Predicting Stock Price Changes of Tehran Artmis Company Using Radial Basis Function Neural Networks

Size: px

Start display at page:

Download "Predicting Stock Price Changes of Tehran Artmis Company Using Radial Basis Function Neural Networks"

Brett Harvey
5 years ago
Views:

Iteratioal Research Joural of Applied ad Basic Scieces 206 Available olie at www.irjabs.

1 Iteratioal Research Joural of Applied ad Basic Scieces 206 Available olie at ISSN X / Vol, 0 (8): Sciece Explorer Publicatios Predictig Stoc Price Chages of Tehra Artmis Compay Usig Radial Basis Fuctio Neural Networs Amirhossei Ghazavi *, Mohammad Aliyari, Mohammad Reza Mohammadi Departmet of Computer, Sciece ad Research Brach, Islamic Azad Uiversity, Tehra, Ira *Correspodig Author A.Ghazavi@srbiau.ac.ir ABSTRACT : Aalyzig ad predictig chages i stoc prices are of most sigificat capabilities to eter the stoc maret. Due to their relative abilities i recogizig the behavior ad chages of stoc prices, techical ad fudametal aalyses are amog the most basic ad most used statistical methods i the stoc maret. Due to the ability of eural etwors to uderstad differet behavior patters, the results of techical aalysis are predicted o a daily basis with a higher speed ad accuracy usig radial basis fuctio (RBF) eural etwors ad Leveberg learig algorithm. Keywords: Predictio, Radial basis fuctio eural etwors, Time series, Stoc maret. INTRODUCTION Oe of the highest paid jobs i the world is to buy, sell ad hold securities of publicly traded compaies i the stoc marets over the world. The most importat factor to esure success i stoc maretig is correct idetificatio ad aalysis of the fudametal ecoomic chages withi the compay as well as fudametal chages i the atioal ecoomic situatio ad fially ivestigatio of the chages i stoc price. Fudametal aalysis icludes ivestigatio of the ecoomic situatio of the compay i terms of umber of employees ad the Board of Directors ad such issues as well as the status of atioal ecoomic ifrastructure. O the other had, techical aalysis icludes uderstadig of the patter of stoc price chages i differet periods. Obviously, if both fudametal ad techical aalyses are cosidered by a aalyst, the aalysis will be of higher reliability. The idea behid techical aalysis is that share prices move i treds dictated by the costatly chagig attitudes of ivestors i respose to differet forces. Based o this idea, eural etwors were used to examie the chages i Tehra Artemis stoc price from 2008 to 206. The use of soft computig techiques such as eural etwors without gettig ito complex statistical issues of techical aalysis eables us to achieve daily buy ad sell sigals with a higher speed ad a acceptable accuracy [2, 0]. The paper icludes the followig sectios: Sectio II itroduces the eural etwors, radial basis fuctio (RBF) eural etwors ad Leveberg learig algorithm. Tehra Artmis dataset is aalyzed ad proposed method is discussed i Sectio III. Fially, the proposed algorithm is implemeted ad results are aalyzed. Artificial Neural Networs A artificial eural etwor (ANN) is a simplified model of a cetral system ispired by the huma brai. Usig sophisticated computatioal structure of commuicatios withi euros, it creates the ability to react to chages ad compliace with data eviromet. Artificial eural etwors trasfer owledge or rules withi the data to the etwor by processig existig data. Based o calculatios o the traiig data set, the etwor lears the geeral rules. Brai-based learig ad its applicatio are a relatively ew research area. Literature shows that it is worth spedig time o most claims of brai-based learig advocates. I this regard, various methods have bee proposed for brai-based traiig. The brai ca act simultaeously i differet ways. The brai costatly processes thoughts, emotios ad imagiatios. Furthermore, the brai is a europhysiological system that iteracts ad exchage with the surroudig eviromet. Learig capability is oe of the most importat features of eural etwors. Learig capability is the ability to adjust etwor parameters through time i differet situatios with the aim that the etwor traied for a specific situatio ca maitai its effectiveess despite small chages i evirometal coditios. After proper traiig, the etwor ca use of iterpolatio process to provide a appropriate output i dealig with a ew iput that ot ecoutered durig traiig. Fig. shows the geeral structure of artificial eural etwors. Neural etwors traied with gradiet descet based algorithms [, 3, 4, 0-20] ad evolutioary algorithms [6-9].

