that predicts stock prices [3]. Forecastig or predictig stock prices may be doe followig oe or a combiatio of four approaches: fudametal aalysis, techical aalysis, time series forecastig ad machie learig. Each approach has its ow virtues as well as limitatios. A artificial eural etworks (ANNs) or eural etworks (NNs) are computatioal theoretical modellig tools that have recetly composed of several highly itercoected computatioal uits called artificial euros or odes. Each ode performs a simple operatio o a iput to geerate a output that is forwarded to the ext ode i the sequece. This parallel processig allows for great advatages i data processig aalysis ad kowledge represetatio [4,5]. ANNs are widely applied i various braches such as computer sciece, egieerig, medical ad crimial diagostics, biological ivestigatio, aalysig the busiess data, ad ecoometric aalysis research. They ca be used for aalysig relatios amog ecoomic ad fiacial pheomea, forecastig, data filtratio, geeratig fiacial time-series, ad optimisatio [6,7,8,9]. ANN-based models are empirical i ature; however they have bee show to be able to decode o-liear fiacial time-series data, which sufficietly describe the characteristics of the stock markets [10]. Numerous researches such as: [11] ad [12] have bee made to compare ANNs with real-world data aalysis statistical tools. Although, ANNs has have bee successfully applied to loa evaluatio, sigature recogitios, fiacial time-series aalysis forecastig ad may other difficult patter recogitio problems that do ot have a algorithmic solutio or the available solutio is too complex to be foud ad complex dimesioality [13,11,12,14]. If stock market retur fluctuatios are affected by their recet historic behaviour, ANN model which ca prove to be better learig of predictors market price idex [15]. ANN model ofte exhibit icosistet ad upredictable performace o oisy data, while predictig the fiacial market s movemets is more difficult is as cosidered by some researchers [16,17,18]. It is of iterest to study the extet of stock price idex movemet predictability usig data from emergig markets such as of the Libya Stock Market (LSM). Sice its establishmet i 3rd Jue 2006 by decisio o. (134) of the Geeral People s Committee (GPC), to form a joit stock compay with capital of 20 millio Libya diars (LYD), divided ito 2 millio shares with a omial value of 10 LYD per share, as the LSM has preseted a outstadig growth as a emergig market. The LSM is characterised with high volatility i the market returs. Such volatility attracts may local ad foreig ivestors as it provides high retur possibility. The umber of compaies listed i the LSM icreased to 12 i 2010 while it was 598

3 8 i The total tradig volume icreased rapidly from 37.4 millio Libya diars i 2008 to reach millio LYD i 2010, ad 1.5 millio LYD shares were traded ad icreased to 12.9 millio i The lowest total market capitalisatio was 330 millio LYD i 2007 with highest market capitalisatio of 3.6 billio LYD i 2010 [19]. The LSM idex startig from a base of 1000 i April 2008 ad closig at i Jue after a high of ad a low of poits, dow by poits, or 20%, it s clear that the secod quarter of the year 2008 was relatively volatile ad perhaps the most sigificat reasos behid the declie durig I year 2010 the LSM idex has jumped by 80% o April, reachig 1600 poits, which represet the highest level sice 2008, the it closed the year at 1354 poits of the total market capitalisatio, traded value, umber of shares traded ad umber of trades realised i the market [19]. The Libya stock market remais small size ad largely uderdeveloped, iefficiet, illiquid whe compared with other emergig Arab stock markets. The total period of (02/01/ /12/2010) is i etwork traiig ad validatio values are obtaied with differet combiatios of parameters for testig the ANN model (total of data poits). The secod sub-period of (02/01/ /03/2013) is i sample period separately as a model iput is used ad predictio rate is calculated. These data poits are the daily closig stock prices i the currecy of the Libya diars (LYD). The LYD is tied with the USD with a coversio rate of approximately 1 USD = 1.27 LYD. The core objective of this study is to predict the directio of movemet i the daily LSM idex ad aswer the why ad whe these computatioal tools are eeded, the motivatio behid their developmet, ad their relatio to biological systems ad other modellig methodologies, the various learig rules ad ANNs forecast is compared with the statistical forecastig result. To the best kowledge of the author, there is o study that deals with predictio usig artificial eural etworks (ANNs) i Libya Stock Market. The major cotributios of this study are to demostrate ad verify the predictability of stock price idex directio usig ANN model ad the fiacial statistical techique, ad the to compare the empirical results of these two techiques. The rest of the paper is orgaised ito seve sectios as follows. Sectio 2 provides a brief overview of the theoretical literature. Sectio 3 proposes empirical chose methodology. Sectio 4 describes the research data aalysis ad experimets. I Sectio 5, descriptive statistics are used simply to describe the sample. I Sectio 6, the empirical results are summarised ad discussed. The last sectio provides a brief coclusio ad future research. 2. LITERATURE REVIEW I recet years, there have bee a growig umber of studies lookig at the directio of movemets of various kids of eural etwork computig to traditioal statistical methods of aalysis [20]. Both academic researchers ad practitioers have made tremedous efforts to predict the future movemets of stock market prices idex or its retur ad devise fiacial tradig strategies to traslate the forecasts ito profits [21]. I the followig sub-sectios 2.1 ad 2.2, we focus the review of previous studies o ANN model applied to fiacial market prices idex predictio. It is worth metioig here that, at the time of this research, literature review revealed that there is o reported research that applied the ANN for the predictio of stock price movemets i Libya stock market. I additio, there is o research foud that deals with predictio usig artificial eural etworks (ANNs) to predict stock price movemets i Libya stock market. 599

