In this paper, we use Nonlinear Autoregressive Network with eogenous inputs!NA"# recurrent dynamic neural network to forecast next day stock price. $inding optimi%ed model for the neural network to get the best performance compared to other models is difficult task. $irstly, we try to analy%e the effects of the number of neurons present in neural network on the performance of the network.
In this paper, we use Nonlinear Autoregressive Network with eogenous inputs!NA"# recurrent dynamic neural network to forecast next day stock price. $inding optimi%ed model for the neural network to get the best performance compared to other models is difficult task. $irstly, we try to analy%e the effects of the number of neurons present in neural network on the performance of the network.
In this paper, we use Nonlinear Autoregressive Network with eogenous inputs!NA"# recurrent dynamic neural network to forecast next day stock price. $inding optimi%ed model for the neural network to get the best performance compared to other models is difficult task. $irstly, we try to analy%e the effects of the number of neurons present in neural network on the performance of the network.
Saurabh Jha School of Computing Science and !ngineering "ellore #ntitute of $echnolog% Chennai& #ndia aurabh'(ha)*+*,-it'ac'in Aman Singh School of Computing Science and !ngineering "ellore #ntitute of $echnolog% Chennai& #ndia aman'ingh)*+*,-it'ac'in AbstractThe prediction of next day stock price performance is a complex problem. Technical analysis and fundamental analysis has helped form basis for stock price expectation and decision making. In this this paper, we use Nonlinear Autoregressive Network with eogenous inputs !NA"# recurrent dynamic neural network to forecast next day stock price. $inding optimi%ed model for the neural network to get the best performance compared to other models is difficult task. The purpose of this paper is to find and optimi%e the parameters of the network model. &e make a two fold study. $irstly, we try to analy%e the effects of the number of neurons present in neural network on the performance of the network. It is observed that a network with ' ( ) neurons gives the best result for single stock price prediction. *econdly, training the network is a time and resource consuming task which depends on number of neurons and features. This issue is addressed by running +rincipal ,omponent Analysis !+,A# on the features and paralleli%ing the training phase of the network model on ,+-. KeywordsStock price prediction, Neural networks, Performance analysis, Machine Learning, parallel programming #' #N$RO./C$#ON $he main characteritic of an% financial aet i it return' Return can be ma0imized if the capital ha been in-eted in the right tock at the right time' Anal%i of the tock1 performance during the da% can gi-e a fair idea of the market -iew about the compan%' 2or e0ample 3 if a tock of compan% A climb during mid4da% trading and end on a high then it can be aid that the in-etor are in-eting in the compan% wherea& if the oppoite pattern i ober-ed then it can be aid that the in-etor are pulling the mone% out of the compan%' $hi i a negati-e pattern and there i a -er% high poibilit% that the trend o ober-ed on a gi-en da% can continue on to the ne0t da% and affect the tock price' $hi gi-e the prediction of the ne0t da%1 trading price great importance' A right prediction can help in-etor make mone% and in-et more on the tock wherea an error can lead to loe for the in-etor' 2or a long term in-etor& intra4da% trading will not be a concern a hi target are planned for a longer duration' 5ut& for a hort term in-etor& intra4da% trading hold ke% -alue' Prediction of price help people take 6uick deciion on bu%ing and elling of tock' Alo& thi prediction help the tock broker form an idea about the compan%1 mo-ement in the market and the% tr% to gi-e an impro-ed ad-ice to their client' Man% in-etment companie in-et their capital in the tock market with calculated rik and e0pected outcome' .ifferent tatitical and technical tool ha-e been hitoricall% ued to predict trend in time erie and aid in-etment deciion' .ue to the criticalit% of thee deciion and the -olume of in-etment in-ol-ed& there i coniderable interet in modification and ad-ancement of prediction methodologie' $hu& tock market prediction ha pro-ed to be -er% important tool that i ued in tock market anal%i' #n thi tud%& we ued -ariou technical indicator to predict ecurit% price of a tock uing technical anal%i' $he technical anal%i i baed on the aumption that hitor% tend to repeat itelf in the tock e0change' #f a certain pattern of acti-it% ha in the pat produced certain reult nine out of ten& one can aume a trong likelihood of the ame outcome whene-er thi pattern appear in the future' .ow theor%7++8 i a tetament to thi aumption' #n thi paper& we ue NARX neural network to model financial tock market to predict tock price' $hi paper make the following contribution' .etermine the optimum et of feature for training' .etermine the optimum number of neuron in the NARX network for tock price prediction' .etermine the optimum plit between train4tet44-alidation et in data' Compare the run time of parallel implementation -eru erial implementation of the model' $hee anal%i are pre4work to our future goal which i to claif% policie of tock market egment to increae the efficienc% of the tock price prediction' Stock data can be downloaded from -ariou ource' 9owe-er& we choe :ahoo financial er-er78 to download tock data' 9itorical data on da% b% da% bai can be eail% downloaded for a tock from thee er-er' $he data i a-ailable in Comma Separated "alue ;CS"< file format' #t can alo be downloaded through MA$=A5 financial tool bo0' >e ued the tock data from *+4*+4)**? to *+4*+4)*+@' ##' R!=A$!. >ORA $he core premie of tock market price prediction deri-e it method and theorie from @ chool of thought' $he firt chool belie-e that no in-etor can& baed on hitorical data and theorie& achie-e more than a-erage reult with hi in-etment' $hi i upported in turn b% a few theorie uch a the Random Walk Hypothesis and the Efficient Market Hypothesis 7+8. Random >alk 9%pothei tate that there i no uch pattern in the mo-ement of tock price and hence it cannot be predicted' !fficient Market 9%pothei tate that financial market and %tem are informationall% efficient& hence getting return more than the a-erage on rik ad(uted in-etment continuoul% i not poible' $he econd chool belie-e that fundamental anal%i can pro-ide a uitable tool to predict the mo-ement of tock market' 2undamental anal%i deal with the baic fundamental factor that concern the operation of a compan%' $hee can be profit& di-idend& new contract& merger& lo& failure of a product or er-ice 7@87B8' $hee baic indicator gi-e an idea of the compan%C functioning and in a wa% reflect the performance of the compan%' $he third chool follow the technical anal%i' $echnical anal%i i baed on the core premie that tock -alue follow a certain a pattern and if thi pattern can be detected then their future -alue can be predicted 7)8' $hi pattern can be pat data or data baed on certain e-ent or particular happening in the compan% or it en-ironment' $he price and -olume information on a compan% can help in prediction' #f the tock touche a particular high and come down time and again then it can be termed a a reitance price or a reitance le-el for the tock and the in-etor can ell a oon a the tock reache that -alue' Same can be aid for the pattern found for upport price or upport le-el' Such pattern gi-e a trend that can be mapped with the tockC mo-ement' Denerall% fi-e baic parameter are deri-ed for thi purpoe out of a tockC data E4 open price& cloe price& highet traded price& lowet traded price and -olume traded' Man% theorie e0it in the market baed on which in-etor mo-e their capital in the market like 4 F#f a tockC recent mo-ing a-erage ha been better than pat a% )*4@* da% mo-ing a-erage and both a-erage are headed north then we can bu%G 7?8' Sometime there are no e0planation for the additional parameter added due to which thi method i not con-incing e-er% time' Current in-etor keep a broad approach where the% take reult from both the anal%i and then make a deciion' 5oth $echnical anal%i and 2undamental anal%i can be imulated on Neural Network' A far a their implementation on Neural Network goe& a few algorithm and techni6ue ha-e been found to be preent in mot of the cae' $hee include the Backpropagation Algorithm& Piecewise Linear Representation, Principal omponent Analysis, !"