2 Itl. Res. J. Appl. Basic. Sci. Vol., 0 (8), , 206 Radial Basis Fuctio(RBF) Neural Networs I this type of etwors, the respose of each euro i the middle activatio layer is a fuctio of the Euclidea distace of the iput from the ceter of cluster of that euro. Geerally, RBF eural etwors mae use of complex patters for classificatio based o oliear mappig from a -dimesioal iput space to a m-dimesioal space so that the liear classificatio error of the mapped data is less tha that of the iitial data. The first layer output is calculated from the followig equatio. xc j 2 ( ) 2 () o j e j Where the ceter of the j-th is cluster ad is the stadard deviatio of the j-th euro. Figure. The geeral structure of artificial eural etwors. Figure 2. The geeral structure of RBF eural etwors. Leveberg learig algorithm I this article, we used Leveberg learig algorithm with traiig parameter μ. The method is discussed i detail i [9]. The LM traiig algorithm is as follows: 973

3 Itl. Res. J. Appl. Basic. Sci. Vol., 0 (8), , 206 W W ( J J I J e (2) T ) J c c c c w w w w Where J is the Jacobia matrix, I is the uit (idetity) matrix ad is the error i the -th step. Data aalysis ad proposed algorithm As metioed, applied data aalysis of the shares of ay compay should be based o both techical ad fudametal cocepts. Fudametal aalysis refers to aual fiscal reports ad covers small ad large corporate evets icludig mior chages i staff ad director ad geeral chages i ecoomic treds ad political issues, especially i the compay s field of activity. Daily predictio of stoc price chages is divided ito two parts. The first part cocers aual ad log-term aalysis ad thus fudametal aalysis provides the aalyst with a stroger loo. Fudametal aalysis ofte deals with log-term ivestmets. Short-term ivestmet ad specialized buyig ad sellig shares are called swig. The most importat factor i swig is correct idetificatio of daily buy ad sells sigals. Buy sigals are itesified whe stoc dowtred is broe ad become a uptred. The opposite holds true for a sell sigal. Accordigly, the eural etwor algorithm should be able to detect local extreme i the test data. I other words, detectig the upward or dowward tred of stoc price i comig days will ot be able to meet techical aalysis requiremets. At the first glace, a repeatig tred caot be foud i the stoc dataset of Tehra Artmis Costructio Compay. However, with a more detailed study ad a closer loo to fudametal data, the repeatig behavior of buyers ad sellers of shares i differet itervals ca be foud. As oted i the first sectio, the chages i our data i differet time itervals show a similar behavior. Amog differet eural etwor algorithms, the structure that aalyzes ad trais the data periodically should be selected. Radial basis fuctio eural etwor is a appropriate optio i this case. Due to the ature of RBF eural etwors, they show good performace i terms of iterpolatio ad classificatio. This ca be attributed to this fact that etwor iitializatio ad classificatio of same data are performed alog with extractio of cluster ceters ad sigmoid for each cluster. Figure 3. Daily stoc price of Tehra Artmis Compay. Experimetatio Ad Results The proposed algorithm was implemeted for predictig chaotic time series of Tehra Artmis Compay. Fially, the results of proposed algorithm were compared with those obtaied from other commo procedures. Fig.4 to Fig.9 ad Table show the results obtaied from various eural etwors with differet traiig methods. It is oteworthy that the raw data has bee ormalized. The structures iclude 2-9-, 2-5-, structures with GD, Rough (periodic owledge), Jorda, Elma ad Elma Jorda etwors. 974