4 2.1 Artificial Neural Networks (ANNs) Artificial eural etworks (ANNs) or eural etworks (NNs) is a itercoected group of atural euros that grew out of research i artificial itelligece specifically attempts to mimic the fault-tolerace ad capacity to lear of biological eural systems by modellig the low-level structure of the brai [22], which the brai basically lears from experiece. Accordig to [23] a artificial eural etwork is a relatively crude electroic model based o the eural structure of the brai. The huma ervous system cosists of billios of euros of various types ad legths relevat to their locatio i the body [24,25]. The Fig. 1 shows a schematic of a oversimplified biological euro with three major fuctioal uits [26,27,28]: dedrites, cell body ad axo. The cell body has a ucleus that cotai iformatio about heredity traits; the dedrites receive sigals from other euros ad pass them over to the cell body; ad the axo, which braches ito collaterals, receives sigals from the cell body ad carries them away through the syapse to the dedrites/syapses of eigh borig euros or erve cell. It is the fial part of a euro to receive a electrical impulse ad is also the area where the impulse is coverted to a chemical sigal. The axo termial trasfers iformatio from its euro ito aother euro, though it does ot come ito physical cotact with the other euro [29,27]. Each axo termial braches off from a euro like figers o a had. Electrical iformatio travels through a euro extremely quickly [30]. While it is i the axo of the erve, this sigal is i the form of a electrical pulse. These pulses are very small, betwee 50 ad 70 millivolts each. Oce the electrical sigal reaches the axo termial, the iformatio is coverted ito a chemical sigal kow as a eurotrasmitter [31]. The axo termial the seds the chemical sigal ito the dedrite of the ext euro, which the coverts this iformatio back ito a electrical sigal ad seds it dow to the ext euro [32]. This basic system of sigal trasfer was the fudametal step of early euro-computig developmet ad the operatio of the buildig uit of artificial eural etworks. Fig. 1. Schematic of biological euro Source: adapted from (33,28] 600

5 Several ecoomists advocate the applicatio of eural etworks to differet fields i fiacial markets ad ecoomic growth methods of aalysis [34,35,36,37]. ANN approach has bee demostrated to provide promisig results i predict the stock market idex or its price retur [27]. [38] apply a eural etwork system to model the tradig of S&P 500 idex futures. The results of the study preseted that the eural etwork system outperforms passive ivestmet i the idex. Based o the empirical results, they favour the implemetatio of eural etwork systems ito the maistream of fiacial decisio makig. [39] I his effort to predict stock market, he had created a three multi-layer, back propagatio artificial eural etworks usig macro-ecoomic idicators as iputs. His data set cosisted of values for (6 variables) over a (150 moths) period extedig from Jue 1982 to December His etworks were traied o the first (75 moths) of the data set, ad the tested o the remaiig (75 moths). [40] compare eural etworks to discrimiate aalysis with respect to predictio of stock price performace. Empirical aalyses fid that a eural etwork models perform better tha discrimiate aalysis i predictig future assigmets of risk ratigs to bods. Che et al. [21] tried to predict the directio of Taiwa Stock Exchage Idex retur. The probabilistic eural etwork (PNN) is used to forecast the directio of idex retur. Statistical performace aalysis of the PNN forecasts is compared with that of the geeralised methods of momets (GMM) with Kalma filter ad radom walk. Empirical results showed that PNN demostrate a stroger predictive power tha the GMM Kalma filter ad the radom walk predictio models. [41] qualified eural etworks based o various techical market idicators to estimate the directio of the Istabul Stock Exchage (ISE) 100 Idex. The approach idicators used are MA, mometum, RSI, stochastic (K%), movig average covergece-divergece (MACD). The empirical results of the study preseted that the directio of the ISE 100 Idex could be predicted at a rate of 60.81%. [42] used ISE-30 ad ISE-ALL idices to see the performaces of several eural etwork models. While, the predictio performace of eural etwork models for daily ad mothly data failed to outperform the lier regressio model, these models are able to predict the directio of the idexes more accurately. [43] tested some smaller compaies o UK idexes FTSE 100, FTSE 250 ad FTSE Small Cap ad cocluded that i these markets the higher predictive ability of techical tradig rules exists. Several researches ted to crossbreed umerous artificial itelligece (AI) techiques approach used to predict stock market performace [44,45,46,47,48,49,50,16,21,51,52, 53,47,54]. [40] applied a hybrid AI approach to predict the directio of daily price chages i S&P 500 stock idex futures. The hybrid AI approach itegrated the rule-based systems ad the eural etworks techique. Empirical results demostrated that reasoig eural etworks (RN) outperform the other two ANN models (back-propagatio etworks ad perceptro). Empirical results also cofirmed that the itegrated futures tradig system (IFTS) outperforms the passive buy-ad-hold ivestmet strategy. [55] compared artificial eural etwork (ANN) performace ad support vector machies (SVM) i predictig the movemet of stock price idex i Istabul Stock Exchage (ISE) Natioal 100 Idex. Experimetal results showed that average performace of ANN model at a rate of 75.74% was foud sigificatly better i predictio tha SVM model at a rate of 71.52%. [ 56] compared the ability of differet mathematical models, such as ANN, (ARCH) ad (GARCH) models, to forecast the daily exchage rates Euro/USD usig time series data of Euro/USD from December 31, 2008 to December 31, Empirical aalyses fid that the ARCH ad GARCH models, especially i their static formulatios are better tha the ANN for aalysig ad forecastig the dyamics of the exchage rates. 601

