TOWARDS ENHANCING STOCK MARKET WATCHING BASED ON NEURAL NETWORK PREDICTIONS

Size: px

Start display at page:

Download "TOWARDS ENHANCING STOCK MARKET WATCHING BASED ON NEURAL NETWORK PREDICTIONS"

Hubert Copeland
5 years ago
Views:

1 TOWARDS ENHANCING STOCK MARKET WATCHING BASED ON NEURAL NETWORK PREDICTIONS 1 Mohammed Awad, 2 Aseel Kmail 1 Departmet of Computer Systems Egieerig, Egieerig ad Iformatio Techology, Arab America Uiversity, Palestie 1 Departmet of Computer Sciece, Egieerig ad Iformatio Techology, Arab America Uiversity, Palestie 1 mohammed.awad@aauj.edu, 2 aseel.kmail@aauj.edu Abstract Uder the growth of the stock market sector ad the widespread of stock market applicatios, the stock market predictio has become oe of the most importat ad challegig tasks i the stock market. May data miig techiques are exploited to predict the stock prices i order to help ivestors i makig ivestig decisios. Oe of the most commo ad widely used techiques is Artificial Neural Networks (ANN). I this paper, we aim to preset a model for stock market predictio based o artificial eural etworks. This model uses the variables of techical aalysis of stock market idicators for predictig stock market prices. The proposed model is tested ad evaluated usig Palestie Exchage Tradig (PEX) data, the experimetal validatios show satisfactory results that help ivestors ad traders to make qualitative decisios. The proposed model employs a adaptive process for optimizig eural etwork weights based o back-propagatio learig strategy. The proposed model improves the effectiveess of forecastig the stock prices of Palestie Exchage Tradig (PEX). Keywords: Predictio, Stock Market, Neural Networks INTRODUCTION Usig forecastig techiques ca obtai better results agaist the eed of people to reduce the risk i decisio-makig ad risk aversio as to the choices they have to make. For a log time, people have sought greater access to iformatio, eablig them to make decisios i a correct way, where the possibilities of "mistakes" are the miimum ad success i decisiomakig is as high as possible [1, 2]. Recetly, the stock market predictio has become oe of the most importat issues i the fiacial market. This is because the stock prices predictio helps ivestors ad traders i makig correct qualitative decisios. The stock market ca be defied as a public market for tradig compay's stock at a approved stock price. Palestie Exchage (PEX) is the formal stock market at Palestie, which is oe of the most importat compaies i the ecoomic sector. Ivestig i PEX is doe by either to be a parter i a compay posted i PEX or buyig bods ad gettig a periodical retur o ivestmet (ROI). The mai motivatio of this proposed model is the possibility of makig profits by ivestig i the Palestie stock market without risk too high, usig a eural etwork model. It aims to study the ability of eural etworks i predictig stock, ad check if beefits are obtaied if ivestmet decisios are made based o these. Neural etworks have prove effectiveess i other disciplies ad there are reasos to thik that it ca be successfully applied to the predictio of fiacial series. I geeral, it iteds to make ivestmet decisios based o predictios usig eural etworks, which ifluece the profit. Furthermore, predictios of iformatio o loger are useful iformatio that ca be used to supplemet the isight ad market kowledge. Based o discussio with PEX clerks ad may ivestors there is o predictio system for forecastig the stock prices of compaies i Palestie. Our goal is to start buildig a system that provides its users with the ability to predict 1

