Evaluation of Artificial Immune System with Artificial Neural Network for Predicting Bombay Stock Exchange Trends

Transcription

1 Joural of Computer Sciece 7 (7): , 20 ISSN Sciece Publicatios Evaluatio of Artificial Immue System with Artificial Neural Network for Predictig Bombay Stock Exchage Treds M. Guasekara ad 2 K.S. Ramaswami Departmet of Computer Applicatios, Park College of Egieerig ad Techology, Coimbatore, Idia 2 Departmet of Mathematics, Coimbatore Istitute of Techology, Coimbatore, Idia Abstract: Problem statemet: The purpose of this study is to develop a artificial immue system for recogizig stock market treds ad predict upward ad dowward directios of stock market. This study compared two predictio models, a Artificial Immue System (AIS) ad Artificial Neural Network (ANN) for predictig the future idex value, tred of Idia stock market ad discovers the best predictio model. Approach: AIS is a efficiet system for predictig tred due to its high capability of learig ad retaiig iformatio i memory. Our proposed system was tested usig SENSEX (Sesitive Idex) data from Bombay Stock Exchage (BSE) of Idia. Results: Performace of models have bee evaluated o the basis of the simulatio results doe o MATLAB. Experimets have bee performed for both methods o well-kow techical idicators ad compared their results with SENSEX data. Coclusio: Artificial Immue System is more efficiet tha Artificial Neural Network. Key words: Artificial immue system, artificial eural etwork, techical idicators, bombay stock exchage, stock market INTRODUCTION Recet research activities o the Artificial Immue System (AIS) have established that immue system have powerful predictio ad patter recogitio capabilities. I recet years, dealig with ucertaity is the big issue. We implemet Artificial Immue System (AIS) is a ew kid of atural computatio method to deal this ucertaity efficietly for predictig the stock tred. Artificial Immue System (AIS) is a computatioal itelligece model to formulate adaptive systems capable of performig a wide rage of tasks i various egieerig applicatios, such as abormality detectio, patter classificatio ad recogitio. AIS are motivated by biological immue system, which have the immue system s characteristics of learig ad memory. A eural etwork has the capability to process the iput variables i parallel ad ca maage large amouts of data i rapid maer (Hsieh, 200). The greatest stregth of eural etwork is its skill to discover patters easily. Artificial Neural Networks (ANN) has bee widely accepted to deal with stock market predictio (Chag et al., 2009; Majee ad Roy, 200). Techical idicators are used to determie the behavior of ivestors ad their impact o the future price movemet of the stock from past movemets. A well performig stock portfolio, that allows ivestors to diversify away usystematic risk ad ehace the quality of ivestmet by sellig ad buyig the stock (Vazakidis ad Adamopoulos, 2009). The mai data of our studies are the price histories of stock, with time ad volume iformatio. Dow ad up treds of the price o the stock market depeds o supply ad demad of stocks. The SENSEX is a weighted average of 30 shares market value of the Bombay Stock Exchage (Idia). The SENSEX was itroduced by Bombay Stock Exchage (BSE) i 986. The BSE SENSEX is cosidered as a bechmark i this experimet. The mai objective of the study is to develop predictive model for fidig the ext day s close value of SENSEX. The remaider of this study is orgaized as follows: Iitially differet predictio models are described. After that Artificial Immue System ad Artificial Neural Network are tested ad compared with BSE Sesex i order to predict the tred of Idia market. The the experimetal results ad umerous discussio issues are preseted. Fially the coclusios are summarized relatively. Correspodig Author: M. Guasekara, Departmet of Computer Applicatios, Park College of Egieerig ad Techology, Coimbatore, Idia 967