6 2.2 Learig Paradigms i (ANNs) The ability to lear is a peculiar feature pertaiig to itelliget systems, biological or otherwise. I artificial systems, learig (or traiig) is viewed as the process of updatig the iteral represetatio of the system i respose to exteral stimuli so that it ca perform a specific task. This icludes modifyig the etwork architecture, which ivolves adjustig the weights of the liks, pruig or creatig some coectio liks, ad/ or chagig the firig rules of the idividual euros [25]. ANN approach learig has demostrated their capability i fiacial modellig ad predictio as the etwork is preseted with traiig examples, similar to the way we lear from experiece. I this paper, a three-layered feed-forward ANN model was structured to predict stock price idex movemet is give i Fig. 2. This ANN model cosists of a iput layer, a hidde layer ad a output layer, each of which is coected to the other. At least oe euro would be employed i each layer of the ANN model. Iputs for the etwork were twelve techical idicators which were represeted by twelve euros i the iput layer (see Table 2). The uits i the etwork are coected i a feed forward maer, from the iput layer to the output layer of the ANN model with coectivity coefficiets (weights), [57]. The weights of coectios have bee give iitial values. The error betwee the predicted output value ad the actual value is back-propagated through the etwork for the updatig of the weights. This method is prove highly successful i traiig of multi-layered eural etworks. The etwork is ot just give reiforcemet for how it is doig o a task. Iformatio about errors is also filtered back through the system ad is used to adjust the coectios betwee the layers, thus improvig performace. This a supervised learig procedure that attempts to miimise the error betwee the desired ad the predicted outputs [58]. If the error of the validatio patters icreases, the etwork teds to be over adapted ad the traiig should be stopped. Fig. 2. A Neural etwork with three-layer feed forward Source: author,

7 The most typical activatio fuctio used i eural etworks is the logistic sigmoid trasfer fuctio. This fuctio coverts a iput value to a output ragig from 0 to 1. The effect of the threshold weights is to shift the curve right or left, thereby makig the output value higher or lower, depedig o the sig of the threshold weight. The output values of the uits are modulated by the coectio weights, either magified if the coectio weight is positive ad greater tha 1.0, or beig dimiished if the coectio weight is betwee 0.0 ad 1.0. If the coectio weight is egative or (value < 0) the tomorrow close price value < tha today s price (loss). While, If the output value is (smaller tha 0.5 or value > = 0) the tomorrow close price value remais same as today s price (o loss); otherwise, it is classified as a icreasig directio i movemet. If (value > 0.5) the the tomorrow close price value > tha today s price (profit). As show i Fig. 2, the data flows from the iput layer through zero, oe, or more succeedig hidde layers ad the to the output layer. The back-propagatio (BP) algorithm is a geeralisatio of the delta rule that works for etworks with hidde layers. It is by far the most popular ad most widely used learig algorithm by ANN researchers [13]. Its popularity is due to its simplicity i desig ad implemetatio. The idea is to trai a etwork by propagatig the output errors backward through the layers. The errors serve to evaluate the derivatives of the error fuctio with respect to the weights, which ca the be adjusted. It ivolves a two stage learig process usig two passes: a forward pass ad a backward pass. The basic back propagatio algorithm cosists of three steps (Fig. 2). Although, the most commercial back propagatio tools provide the most impact o the eural etwork traiig time a performace. The output value for a uit is give by the followig Equatio: y f (h j ) f ( w i, j x i, j ) i 1 1 w ij x i 0 w ij x i (i 1, 2,..., ) (1) where y the output value is computed from set of iput patters, w th x i of ith uit i a previous layer, ij is the weight o the coectio from the euro i to j, j is the threshold value of the threshold fuctio f, ad is the umber of uits i the previous layer. The fuctio ƒ(x) is a sigmoid hyperbolic taget fuctio [59,60]: f ( x ) tah( x ) 1 e z 1 e z 0 threshold : f ( x ) 1 if x 0,1 otherwise (2) where ƒ(x) is the threshold fuctio remais the most commoly applied i ANN models due to the activatio fuctio for time series predictio i back-propagatio: y f w i x i u w i x i i 1 i 1 (3) 603

8 Oce the output has bee calculated, it ca be passed to aother euro (or group of euros) or sampled by the exteral eviromet. I terms of the weight chage, formula equatio is give as: w ij j x i w ij, the (4) where η is the learig rate (0<η<1), δj is the error at euro j, the weights vector. This rule of LMS ca also be rewritte as: x i is a iput vector ad w i w i t i x iw i x i (5) Although a high learig rate, η, will speed up traiig (because of the large step) by chagig the weight vector, w, sigificatly from oe phase to aother. Accordig to [61] suggests that 0.1,1.0, Zupa ad Gasteiger [62] recommed η [63] recommeds η [0.0, 1.0]. [0.3, 0.6], ad Fu 3. EMPIRICAL METHODOLOGY It is difficult to desig artificial eural etwork (ANN) model for a particular forecastig problem. Modellig issues must be cosidered carefully because it affects the performace of a ANN. Oe critical factor is to determie the appropriate architecture, that is, the umber of layers, umber of odes i each layer. Other etwork desig decisios iclude the selectio of activatio fuctios of the hidde ad output odes, the traiig algorithm, ad performace measures. The desig stage ivolves i this study to determie the iput (idepedet) ad output (depedet) layers through the hidde layers i the case where the output layer is kow to forecast future values. Output of the etwork was two patters (0 or 1) of stock price directio. The output layer of the etwork cosisted of oly oe euro that represets the directio of movemet. The umber of euros i the hidde layer was determied empirically. The determiatio of the formulatio betwee iput ad output layers is called learig ad through the learig process, model recogises the patters i the data ad produces estimatios. The architecture of the three-layered feed-forward ANN is explaied i Fig. 2. The etire data set covers the period from Jauary 2, 2007 to December 30, 2010 for etwork traiig ad validatio values, while data from Jauary 2, 2007 to March 28, 2013 to test the predictive ability of the etwork. For this study the artificial eural etworks are capable estimatio models for fiacial ad statistical performace to test the power of eural etwork i the predictio of stock price ad its movemet. This process ca be described below: 3.1 Statistical Performace Evaluatio of the Model I order to estimate the forecastig statistical performace of some methods or to compare several methods we should defie error fuctios. [64] advised to use the followig forecast accuracy measures: Mea Error (ME), Mea Absolute Error (MAE), Mea Squared Error (MSE), Root Mea Squared Error (RMSE), Stadard Deviatio of Errors (SDE), Mea Per cet Error (MPE) ad Mea Absolute Per cet Error (MAPE), etc. I our study we use four performace criteria amely mea absolute error (MAE), root mea square error (RMSE), mea absolute percetage error (MAPE) ad goodess of fit R 2. The back-propagatio 604