2 the stock market prices based o ANN. Though stock market field suffers from upredictability [3, 4], very high umber of research papers have bee performed to produce efficiet techiques ad approaches that ehace the predictio process. These techiques iclude ANN [5], fuzzy time series predictios [6], data miig [7] ad Hidde Markov models [8]. These itelliget systems were ot used before i predictig the stock prices i the Palestiia stock market. As we use a method that depeds o eural etworks, we preset some of the ew methods ad techiques that used eural etworks i stock market predictio; as. I [5] the author preset a hybridized method that combies the techical ad fudametal aalysis variables of the stock market as a idicator for predictig the future price of the stock. The author i [9] presets a approach depedig o the basis of filter ad fuctio based clusterig; the importat features i risk ad retur predictio are selected the risk ad retur re-predicted. The methodology proposed i [10] use a multilayer perceptio eural etwork to derive the relatioship betwee variables use as idepedet factors ad the level of stock price idex as a depedet elemet. A ew itelliget model i a multi-aget framework called bat-eural etwork multi-aget system (BNNMAS) to predict stock price was proposed. The model performs i a four layer multiaget framework to predict eight years of DAX stock price i quarterly periods [11]. The study i [13] compares four predictio models, Artificial Neural Network (ANN), Support Vector Machie (SVM), radom forest ad aive-bayes with two approaches for iput to these models applied o Idia stock markets. I this paper, we used differet cofiguratios of multilayer perceptro eural etworks. These cofiguratios rely o backpropagatio learig method applied o Palestie Exchage Tradig. This model start specifyig the problem iputs/outputs, ad start from simplest structure of eural etwork, the by tray ad error we select the suitable activatio fuctio for the hidde layer i the eural etwork, the weights optimized by usig backpropagatio learig, the best cofiguratio is selected by test ad geeralize the etwork. The orgaizatio of the rest of this paper is as follows: Sectio 2 presets the proposed model. Sectio 3, shows some results that cofirm the performace of the proposed method. Some coclusios will be preseted i Sectio THE PROPOSED NNS MODEL The mai motivatio of the proposed model is the possibility of makig profits by ivestig i the Palestie stock market without a high risk, usig artificial eural etworks. The proposed model aims to use the ability of eural etworks i predictig, ad check the beefits obtaied if the ivestmet decisios are made based o the suggested model. Neural etworks have prove a effective method i may disciplies [9] ad there are may reasos to thik that it ca be successfully applied o the predictio of fiacial series. It will be used to ited to make ivestmet decisios based o predictios, which ifluece the profit [14]. Artificial Neural Networks (ANN), which emulate biological eural etworks, ad They have bee used to lear solvig strategies based o examples of typical behavior patters; these systems do ot require that the task is scheduled to ru, they geeralize ad lear from experiece. The theory of NN has provided a alterative to classical computig systems, for those problems where traditioal methods have delivered ucovicig, or icoveiet results especially i oliear behavior of real systems [15]. Neuro is a processig uit that maps a iput sigal to a output, itegrated with other euros o the same layer of the eural etwork. Multi-layer Feed- forward with backpropagatio Neural etworks (MFFNNBP), is multi-layer perceptro eural etwork that passes i iputs ad their weights from oe layer to the ext oe through feed forward process, ad the it perform the weights update to be back-propagated to the previous layers i order to recalculate ad measure the update coditios like (certai error value, or umber of iteratios [15]. Fig. 1. MLP Neural Network 2

3 Perform a mappig process betwee iputs/output data called traiig of the eural etwork, the output of the simple eural etwork is give by the followig expressio: m Y F( w. X bias ) (1) i ji i i j 1 Where w ij is the weights coectio ad X j are the value of the i th iputs for simple form of the NN, bias i is the NN bias, m is the umber of euros i each hidde layer ad F is the activatio fuctio. To predict the future output depeds o curret ad previous output the model use the error result, which comparig the actual output of the NN with the desired output i the learig process, this error is calculated usig the followig expressio: Error Yid Yia (2) The traiig process cotiues to adjust the weights util the error criteria is satisfied the weight update is performed by the equatio 2: w 1. Error. xi i (3) Where α is the learig rate. I Multilayer Perceptro Neural Networks (MLPNN), the output of a layer will be a iput for the ext layer passig from the iput layer to the output layer; the equatios used for this procedure are illustrated as follows: 2 output f ( out1. wjk ) (4) j 1 Where the output of the first hidde layer out 1, which calculated usig the followig expressio: 1 out1 f ( xi. wij ) (5) j 1 Where f 1 ad f 2 are the activatio fuctios for output layer ad hidde layer, which calculated as i the followig expressios: f 1 1 f 2 1 x e (6) x (7) Where, x = iput vector. Depedig o equatios above, the weights are updated use as the followig expressio: de( wjk ) wjk (8) dwjk Where µ is the learig rate (ormally betwee 0 ad 1). The fial output depeds o all earlier layer's output, weights, ad the algorithm of learig used [15]. The backpropagatio process calculates the gradiet decet error betwee the desired ad the predicted output cosiderig the ew weights each time, this gradiet is usually use i a simple stochastic gradiet descet algorithm to fid the weights that miimize the error. The geeral steps of the eural etwork traiig appear i the followig pseudocode: Iitialize etwork weights (ofte-small radom values) For Each traiig example fid: Real Output of NN Error (Real - Target) at the output uits Compute wi 1 for all weights from hidde layer to output layer Compute wi 1 for all weights from iput layer to hidde layer Update etwork weights depeds o the error value. Retur the process util the termiatio coditio satisfied. From the above algorithm that the backpropagatio process calculates the gradiet decet error betwee the desired ad the predicted output cosiderig the ew weights each time, this gradiet is almost always used i a simple stochastic gradiet descet algorithm to fid weights that miimize the error [15]. The architecture of the proposed model for predictig the prices of the stock market based o exploitig artificial eural etworks (ANN). As show i figure 1, the proposed eural etwork cosists of three mai layers [5]: Iput layer: it cosists of N uits, such that each uit X i [X 1, X 2,,,,,X ]. Hidde layer: it cosists of P processig uits, such that each uit K m [K 1, K 2,,,,, K m ]. Output layer: it fids the output Y accordig to the followig equatio: 3