2 MATERIALS AND METHODS Number of researchers has proposed to aalyze ad predict stock market activity for the last two decades. Liear methods domiated predictio for the past several years. Liear methods are simple to recogize ad iterpret ad they are also comparatively easy to create ad deploy. Liear models also have serious problems o the other side which are uable to idetify o-liear relatioships i stock market data. Artificial Neural Network (ANN) model: Artificial Neural Network (ANN) methodology allows us to develop oliear systems receivig large quatities of iputs that developmet based o oly istaces of iput-output relatios (Vazakidis ad Adamopoulos, 2009). A major applicatio field of ANNs is forecastig. ANN is especially best fit for discoverig accurate solutios i a eviromet characterized by oliear, irrelevat, complex, partial or oisy data (Eletter et al., 200). However, several limitatios happeed eve if they were recogized with multi dimesioal oliear models. The multi layer feed forward eural etwork model with back propagatio algorithm for traiig is employed for predictig stock market is show i Fig.. A Neural etwork essetially has three layers which are the iput layer, the output layer ad the hidde layer i betwee. Neural etwork is made up of itercoected euros ad each euro has the relatioship with eighborig euros. The relatioship betwee the euros is measured by weight. The iput layer x of euros acquires the data directly from the stock exchage. After computatio has bee performed by the hidde layer y, resultat outputs are trasferred to output layer z. Output of ode i hidde layer is derived from iputs, weight of coectios ad a oliear fuctio called as activatio fuctio (Solaimai, 2009). Actual output of ode i hidde layer y is defied as: AO = f (sum ) () y sum x y = w xyao x (2) x= w xy = The weight of the coectios betwee iput ode x ad hidde ode y AO x = The actual output of ode x ad f is the activatio fuctio J. Computer Sci., 7 (7): , 20 Fig. : Artificial eural etwork Back propagatio algorithm adjusts the weights of coectios slightly. Output layer has oly oe euro. We ca compare the actual output AO ad the predicted output PO for the output layer z ad variatio V betwee these values is measured as: V = (PO AO ) 2 z z (3) 2 z = Back propagatio algorithm decides to icrease or decrease weights w xy for miimizig the variatio: w xy = w xy ± w xy (4) where, w xy is the weight of the lik betwee the ode x ad ode y. O eural etwork, umber of suitable euros is preseted i the iput ad hidde layers ca modify the predictio performace (Alsmadi ad Omar, 200; Dastorai et al., 200). Artificial immue system model: I recet years, Artificial Immue System (AIS) is itroduced amog the differet categories of predictio methods. Artificial Immuology cocept was proposed by Professor Forrest, Uiversity of New Mexico i 997 (Wag et al., 2009). Artificial Immue System (AIS) is a ew computatioal itelligece method ispired by biology immue system (Deg ad Gao, 2009). A huma immue system cosists of B-cells that recogize atiges which eterig from exteral eviromet. B-cells are a type of immue cell, depedig o protei molecules called atibody (Golzari et al., 20). A B-cell idetifies the atige whe its atibodies appear ito cotact with a atige of complemetary shape (Do et al., 2009). I a discrete immue etwork, cells ad molecules ca elarge or reduce i umber ad may also chage their 968

3 J. Computer Sci., 7 (7): , 20 behavior to develop their similarity. The discrete immue etworks adjust may elemets ad adjust the costructio of these elemets. This approach is facilitatig to solve may problems. It obviously deals the relatioship with the exteral eviromet (atiges) as the discrete immue etworks plaed at problem solvig. A pseudo-code for AIS algorithm is give below:. Geerate a iitial atibodies 2. while stoppig criteria is ot met do 3. Estimate the fitess value of each atibody agaist the atige 4. Fid the better fitess value amog atibodies 5. while solutio fitess value do 6. Apply mutatio for gettig exact fitess value for solutio 7. ed while 8. produce cloed atibodies 9. if solutio = fitess value of cloed atibody 0. the produce more cloed atibody ad set cloed atibody as atibody. ed if 2. ed while 3. retur the best solutio foud I Artificial Immue Network, each elemet correspods to a B-cell composed of a atibody. Euclidea distace is a distace betwee each atige ad atibody (or a B cell) is give by Eq. 5: ED = (AG AB ) (5) ij ix jx x AG ix = x th attribute value of atige i AB jx = x th attribute value of atibody j accordig to the idea metioed above ad it is illustrated i followig Fig. 2. Data refiig is very importat stage, the Sesig phase filter uecessary data ad decrease data quatity for extractig relevat iformatio from the historical stock data. The Detectio phase validates the outcome of sesig phase by utilizig kowledge database (Prado et al., 200). The Defed phase has higher perceptio level that aalyzes a solutio sythetically ad provides a exact decisio. Future tred ca be efficietly predicted i this way. The SENSEX idex of curret stock market has the rise ad fall i price ad volume may chage the result of the detectio. To overcome this problem, AIS system updates the kowledge database ad the detectors, so the system will adjust itself to discover effectively the chages of exteral factors. A fiacial applicatio has bee developed to predict treds usig tradig rules that suggests ivestors whe should buy or sell various stocks (Widiputra et al., 2009). Techical aalysis ivestigates patters ad movemets i price ad volume charts by utilizig ifereces hidde i past tradig activities. It believes that past history will repeat i future ad the relatioship betwee price ad volume clearly expresses share market performaces (Oskooe, 200). Techical aalysts use techical idicators, which are mathematical formulas that predict the market tred easily (Oskooe, 200). Amog the various techical idicators, we utilize the Simple Movig Average (SMA) ad Expoetial Movig Average (EMA) betwee differet movig averages ad Relative Stregth Idex (RSI) for the price aalysis. Arms Idex ad Moey Flow Idex are used for aalyzig volume (Chag et al., 2009; Abbodate, 200). The probability of selectig a atibody from the populatio of atibodies AB accordig to the populatio of atiges AG is give as: p(j) = atibody( j) j= atibody( j) (6) j = The series of atibody or affiity = The populatio of atibodies This study has developed a geeral artificial immuity system framework to defese curret stock data Fig. 2: Geeral framework of artificial immue system 969