9 learig algorithm was used to trai the three-layered feed-forward ANN structure i this study were the most used error fuctios is as followig: The mea absolute error is a average of the absolute errors E ( Pi p i ), where Pi ad p i are the actual (or observed) value ad predicted value, respectively. Lesser values of these measures show more correctly predicted outputs. This follows a log-stadig traditio of usig the ex-post facto perspective i examiig forecast error, where the error of a forecast is evaluated relative to what was subsequetly observed, typically a cesus based bechmark [65,66]. The most commoly used scale-depedet summary measures of forecast accuracy are based o the distributios of absolute errors ( E ) or squared errors 2 (E ) observatios () is the sample volume. The mea absolute error is give by: Mea A bsolute Error ( MA E ) ( E / ) (i 1, 2,..., ) i 1 (6) The MAE is ofte abbreviated as the MAD ( D for deviatio ). Both MSE ad RMSE are itegral compoets i statistical models (e.g., regressio). As such, they are atural measures to use i may forecast error evaluatios that use regressio-based ad statistical. The square root of the mea squared error as follows: Mea Square Error ( MSE ) ( E 2 / ) (i 1, 2,..., ) i 1 Root Mea Square Error ( RMSE ) Sqrt ( E 2 / ) (i 1, 2,..., ) i 1 (7) If the above RMSE is very less sigificat, the predictio accuracy of the ANN model is very close to 100%. Sice percetage errors are ot scale-idepedet, they are used to compare forecast performace across differet data sets of the area usig absolute percetage error give by A PE ( Pi Pi ) *100. Like the scale depedet measures, a positive value of APE is derived by takig its absolute value A PE observatios (). This measure icludes: M ea A bsolute Percetage Error ( M A P E ) ( A PE / ) i 1 (i 1, 2,..., ) (8) The use of absolute values or squared values prevets egative ad positive errors from offsettig each other. However, there are other error measures (e.g. Theil-U or LINEX loss fuctio) but they are less ituitive ad ifrequetly used [67]. All these features ad more make MATLAB a idispesable tool for use i this work. G oodess of Fit ( R 2 ) 1 ( E 2 ) / ( e i2 ) i 1 i 1 ( i 1, 2,..., ) (9) 605

10 where e i p i p i, is the forecast error values. p i, the actual values ad p i, deote the 2 predicted values. The more R correlatio coefficiet gets closer to oe, the more the two data sets are correlated perfectly. As the aim of all of the predictio system models proposed i this study is to predict the directio of the stock price idex forecastig, the correlatio betwee the outputs do ot directly reflect the overall performace of the etwork. 3.2 Fiacial Performace Evaluatio of the Model I order to evaluate the fiacial performace of the model, the correct predicted positios by the model have bee compared. Predictio rate (PR) is evaluated used i the formula to calculate the predictio accuracy ad is as follows: Predictio Rate ( PR ) 1 Ri i 1 (i 1, 2,..., ) (10) where R i the predictio result is for the i tradig day is defied by: th 1 if PO i A O i Ri otherw ise 0 PO i is the predicted output from the model for the ith tradig day, ad A O i is the actual th output for the i tradig day, the total predicted outputs. The error level was determied 5% ad it meas that those outputs with the error level less tha the defied value are cosidered as correctly predicted values. 4. DATA ANALYSIS AND EXPERIMENTS The data used i this study iclude total stock price idex which is composed of closig price, the high price ad the low price of total price idex. It should be remided that the total period of (02/01/ /03/2013) is divided ito two sub-periods of traiig ad validatio period. First sub-periods of (02/01/ /12/2010) is i etwork traiig ad validatio values are obtaied with differet combiatios of parameters for testig the models. The secod sub-period of (02/01/ /03/2013) is i sample period for testig predictio rate model iputs. The whole data i the statistical populatio were employed i the aalysis ad this leads to o-selectio of a specified samplig method. The total umber of sample is 763 tradig days. The umber of sample with icreasig directio is 443 while the umber of sample with decreasig directio is 320. That is, 58% of the all sample have a icreasig directio ad 42% of the all sample have a decreasig directio. The research data used i this study is the directio of daily closig price movemet i the LSM Idex. The umber of sample for each year is show i Table