##y $e"ral $etworks, %opology Represented $etworks. #n our anal%i we are uing the =e-enberg 3 Mar6uartdt back propagation algorithm in NARX network7H8 e0plained in ection B' $hi algorithm& help in modification of the weight aigned to the neuron& in cae of error in the predicted and ober-ed -alue' ###' $!C9N#CA= #N.#CA$ORS AN. 2!A$/R! S!=!C$#ON Principall%& technical anal%i i baed on hitorical data& current price and the -olume of the hare traded' #n imple word& $echnical anal%i i baed on model and trading rule that ha-e been deri-ed uing the pat data obtained from the market like mo-ing a-erage& relati-e trength inde0 and certain other indicator' $hi anal%i alo look for chart pattern uch a triangle& trend line' 2or our model& data wa downloaded from %ahoo er-er with cutomized MA$=A5 cript' 9owe-er& the hitorical data for #ndian tock market can alo be downloaded from 5S! or NS! webite' $he data i a-ailable in comma eparated -alue ;CS"< for all tock on da% b% da% bai' One need to pre4proce thee downloaded file to e0tract pecific tock data o-er a period of time' $herefore& we ued %ahoo er-er with our own cutomized cript to download onl% meaningful and pecific data of tock' $he hitorical tock data generall% ha-e the following feature ;attribute<E Stock Name& .ate& Open& 9igh& =ow& Cloe& Pre-iou Cloe& "olume' $here are o-er +** technical indicator being ued in tock market toda%' 9owe-er& e0pert rel% onl% on mall et of feature to determine the beha-ior of tock market' #n thi tud% we choe the following technical indicator for our initial anal%iE A. Bollinger Bands &BB' 5ollinger 5and are -olatilit% band placed abo-e and below a mo-ing a-erage' "olatilit% i baed on the tandard de-iation& which change a -olatilit% increae and decreae' $he band automaticall% widen when -olatilit% increae and narrow when -olatilit% decreae' $he purpoe of the band i to pro-ide a relati-e definition of high and low' $he% comprie a et of @ cur-e& drawn in relation to the price of ecuritie' $he middle band ignifie an intermediate trend 3 imple mo-ing a-erage generall%& which er-e a the bae for the high and low band' $he high band hold a ( times tandard de-iation abo-e the mo-ing a-erage while the low band hold a ( times tandard de-iation below the mo-ing a-erage' B. Relati)e *trength +nde, &R*+' Relati-e Strength #nde0 i a momentum ocillator that meaure current price trength in relation to pre-iou price' RS# ocillate between zero and +**' $he RS# compute momentum a the ratio of higher cloe to lower cloeE tock which ha-e had more or tronger poiti-e change ha-e a higher RS# than tock which ha-e had more or tronger negati-e change' . Mo)ing A)erage &MA' A mo-ing a-erage& alo called rolling a-erage& mo-ing mean& rolling mean& or running a-erage& moothen the price data to form a trend following indicator' $he% do not predict price direction& but rather define the current direction with a lag' Mo-ing a-erage lag becaue the% are baed on pat price' A imple mo-ing a-erage i formed b% computing the a-erage price of a ecurit% o-er a pecific number of period' !0ponential mo-ing a-erage reduce the lag b% appl%ing more weight to recent price' $he weighting applied to the mot recent price depend on the number of period in the mo-ing a-erage' #t can be calculated uing the formula gi-en below 4 "rrent EMA- &&Price&c"rrent' . pre)io"s EMA'' / m"ltiplier' 0 pre)io"s EMA Mot mo-ing a-erage are baed on cloing price' Mo-ing a-erage are generall% ued to meaure momentum and define area of poible upport and reitance' $%picall%& upward momentum i confirmed when a hort4term a-erage ;e'g'+?4da%< croe abo-e a longer4term a-erage ;e'g' ?*4da%<' .ownward momentum i confirmed when a hort4term a-erage croe below a long4term a-erage' 1. 2ol"me Rate of hange &2R3' $he "olume Rate of Change indicator meaure the percentage change of current -olume a compared to the -olume a certain number of period ago' $he "olume Rate of Change indicator can be ued to confirm price mo-e or detect di-ergence' $he -olume ROC i calculated uing the formula below' 2R3-7&"rrent 2ol"me 4 2ol"me.n periods ago' . 