4 J. Computer Sci., 7 (7): , 20 Table : Techical idicators used as iput variables S. No Techical idicators Descriptio Formula Simple Movig Average (SMA) Calculates the average price over a specified movig time period. (/ m) x 2 Expoetial Movig averages the last x days closes but EMA = (CP curret-cp previous)* 2/ (+) Averages (EMA) assigs a greater weight to the more + CP previous recet prices makig it more sesitive Where CP Closig Price to curret price actio 3 Moey Flow Idex Measures the stregth of moey Positive Moey Flow Moey Flow Ratio = Negative Moey Flow flowig ito ad out of a stock. Where It also takes both price ad Moey flow = Avg. Price* Day s Volume volume ito accout. Day's High + Day's Low + Close Avg. Price = 3 4 Arms Idex Measures relative volume flows Arms Idex= ( AI/DI ) (AV/DV) AI = No. of Advacig issues AV = Advacig Volume DI = No. of Decliig issues DV = Decliig Volume If AI >.0, Market is dow tred AI <.0, Market is up tred 5 Relative Stregth RSI measures the velocity of price RSI=00-[00/ (+RS)] Idex (RSI) movemets ad determie the where: overbought ad oversold coditio RS = (Avg. of -day up closes)/ (Avg. of -day dow closes) = days (9-5 days) t i= t m+ i We first show how parameters are used to cofigure the system. The iput variables used i this study are based o well-kow techical idicators. Both immue etwork ad artificial etwork use these techical idicators for predictig the stock market tred. We have cosidered 5 importat techical idicators with brief descriptio ad formula, as show i Table. To evaluate the performace of both models for calculatig the followig statistical metrics, amely Root Mea Squared Error (RMSE), Mea Absolute Error (MAE) ad Mea Absolute Percetage Error (MAPE) of the traied forecastig model for the test data (Sigh ad Ahmad, 20; Assis et al., 200). Root Mea Squared Error (RMSE) is described as follows: RMSE = (A P ) x 00 2 i i (7) i = Mea Absolute Error (MAE) is give as: MAE = A P [ i i ] (8) i = Similarly Mea Absolute Percetage Error (MAPE) is defied as: 970 Ai P i MAPE = (9) i= A i A i = Actual value of idex o day i P i = Predicted value o day i of forecastig models If the above RMSE is very less sigificat, the predictio accuracy of the system is very close to 00%. RESULTS The predictio models have bee successively implemeted ad tested o MATLAB based Graphical User Iterface. The system was tested o Idia BSE SENSEX Idex usig historical data from We have used separate data sets for traiig ad testig. Both eural etwork ad immue etwork will be traied usig the data from 2d Jauary 2009 to 3st July The testig period is selected to be from 3rd August 2009 to 30th July 200. We have cosidered the historical data as a source of iformatio o the closig prices of BSE SENSEX idex. Our objective is to compare the predicted results ad values of above metioed methodologies with curret treds of BSE SENSEX idex.