11 Table 1. The umber of sample i the etire data set Descriptio Icrease % Decrease % Total Year Source: author, Total Sice we attempt to forecast the directio of daily price chage i the LSM idex, techical idicators are used as iput variables i the costructio of predictio models. This study selects 12 techical idicators to make up the iitial attributes, as determied by the review of domai experts alog with the previous studies such as [67,17,41,16,25,2,42,55]. Table 2 demostrates the titles of twelve techical idicators ad their formulas, ad summary statistics data for the selected idicators were calculated ad give i Table DESCRIPTIVE STATISTICS As metioed previous i sub-sectio 2.2 ad Fig. 2, the fuctio of hidde layer is tasigmoid ad the trasferred fuctio of output layer is liear i a three layer etwork, where iput layer is simply distributig the iputs i various hidde layer ad o processig takes place there, requires least umber of traiig epochs. The parameters of eural etwork model iclude the umber of euros () i the hidde layer, value of learig rate (η), mometum coefficiet (µ) ad umber of traiig epochs (ep) are professioally determie with ANN model parameters usig eural etworks toolbox of MATLAB software to implemet the model. As suggested i the previous sectio 2, literature, a small value of η was selected as 0.1. The levels of the ANN parameters that are tested for choosig the best combiatio is preseted i Table 3. Te levels of euros (), five levels of mometum (µ) ad 120 levels of epochs (ep) were tested i the parameter settig experimets. The parameter levels evaluated i parameter settig a total umber of (10 * 5 * 120 = 6000) treatmets for ANN model. Each parameter combiatio was applied to the traiig ad validatio data sets ad predictio accuracy of the models were evaluated. Therefore, the traiig ad validatio performace were calculated for each parameter combiatio. The parameter combiatio that resulted i the best average of traiig ad validatio performaces was selected as the best oe for the predictio model. 607

12 Table 2. Selected techical idicators ad their formulas Defied variables Accumulatio/distributio oscillator. It is a mometum idicator that associates chages i price It measures the variatio of a security s price from its statistical mea It is a mometum idicator that measures overbought/oversold levels Movig average covergece divergece Code A/D Oscillator Formula equatio Descriptio H t C t 1 H t Lt where C t is the closig price at time t, Lt the low price at time t, H t the high price at time t w here M t ( H t L t C t ) / 3, CCI Commodity chael idex Larry William s (R%) MACD SM ( M t SM t ) 0.015Dt H Ct *100 H L t Dt i 1 i 1 M t i 1 M t i 1, ad SM t MA CD ( )t 1 2 / 1* where DIFF : EMA (12)t EMA (26)t, EMA is ( DIFFt MA CD ( )t 1 ) expoetial movig average, EMA ( k )t : EMA ( k )t 1 * (C t EMA ( k )t 1 ), smoothig factor: 2 / (1 k ), k is time period of k day expoetial It measures the amout that a security s price has chaged over a give time spa movig average. Mometum Ct Ct price day where C t is the closig price at time t, the 608

13 It displays the differece betwee the curret price ad the price days ago Relative stregth idex. It is a price followig a oscillator that rages from 0 to 100. A method for aalysig RSI is to look for divergece i which the security is makig a ew high. ROC Pricerate-ofchage RSI movig Simple MA It compares where a security s price closed relative to its price rage over a give time period Stochastic (K %) Simple average. 10-day Movig average of %K Ct *100 Ct ( i 0Upt 1 / ) / ( i 0 Dwt i / ) C t Lt *100 HH t LLt where Upt meas Dw upward-price- t meas dowwardchage ad price-chage at time t. It shows the average value of a security s price over a period of time. If the value of a security s price over a period of time. If the price moves above its MA, a buy sigal is geerated. If the price moves below its MA a sell sigal is geerated. where LLt ad HH t, mea lowest low ad highest high i the last t days, respectively. ( i 0 K t i %) 1 Stochastic (D%) Movig average of %D. Stochastic slow (D%) Weighted average WMA movig 1 C t C t 1... C t ( i 0 Dt i %) 1 10-day (( ) * ct ( 1) * ct 1... c ( ( 1)... 1) Notes: I this study the origial data were ormalised i a rage of [-1, 1]. 609

14 Table 3. ANN parameter levels tested i parameter settig Parameters Number of euros () Learig rate (η) Mometum costat (µ) Epochs (ep) Level(s) Ta-Sigmoid fuctio trasfer Liear fuctio trasfer 10, 20,..., ,0.2,..., , 0.02,, , 60, 90,120 Source: author calculatio, 2013 Table 4 presets the summary statistics for each attribute. This study is to predict the directios of daily chage of the LSM idex is categorised as 0 or 1. 0 meas that the ext day s LSM idex at time t is lower tha today s LSM idex at time t-1. If the LSM idex at time t is higher tha that at time t-1, directio t is 1. The origial data were scaled ito the rage of [-1:0; 1:0] usig max-mi ormalisatio formula, which is used here to expresses the actual value usig the followig equatio. u ( x i - x i,mi ) ( hi l i ) l i ( x i,max - x i,mi ) (i 1, 2,..., ) where u ad x i represet ormalised ad actual value respectively. (11) x i,mi ad x i,max represet miimum ad maximum values of the attribute x i. hi upper boud of the ormalisig iterval ad l i lower boud of the ormalisig iterval. Max-mi ormalisatio plas a value u of x i i the rage ( hi l i ) i.e. (-1.0; 1.0), i this case. As a value greater tha 0 represets a buy sigal while a value less tha 0 represets a sell sigal. (i 1, 2,..., ) the umber of observatios. The aim of liear scalig is to idepedetly ormalise each feature compoet to the specified choice. It esures the larger value iput attributes do ot overwhelm smaller value iputs, the helps to reduce predictio errors [16,18]. As we metioed earlier i Table 2, twelve techical idicators are as iput variables. The Mea ad Stadard Deviatio of iput variables is show i Table