4 Itl. Res. J. Appl. Basic. Sci. Vol., 0 (8), , 206 This reflects data iterpolatio accordig to the behavior of closer data or, from the stoc diagram view, repetitive behavior of buyig ad sellig securities i differet time itervals i the lac of ay fudametal chage i the ecoomic process. Accordig to the above argumets, RBF eural etwors detect the behavior of both test ad trai datasets. The most importat poit is accurate detectio of local extreme of test data by the eural etwor. It should be oted that if we oly suffice to determie the behavior of trai ad test data ad low levels of errors i test ad trai datasets, the project will ot have ay practical value. The importat poit to detect data tred is accurate idetificatio of local extreme ow as top buy ad sale sigals i techical aalysis. Herei, Leveberg learig algorithm is used to trai the RBF eural etwors for accurate detectio of local extreme. The stoc price i the comig day was predicted usig the stoc price i the last two days. The output data () is calculated usig the iput data (-) ad data (-2). The proposed method used RBF eural etwors ad Leveberg learig algorithm ad reached good results oly with 50 iteratios. I additio, it detects ad predicts local extreme with a high accuracy. A very high speed with a high accuracy of are of positive aspects of this structure i predictig the data tred. I the (2-5-) RBF etwor traied by LM method, all the ceters of clusters ad stadard deviatios are iitialized at the start of the algorithm usig K-Meas clusterig algorithm. The fial layer weights i the rage [-, ] are radomly iitialized with a uiform distributio. I the proposed method, the ceters of clusters, sigmoid, ad weights have bee traied at each step of the algorithm. Table. The best results of proposed algorithm ad other algorithms Best MSE of Best MSE of structure Epoch trai test MLP 2-9- MLP Elma- Jorda propose Figure 4. The best result obtaied from the gradiet descet algorithm with the structure (2-9-) - Trai Figure 5. The best result obtaied from the gradiet descet algorithm with the structure (2-9-) - Test 975

5 Itl. Res. J. Appl. Basic. Sci. Vol., 0 (8), , 206 Figure 6. The best result obtaied from the gradiet descet algorithm with the structure (2-6-3-) Trai Figure 7. The best result obtaied from the gradiet descet algorithm with the structure (2-6-3-) Test Figure 8. The best result obtaied from the Elma Jorda algorithm with the structure (2-5-) Trai Figure 9. The best result obtaied from the Elma Jorda algorithm with the structure (2-5-) - Test 976

6 Itl. Res. J. Appl. Basic. Sci. Vol., 0 (8), , 206 Figure 0. The best results of the proposed algorithm - Trai Figure. The best results of the proposed algorithm - Test Figure 2. Regressio results of RBF etwor traiig by Leveberg learig algorithm- Test 977