5 J. Computer Sci., 7 (7): , 20 above Table 2. From above empirical results, it is very evidet that the AIS model gives better results o the problem of stock market forecastig. As a result, the Artificial Immue System ca be used to predict the stock market tred efficietly. CONCLUSION Fig. 3: Actual ad projected value of BSE SENSE X geerated by the AIS ad ANN Table 2: Evaluatio of AIS ad ANN usig test data ERRORS Methods RMSE MAE MAPE AIS ANN Comparisos ad results of both AIS ad ANN models were evaluated by estimatig the error betwee the curret price of closig ad the predicted closig price. Figure 3 shows the predicted results of AIS ad ANN models ad performace of predicted models were compared with actual BSE SENSEX idex. The above graph demostrates the predicted values of two predictio models durig the period The actual values for the period are also illustrated. We evaluated the forecast accuracy measures for this period. The predictio performace is evaluated by cosiderig the variatio betwee the predicted value ad actual value. Table 2 obviously shows the results have bee obtaied by two differet models, such as AIS ad ANN for predictig BSE SENSEX idex. DISCUSSION I this study, we have proposed the Artificial Immue System for predictig the stock market movemet. Our ew AIS predictio model is compared with aother ANN predictio model ad bechmark idex SENSEX. We clearly idetified that the predicted AIS data is very close to actual SENSEX data tha predicted ANN data from the above Fig. 3. The statistical measures of predictio error- RMSE, MAE ad MAPE of AIS were sigificatly lower tha those obtaied with ANN as idicted i the 97 Numerous soft computig approaches have bee applied effectively i predictig the movemet of stock tred ad produced better performace. I this study, we have proposed a immue based computatioal strategy for predictig the stock tred which icludes sesig, detectig ad defedig. The experimetal studies have bee coducted o the Bombay Stock Exchage (BSE SENSEX Idex, Period: August 2009-July 200) datasets. Most of the research papers proposed the ANN model for predictig o-liear data such as stock market. Recetly, AIS has the ability to predict oliear data very well. This study is evaluated ad compared the performace of ANN ad AIS models. The AIS model has the lowest value of RMSE, MAE ad MAPE. Our experimetal results propose that the AIS model ca provide better predictios tha the stadard ANN model. We will improve optimal way to fuse the various itelliget systems usig a hybrid approach i future work. We will cosider differet methodologies to overcome the drawbacks of a specific itelliget system. We will also develop multi-layer presetatios of AIS for hadlig the abormal characteristics of stock market. REFERENCES Abbodate, P., 200. Tradig volume ad stock idices: A test of techical aalysis. Am. J. Eco. Bus. Admi., 2: DOI: /ajebasp Alsmadi, M.K. ad K.B. Omar, 200. Fish recogitio based o robust features extractio from size ad shape measuremets usig eural etwork. J. Comput. Sci., 6: DOI: /jcssp Assis, K., A. Amra, Y. Remali ad H. Affedy, 200. A compariso of uivariate time series methods for forecastig cocoa bea prices. Treds Agric. Eco., 3: DOI: /tae Chag, P.C., C.Y. Fa ad C.H. Liu, Itegratig a piecewise liear represetatio method ad a eural etwork model for stock tradig poits predictio. IEEE Tras. Syst., Ma Cyberetics-Part C: Appli. Rev., 39: DOI: 0.09/TSMCC

6 J. Computer Sci., 7 (7): , 20 Dastorai, M.T., A. Talebi ad M. Dastorai, 200. Usig eural etworks to predict ruoff from ugauged catchmets. Asia J. Applied Sci., 3: DOI: /ajaps Deg, L. ad D.Y. Gao, Research o immue based adaptive itrusio detectio system model. Proceedigs of the Iteratioal Coferece o Networks Security, Wireless commuicatios ad Trusted Computig, Apr , IEEE, Hubei, Chia, pp: DOI: 0.09/NSWCTC Do, T.D., S.C. Hui amd A.C.M. Fog, Associative classificatio with artificial immue system. IEEE Tras. Evolut. Comput., 3: DOI: :02:03.0 Eletter, S.F., S.G. Yasee ad G.A. Elrefae, 200. Neuro-based artificial itelligece model for loa decisios. Am. J. Eco. Bus. Admi., 2: DOI: /ajebasp Golzari, S., S. Doraisamy, M.N. Sulaima ad N.I. Udzir, 20. A efficiet ad effective immue based classifier. J. Comput. Sci., 7: DOI: /jcssp Hsieh, K.L., 200. Employig artificial eural etworks ito achievig parameter optimizatio of multi-respose problem with differet importace degree cosideratio. Iform. Techol. J., 9: DOI: /itj Majee, N.C. ad A.B. Roy, 200. Asymptotic behavior of a artificial eural etwork defied o multipartite directed graph. OLie J. Biol. Sci., 0: DOI: /ojbsci Oskooe, S.A.P., 200. Emergig stock market performace ad ecoomic growth. Am. J. Applied Sci., 7: DOI: /ajassp Prado, R.P., N. Garcia-Galá, S. Exposito ad J. Yuste, 200. Kowledge acquisitio i fuzzyrule-based systems with particle-swarm optimizatio. IEEE Tras. Fuzzy Syst., 8: DOI: 0.09/TFUZZ Sigh, V.K. ad N. Ahmad, 20. Forecastig performace of costat elasticity of variace model: Empirical evidece from Idia. It. J. Applied Eco. Fia., 5: DOI: /ijaef Solaimai, K., A study of raifall forecastig models based o artificial eural etwork. Asia J. Applied Sci., 2: DOI:0.3923/ajaps Vazakidis, A. ad A. Adamopoulos, Stock market developmet ad ecoomic growth. Am. J. Applied Sci., 6: DOI: /ajassp Wag, W., S. Gao ad Z. Tag, Improved patter recogitio with complex artificial immue system. J. Soft. Comp., 3: DOI: 0.007/s Widiputra, H., R. Pears, A. Serguieva ad N. Kasabov, Dyamic iteractio etworks i modellig ad predictig the behaviour of multiple iteractive stock markets. Itell. Syst. Acc. Fi. Mage., 6: DOI: 0.002/isaf