15 Table 4. ANN parameter levels tested i parameter settig Features ame A/D Oscillator CCI Larry William s (R%) MACD Mometum ROC RSI Simple MA (K %) (D%) Slow (D%) WMA Max Mi Mea Source: author calculatio, 2013 Stadard deviatio EXPERIMENTAL RESULTS 6.1 Compariso of Fiacial Performace The total period of (02/01/ /12/2010) is i etwork traiig ad validatio values are obtaied with differet combiatios of parameters for testig the ANN model. The empirical results are preseted i Table 5 ad 6. The secod sub-period of (02/01/ /03/2013) is i sample period separately as a model iput is used ad predictio rate is calculated. The total of 6000 parameter combiatios for the ANN model were tested ad completed as preseted i Table 3. Three parameters are cosidered as the best combiatios ad correspodig predictio accuracies are give i Table 5. Through these parameter combiatios, we ca ow able to reach compariso experimets of the ANN model, based o the data sets preseted i Table 5, icludig the average of traiig ad validatio performaces for each case is calculated. The average of traiig ad validatio performace of the ANN model for these parameter combiatios was varied betwee 87.58% ad 87.99%. It ca be assumed that both the traiig ad validatio performaces of the ANN model are sigificat for parameter combiatio settig data set. The experimets were carried out for each year ad the aalysis results which are revealed separately i Table 6. Table 5. Three parameters are cosidered as the best combiatios of ANN model No η ep µ Traiig Source: author calculatio, 2013 Validatio Average Table 6 shows that the average predicatio rate values (91%) as the measure of fiacial performace of the ANN model for three differet parameter combiatio (0.1; 120; 0.019; 30) is relatively better tha others. Therefore, the predictio rate values performace of this parameter combiatio ca be adopted as the best of the ANN model, sice its average traiig ad validatio performace (82.56%; 93.42%) is relatively greater tha the others. 611

16 As show i Table 6, the best adaptatio of the eural etwork model outputs with predictio rate (PR) values is 91%, which meas 91% of data aalysis is correctly predicted by the ANN model. Table 6 also shows that the predictio rate values performaces are differet for each year. For the selected parameter combiatio, the best predictio rate values rate performace (92%) was obtaied i 2009 ad 2010, while the worst oe (88%) was obtaied i For the other parameter combiatios, the predictio rate values performaces i 2013 were geerally lesser tha the other years. However, therefore, based o the experimetal results give, the best parameter combiatio of ANN model is (η = 0.1; ep = 120; µ = 0.019; = 30) with a average predictio rate values performace (91%). Table 6. Fiacial predictio performace (%) of ANN model Parameter combiatio (η ep µ ) (0.023; 60; 0.037; 20) (Traiig; validatio) Year (80.49; 94.67) Predictio rate Average 0.89 (0.1; 120; 0.019; 30) (Traiig; validatio) (82.56; 93.42) Predictio rate (0.025; 90; 0.041; 30) (Traiig; validatio) (81.99; 93.51) Predictio rate Note: The Libya Stock Market closed followig the eruptio of the Libya civil war i 15 February It remaied closed util reopeig the followig year o 15 March Source: author calculatio, Compariso of Statistical Performace Statistical performace of the three parameter combiatio is compared i Table 6. As 2 already is metioed i sub-sectio 3.1, MAE, RMSE, MAPE ad R, measures are used i order to compare the statistical performace of parameters combiatios. MAE, MAPE ad RMSE are used as error measuremet. Table 7 shows that the error measuremets are used for ANN model i order to compare the statistical performace of parameters 2 combiatios. Goodess of fit R is also referred to as the coefficiet of multiple correlatios. As is show i Table 8, the parameter combiatio (0.1; 120; 0.019; 30) is relatively better tha others. Therefore, this combiatio i terms of fiacial ad statistical performaces is the best oe. I all cases the relatioship stregth betwee parameter combiatio ad 2 forecast accuracy measures such as MAE, MAPE, ad RMSE is strog (R 0.99). MAPE ad RMSE measure the residual errors, which gives a global idea of the differece betwee the predicted ad actual values. Although, the MAE is very similar to the RMSE but it is less sesitive to large forecast errors (Fig. 3). The loger MAE meas higher bias level ad less accurate forecast to predict prices, but it does ot mea that MAE is ot suitable to predict stock market fluctuatios. It seems that MAE method is more suitable to predict stock market fluctuatios rather tha short movig average. Accordig to Table 8, as good as ANN model ca be, is a powerful tool i predictig directio of LSM idex movemet ad the curret study results is i cosistet with the previous studies such as [17,41,42,68,55]. 612

17 Table 7. Statistical performace of ANN model Parameter combiatio (η ep µ ) (0.023; 60; 0.037; 20) Year R2 MAE MAPE E E E E RMSE E E-06 7E (0.1; 120; 0.019; 30) R2 MAE MAPE E E E E E E RMSE 2E-05 4E-07 1E-05 3E E-05 Source: author calculatio, 2013 (0.025; 90; 0.041; 30) R2 MAE MAPE E E E RMSE 6.31E E