activatio fuctios for hidde processig elemets ad output [5]. Fig.

4 P N Y ActivF( w. ActivF w. x )) (9) md im i m 1 i 1 Where, w im is the i th coectio weights betwee iput uits ad hidde processig uits, w md is the coectio weights betwee hidde processig uits ad the output uit, ActivF is the activatio fuctios for hidde processig elemets ad output [5]. Fig.2 Neural Networks System Architecture For our model, we defied the followig four techical variables as iputs: The high price for day i-1 The low price for day i-1 The close price for day i-1 The tradig volume for day i-1 The etwork aims to fid the close price for day i from the data of day i If the etwork still does't perform well, go to step 3 ad try aother activatio fuctio. 10. If the etwork still does't perform well, go to step 2 ad try aother form of the etwork. EXPERIMENTAL RESULTS AND DISCUSSION I this sectio, we detail ad discuss experimets that have bee istatiated to validate our proposal. The proposed model is tested usig Matlab Neural Network Tools Box versio 7. 1 o a pc with dual-core CPU (2.4 GHz) ad 4 GB RAM. The operatig system is Widows 7. I order to carry experimets, we used differet eural etwork cofiguratios such as 4-1-1, 4-3-1, 4-5-1, 4-7-1, ad Our eural etwork model has the form of feed-forward multi-layer perceptro eural etwork that is traied with backpropagatio algorithm. The traiig ad testig data were selected carefully from Palestie Exchage website. The output of our eural etwork model was evaluated by comparig the predicted close prices with actual close prices. I geeral, the algorithm for ANN works as i the followig steps: 1. Uderstad ad specify the problem i terms of iputs ad required outputs. 2. Choose the simplest form of etwork to solve the problem. 3. Choose appropriate activatio fuctio. 4. Choose a suitable etwork structure. 5. Fid the appropriate coectio weights, so that the etwork produces the correct output for each traiig data. 6. Test the etwork geeralizatio ad evaluate it usig ew traiig data. 7. If the etwork does't perform well, go back to step 5 ad try harder. 8. If the etwork still does't perform well, go to step 4 ad try harder. Fig.3. Network predictio for cofiguratio After several experimets of differet etwork cofiguratios, we foud that the most accurate predictio was produced by cofiguratio. Table 1 shows a compariso betwee differet etwork cofiguratios accordig to their level of accuracy. Figure 3-7 depicts the correlatio level of accuracy by comparig the actual stock prices with the predicted values. The eural etwork structure that gives the best results is

Table 1: Sample of experimetal results for differet ANN cofiguratio Sample period Actual value Predicted values with differet ANN cofiguratios 4-1-1 4-3-1 4-5-1 4-7-1 4-10-1 4/5/2014 5.55 5.533 5.

5 Table 1: Sample of experimetal results for differet ANN cofiguratio Sample period Actual value Predicted values with differet ANN cofiguratios /5/ /5/ /5/ /5/ /5/ /5/ /5/ /5/ /5/ From the result i table ad the figures, we ca show the possessig ability of NN to predict, the best results are obtaied with the proposed model cofiguratio as show i figure 1. It is clear that the predictio is beig made with cosiderably good accuracy, but the eural etwork cofiguratio of produce a higher predictig error tha other cofiguratios, these situatios ca occur whe the value of a stock ca be iflueced by multiple factors that were ot take ito accout i modelig. I this case, there could have bee a uusual behavior. Fig.6. Network predictio for cofiguratio Fig.4. Network predictio for cofiguratio. Fig.7. Network predictio for cofiguratio Fig.5. Network predictio for cofiguratio The proposed model, for the period ad the target market, produce better. Moreover, the model developed optimizatio protocol allows the user to determie their ow level of risk, makig this work applicable to a wide rage of ivestors. 5