7 Itl. Res. J. Appl. Basic. Sci. Vol., 0 (8), , 206 CONCLUSION Based o existig soft computig techiques, the radial basis fuctio (RBF) eural etwors traied by Leveberg-Marquardt (LM) learig algorithm were used to predict time series stoc price of Tehra Artmis Compay with a high accuracy of *0-5. REFERENCES M. T.Haga, M. B. Mehaj, "Traiig feed forward etwors with the Marquardt algorithm." IEEE Trasactios o Neural Networs 5(6) (994) R. Lawrece, "Usig eural etwors to forecast stoc maret prices." Uiversity of Maitoba (997). A. Falahiazar, S. Setayeshi, Y. Sharafi,"Computatioal model of social itelligece based o emotioal learig i the Amygdala." Joural of mathematics ad computer sciece 4 (205) Y. Sharafi, S. Setayeshi, A. Falahiazar, "A Improved Model of Brai Emotioal Learig Algorithm Based o Iterval Kowledge." Joural of mathematics ad computer sciece 4 (205) M. T.Haga, M. B. Mehaj, "Traiig feed forward etwors with the Marquardt algorithm." IEEE Trasactios o Neural Networs 5(6) (994) D. Gyaesh, P. K. Pattai, S. K. Padhy, "Artificial eural etwor traied by particle swarm optimizatio for o-liear chael equalizatio."expert Systems with Applicatios 4(7)(204) Yusiog, Joh Paul T. "Optimizig Artificial Neural Networs usig Cat Swarm Optimizatio Algorithm." Iteratioal Joural of Itelliget Systems ad Applicatios 5. (202): 69. Y. Sharafi, M.A. Khaesar, M. Teshehlab, "Discrete Biary Cat Swarm Optimizatio Algorithm." i Computer, Cotrol & Commuicatio (IC4), 203 3rd Iteratioal Coferece o (203). K. I. Shereef, S.S. Baboo, "A New Weather Forecastig Techique usig Bac Propagatio Neural Networ with Modified Leveberg- Marquardt Algorithm for Learig." IJCSI Iteratioal Joural of Computer Sciece 8(6)(20) M. Zeic, "Neural etwor applicatios i stoc maret predictios-a methodology aalysis." proceedigs of the 9th Iteratioal Coferece o Iformatio ad Itelliget Systems (98) 998. A. V. Devadoss, T. A. Ligori, "Stoc predictio usig artificial eural etwors." Iteratioal Joural of Data Miig Techiques ad Applicatios (2) (203) M. Zeic, "Neural etwor applicatios i stoc maret predictios-a methodology aalysis."proceedigs of the 9th Iteratioal Coferece o Iformatio ad Itelliget Systems ( 98)(998). Z. H. Kha, T. Sh. Ali, M. A. Hussai, "Price predictio of share maret usig artificial eural etwor (ANN)." Iteratioal Joural of Computer Applicatios 22(2) (20) M. P. Naeii, H. Taremia, H. Baradara Hashemi, "Stoc maret value predictio usig eural etwors." Computer Iformatio Systems ad Idustrial Maagemet Applicatios (CISIM), 200 Iteratioal Coferece o. IEEE, 200. G. Dutta, "Artificial eural etwor models for forecastig stoc price idex i the Bombay stoc exchage." Joural of Emergig Maret Fiace 5(3) (2006) Y. Shachmurove, D. Witowsa, "Utilizig artificial eural etwor model to predict stoc marets." Uiversity of Pesylvaia, Ceter for Aalytic Research i Ecoomics ad the Social Scieces, K. Tseg, O. Kwo, L. C. Tjug, "Time series ad eural etwor forecast of daily stoc prices." Ivestmet Maagemet ad Fiacial Iovatios 9() (202) R. Aghababaeya, T. Siddiqui, N. AhmadKha, "Forecastig the Tehra Stoc Maret by artificial eural etwor." Iteratioal Joural of Advaced Computer Sciece ad Applicatios (IJACSA), Special Issue o Artificial Itelligece (3) (20). B. W. Wajawa, L. Muchemi, "ANN Model to Predict Stoc Prices at Stoc Exchage Marets." arxiv preprit arxiv: (204). R. Gh. Ahagar, M. Yahyazadehfar, H. Pouraghshbad, "The compariso of methods artificial eural etwor with liear regressio usig specific variables for predictio stoc price i Tehra stoc exchage." arxiv preprit arxiv: (200). 978

Radar emitter recognition method based on AdaBoost and decision tree Tang Xiaojing1, a, Chen Weigao1 and Zhu Weigang1 1

Radar emitter recognition method based on AdaBoost and decision tree Tang Xiaojing1, a, Chen Weigao1 and Zhu Weigang1 1 Advaces i Egieerig Research, volume 8 d Iteratioal Coferece o Automatio, Mechaical Cotrol ad Computatioal Egieerig (AMCCE 7) Radar emitter recogitio method based o AdaBoost ad decisio tree Tag Xiaojig,