18 Table 8. Statistical performace forecastig results of ANN model Model ANN No of Obs. 6 2 R PR 0.91 MAE Source: author calculatio, 2013 RMSE MAPE Fig. 3 shows that the MAE, RMSE ad MAPE calculated for the forecast period usig the empirical atiracial atural etwork model i predictig directio of LSM idex movemet results are give i Table 8. The ANN models correctly predict the sigs of stock price idexes up to 91% for etry data. Fig. 3. MAE, RMSE ad MAPE of eural etwork model Source: author calculatio from Table 8, 2013 The best results of the ANN model i predictig directio of LSM idex is based i the paired samples T-test. The experimetal results of T-test are give i Table 9. Table 9 shows that the mea performaces of ANN model is sigificat level at α = 0.05, which give us some very importat clues. That is, the performace of the ANN shows superior predictig power i forecastig the LSM idex movemet. Table 9. T-test results of fiacial predictio ANN model Model ANN No of Obs. 6 Mea 91% Std.dev 3.28 Source: author calculatio, 2013 t p The ANN model accurately predicted the directio of movemet with predictio rate 91% of data aalysis i its best case, which is a perfectly good outcome. I additio, the detailed empirical aalysis of models parameters ad selectio of efficiet parameter values may result i higher predictio accuracy. The set of techical idicators aalysis adopted i our ANN models cosidered as the most appropriate for stock price predictio i emergig ecoomy such as Libya Stock Market. Although the predictio performace of ANN outperforms studies i the review of literature, it is still likely that the statistical forecast performace of the model is still possible to be improved performace over traditioal techiques by doig the followig: the model parameters should be adjusted by systematic experimetatio comprehesive or the iput variable sets requisite to to be modified by choosig those iput that are more realistic i reflectig of the fiacial market performace. 614

19 7. CONCLUSION AND FUTURE RESEARCH This paper aims to fid the aswer of the followig questio: whether the Libya forecasted LSM idex through the learig procedure techiques of ANN model or ot. The issue of accurately predictig the directio of movemets of the stock market price levels is highly sigificat for formulatig the best market tradig solutios. It is fudametally affectig fiacial trader s decisios to buy or sell of a istrumet that ca be lucrative for ivestors. It ca be cocluded that successful predictio of stock prices may promise attractive beefits for ivestors. LSM idex behaviour, however, is extremely complicated ad very difficult. I additio, stock market is ca be affected by may macro-ecoomic factors such as political evets, ivestors expectatios, istitutioal ivestors choices, firms policies, geeral ecoomic coditios, iterest rates, foreig exchage rates, movemet of other stock market, psychology of ivestors ad cosumer price idex etc. Cosequetly, it is very importat to desig ad develop a model with the capability i predictig the LSM idex behaviour which lears from observed data. This study attempted to predict the directio of stock price movemet i emergig market such as the Libya Stock Market closig price levels usig ANN model based o the daily data from period 2007 to The forecastig 2 ability of the model is accessed usig MAE, RMSE, MAPE ad R, which ca be a future work for iterested readers. The key experimetal results obtaied, ca give some very sigificace coclusios of this study is as follows. Firstly, it shows that how forecastig the stock market price could be achieved by usig the proposed model. This model of ANN showed sigificat performace i predictig the directio of stock price movemet. Thus, ANN model ca be used as a better alterative techique for forecastig the daily stock market prices for this area. The ANN model accurately predicted the directio of movemet with the average predictio rate 91% of data aalysis i its best case, which is a perfectly good outcome. I all cases the relatioship stregth betwee parameter combiatio ad forecast accuracy measures such 2 as MAE, MAPE, ad RMSE is strog (R 0.99). Furthermore, this study proved the sigificace of usig twelve particular techical market idicators which gave also useful results i predictig the directio of stock price movemet. To improve ANN model capabilities, a mixture of techical ad fudametal factors as iputs over differet time period were used to be a effective tool i forecastig the market level ad directio. The limitatios of curret study that provide some suggestios for future research are metioed i the followig: (1) each method has its ow stregths ad weakesses. Future researches are suggested to use techical idicators of this study ad other combiig techiques models by itegratig ANN with other classificatio models such as radom walk [see, e.g., 69], Support Vector Machies SVM, Geetics Algorithm GA ad Geeralised Autoregressive Coditioal Heteroskedasticity GARCH models to predict the LSM idex movemets. The weakess of oe method ca be balaced by the stregths of aother by achievig a systematic effect. (2) Hybrid ANN model ca also be used i the predictio of stock price idexes. Thus, these models grew to iclude: (SVM), (GA), (GARCH) ad fiacial time series predictio. (3) Aother importat issue that should be metioed here is the differeces amog the predictio performaces for each year. It ca be show from the experimetal results that the year of 2011 is ot icluded because Libya had a devastatig political crisis i The crisis i the revaluatio had affected the stock market iitially. The stock market has bee closed for 12-moths. Uder such circumstaces of political crisis, a decrease i the predictio performace of techical idicators ca be cosidered acceptable. 615