6 CONCLUSION Artificial eural etworks have a practice applicatio i the stock market better tha traditioal statistics models because they deped o theoretical assumptios o which are based o the statistics techiques. We preseted i this paper a model for stock market predictio based o artificial eural etworks. The proposed model is created to help ivestors ad traders at Palestie exchage to predict the stock prices for specific compaies i order to make suitable qualitative ivestig decisios. The proposed model was tested ad evaluated usig real-world tradig data. The empirical study compared actual data with predicted data, ad the produced results showed a high level of accuracy i predictio, which is useful i makig decisios. I the future, we aim to icrease the umber of techical parameters that have bee used for predictio ad the test the impact of the ewly added techical parameters o the accuracy of predictio. The proposed model ca be used as a itegral part fiacial operatio as it is used ANN tools to predict applicatio i fiacial decisio makig i the treasury maagemet ad fiacial risk maagemet i Palestie exchage. REFERENCES [1] Araújo, R. d. A., & Ferreira, T. A. A morphological-rak-liear evolutioary method for stock market predictio. Iformatio Scieces, vol. 237, pp [2] W.W.Y. Ng, X-L. Liag, J. Li, D.S. Yeug, P.P.K. Cha, LG-Trader: Stock Tradig Decisio Support based o Feature Selectio by Weighted Localized Geeralizatio Error Model. Neurocomputig, vol. 146, pp [3] M.T. Philip, K. Paul, S.O. Choy, K. Reggie, S.C. Ng, J. Mak, T. Joatha, K. Kai, ad W. Tak-Lam, Desig ad Implemetatio of NN5 for Hog Stock Price Forecastig. Joural of Egieerig Applicatios of Artificial Itelligece, vol. 20, pp [4] T. H. Roh, Forecastig the Volatility of Stock Price Idex. Joural of Expert Systems with Applicatios, vol. 33, pp [5] Adebiyi Ayodele A., Ayo Charles K., Adebiyi Mario O., ad Otokiti Suday O, Stock Price Predictio usig Neural Network with Hybridized Market Idicators. Joural of Emergig Treds i Computig ad Iformatio Scieces. vol. 3, o. 1 pp [6] Tiffay Hui-Kuag yu ad Ku-Huag Huarg, A Neural etwork-based fuzzy time series model to improve forecastig. Expert Systems with Applicatios, vol. 37 o. 4, pp: [7] K. Sethamarai Kaa, P. Sailapathi Sekar, M.Mohamed Sathik ad P. Arumugam, Fiacial stock market forecast usig data miig Techiques". Proceedigs of the iteratioal multicoferece of egieers ad computer scietist s vol. 1, [8] Md. Rafiul Hassa ad Baikuth Nath, Stock Market forecastig usig Hidde Markov Model: A New Approach. I Itelliget Systems Desig ad Applicatios, ISDA'05. Proceedigs. 5th Iteratioal Coferece, pp IEEE [9] Sasa Barak, Mohammad Modarres, Developig a approach to evaluate stocks by forecastig effective features with data miig methods. Expert Systems with Applicatios, vol. 42, o. 3, pp [10] A. Ghezelbash, F. Keyia, Desig ad Implemetatio of Artificial Neural Network System for Stock Exchage Predictio. Africa Joural of Computig & ICT, vol. 7, o. 1, pp [11] R. Hafezi, J. Shahrabi, E.Hadavadi, A bat-eural etwork multi-aget system (BNNMAS) for stock price predictio: Case study of DAX stock price, Applied Soft Computig, vol. 29, pp , [12] J. Patel, S. Shah, P. Thakkar, K. Kotecha, Predictig stock ad stock price idex movemet usig tred determiistic data preparatio ad machie learig techiques. Expert Systems with Applicatios, vol. 42 pp [13] J.T. Yao, C.L. Ta, ad H.L Poh, Neural etworks fortechical Aalysis: A study o KLCI, Iteratioal Joural of Theoretical ad Applied Fiace, vol. 2, o. 2, pp [14] Laboissiere, L. A., Ferades, R. A., & Lage, G. G, Maximum ad miimum stock price forecastig of Brazilia power distributio compaies based o artificial eural etworks. Applied Soft Computig, vol. 35, pp [15] Hayki, S. S. Neural etworks ad learig machies (Vol. 3). Upper Saddle River: Pearso Educatio 6

The Time Value of Money in Financial Management

The Time Value of Money in Financial Management The Time Value of Moey i Fiacial Maagemet Muteau Irea Ovidius Uiversity of Costata irea.muteau@yahoo.com Bacula Mariaa Traia Theoretical High School, Costata baculamariaa@yahoo.com Abstract The Time Value