20 COMPETING INTERESTS Author has declared that o competig iterests exist. REFERENCES Abu-Mostafa YS, Atiya AF. Itroductio to fiacial forecastig. Applied Itelligece. 1996;6(3): Maish K, Themozhi M. Support vector machies approach to predict the S&P CNX NIFTY idex returs. I Proceedigs of 10th Idia istitute of capital markets coferece, Mumbai, Idia; Accessed 12 March Available: Fama E. Efficiet capital markets: a review of theory ad empirical work. Joural of Fiace. 1970;25: Hecht-Nielse R. Applicatios of couter propagatio etworks. Neural Networks. 1988;1: Schalkoff RJ. Artificial eural etworks. McGraw-Hill: New York; Hawley DD, Johso JD, Raia D. Artificial eural systems: a ew tool for fiacial decisio-makig. Fiacial Aalysis Joural. 1990; Shtub A, Versao R. Estimatig the cost of steel pipe bedig, a compariso betwee eural etworks ad regressio aalysis. Iteratioal Joural of Productio Ecoomics. 1999;62(3): Garcia R, Gecay R. Pricig ad hedgig derivative securities with eural etworks ad a homogeeity hit. Joural of Ecoometrics. 2000;94(1-2): Hamm L, Brorse BW. Tradig futures markets based o sigals from a eural etwork. Applied Ecoomics Letters. 2000;7(2): Lapedes A, Farber R. Noliear sigal processig usig eural etworks. IEEE coferece o Neural Iformatio processig systems- Natural ad sythetic; Sharda R, Patil R. Neural etworks as forecastig experts: a empirical test. Proceedigs of the 1990 Iteratioal Joit Coferece o Neural Networks, Washigto DC: USA. 1990;1: Tag Z, Fishwick PA. Feed forward eural ets as models for time series forecastig. ORSA Joural o Computig. 1993;5(4): Bishop C. Neural etworks for patter recogitio. Oxford Uiversity Press: Oxford; Dacha K. Causal modelig of stock market prices usig eural etworks ad multiple regressio: a compariso report. Fiace Idia. 2007;(3): Tag A, Fishwick S. Time series forecastig usig eural etworks vs. Box-Jekis methodology. Simulatio. 1991;57(5): Kim K. Fiacial time series forecastig usig support vector machies. Neurocomputig. 2003;55: Kim K, Ha I. Geetic algorithms approach to feature discretizatio i artificial eural etworks for the predictio of stock price idex. Expert Systems with Applicatios. 2000;19: Maish K, Themozhi M. Forecastig stock idex movemet: a compariso of support vector machies ad radom forest. I Proceedigs of ith Idia istitute of capital markets coferece, Mumbai, Idia; Accessed 12 March Available: ssr.com/abstract= Libya Stock Market Reports; Accessed 2 Jauary Available: 616

21 Hertz J, Krogh A, Palmer RG. Itroductio to the theory of eural computatio. Redwood City: Addiso Wesley; Che AS, Leug MT, Daouk H. Applicatio of eural etworks to a emergig fiacial market: forecastig ad tradig the Taiwa Stock Idex. Computers ad Operatios Research. 2003;30(6): Pesara MH, Timerma MA. Recursive modellig approach to predictig stock returs. Uiversity of Cambridge Press: Cambridge; Aderso JA, Rosefeld E. Neurocomputig: foudatios of research. MIT Press: Cambridge, MA; Rumelhard DE, Mcclellad JL. Parallel Distributed Processig - Exploratios i the Micro-structure of Cogitio. Foudatios. 1986;1:294. Armao G, Marchesi M, Murru AA. Hybrid geetic-eural architecture for stock idexes forecastig. Iformatio Scieces. 2005;170:3-33. Chiag WC, Urba NL, Baldridge GW. A eural etwork approach to mutual fud et asset value forecastig. Omega, The Iteratioal Joural of Maagemet Sciece. 1996;215(2): Basheer IA, Hajmeer M. Artificial eural etworks: fudametals, computig, desig ad applicatio. Joural of Microbiological Methods. 2000;34:3-31. Breedlove SM, Watso NV, Rosezweig MR. Biological Psychology: A Itroductio to Behavioral, Cogitive, ad Cliical Neurosciece. 6th ed. Siauer Associates, Ic. USA; Jai AK, Mao J, Mohiuddi KM. Artificial eural etworks: a tutorial. Comput. IEEE. 1996; Zhou WJ, Jie WJ, George ZZ, Po Guo S. Forecastig stock idices with back propagatio eural etwork. Expert Systems with Applicatios. 2011;37: Rasbad MN. The axo iitial segmet ad the maiteace of euroal polarity. Nature Reviews Neurosciece. 2010;11: Silverthor DU. Huma Physiology: a Itegrated Approach. Pearso. ISBN :253. Brow ad Bechmark. Itroductory Psychology Electroic Image Bak. Times Mirror Higher Educatio Group Ic; Bieres HJ. Commet o artificial eural etworks: a ecoometric perspective. Ecoometric Reviews. 1994;13(1): Kua, C.M. ad White, H. Artificial eural etworks: a ecoometric perspective. Ecoometric Views. 1994;13(1):1-91. Kaastra I, Boyd MS. Forecastig futures tradig volume usig eural etworks. Joural of Futures Markets. 1995;15(8): Trippi RR, DeSieo D. Tradig equity idex futures with a eural etwork. Joural of Portfolio Maagemet. 1992;19(1): Witkowska D. Neural etworks as a forecastig istrumet for the Polish Stock Exchage. Iteratioal Advaces i Ecoomic Research. 1995;1(3): Harvey CR. Predictable risk ad returs i emergig markets. The Review of Fiacial Studies. 1995;8: Yoo Y, Swales G. Predictig stock price performace. Proceedig of the 24th Hawaii Iteratioal Coferece o System Scieces. 1997;4: Diler AI. Predictig directio of ISE atioal-100 idex with back propagatio traied eural etwork. Joural of Istabul Stock Exchage. 2003;7(25-26): Altay E, Satma MH. Stock market forecastig: artificial eural etworks ad liear regressio compariso i a emergig market. Joural of Fiacial Maagemet ad Aalysis. 2005;18(2):