56 / 577 $he -olume rate of change i a powerful indicator when etimating a ecurit%1 abilit% to puh through ke% reitance' A ma(or problem faced while calculating "olume Rate of Change i fi0ing a time period' $he e0act time period calculation are needed for a correct prediction a otherwie the reading then ma% not be conitent with the mo-ement in the market' E. haikan 3scillator &3' A Chaikin ocillator i formed b% ubtracting a +*4da% e0ponential mo-ing a-erage from a @4da% e0ponential mo-ing a-erage of the accumulationIditribution inde0' $hi make it an indicator of an indicator' $he concept of thi ocillator i baed on three main thee' !irst8 if a hare or an inde0 i higher when it cloe than it wa during the da% ;%ou can calculate the a-erage -alue a 7ma0Jmin8 I )<& it mean that it wa a da% of accumulation' "ice -era& if a hareC cloing price i lower than the a-erage le-el of the da%& it mean that ditribution took place' *econd8 table price growth i accompanied b% increae in trade -olume and trong accumulation of the -olume' A the -olume i like fuel that feed market growth& the lag of -olume along with the growth of price how that there inCt enough fuel to continue the rie' %hird8 with a ChaikinC ocillator %ou can trace back the -olume of mone% reource coming in to the market and lea-ing it' Comparing the d%namic of -olume and price allow finding out peak and foundation of the market& both hort and medium4term' !. Williams Acc"m"lation 1istri9"tion &WA1' >illiamC A. i the accumulated um of poiti-e KaccumulationalK and negati-e KditributionalK price mo-ement' 2or e0ample& if the current cloing price i higher than the pre-iou one& >IA. increae b% the difference between the current cloing price and the true minimum' #f the current cloing price i lower than the pre-iou one& >IA. decreae b% the difference between the current cloing price and the true ma0imum' $he term KaccumulationK denote a market controlled b% purchaer and the term KditributionK mean that eller control the market' .i-ergence between the indicator and the price are ignal' #n other word& when a di-ergence appear& the price change it direction according to the indicator' #f the price reache a new ma0imum& but the accumulationIditribution indicator cannot reach a new ma0imum& it mean that the ecurit% i ditributing itelf' #t i a ignal for ell' #f the price reache a new minimum& but the accumulationIditribution indicator cannot reach a new minimum& it mean that the ecurit% i accumulating' #t i a ignal for bu%' >e calculate 55& RS# and MA for opening price& cloing price and high price of the ecurit%' #n thi wa%& there are total of +L feature initiall% in the model' Neural network training i a comple0 tak and increaing the number of feature in the network lead to poor performance of the network' $hi alo lead to the confuion in the weight of the neuron in the network' $hu decreaing the dimenionalit% of the training e0ample that i limiting the number of feature in the network i important' Principal Component Anal%i ;PCA< i one uch method' .ata mut be normalized before uing them in the network or running principal component anal%i' So that each feature contribute e6uall% in the training of the network' >e ober-e that i0 PCA feature are able to catch more than LL'?M of the -ariation in data' $hu& we chooe firt i0 PCA feature a training input to the NARX network' $he PCA anal%i on different ecurit% i hown in $able + below' $able + 3 PCA Anal%i on Stock NTPC.NS Reliance.NS TCS BO INFY.NS 0.7786 0.7972 0.7839 0.7576 0.9241 0.9476 0.9293 0.9111 0.9601 0.9779 0.9651 0.9597 0.9781 0.9848 0.9829 0.9861 0.9863 0.9908 0.9906 0.9934 0.9938 0.9954 0.9958 0.9963 0.9971 0.998 0.9989 0.9988 0.9989 0.9994 0.9996 0.9996 0.9995 0.9997 0.9998 0.9998 0.9997 0.9998 0.9999 0.9999 0.9998 0.9999 0.9999 0.9999 0.9999 0.9999 1 0.9999 0.9999 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 #"' M!$9!.O=OD: Our methodologie in-ol-e two phae' #n the firt phae we elect the principal feature a ummarized in 2ig' +' All the technical indicator are calculated uing MA$=A5 predefined function a-ailable in the financial toolbo0' $he ne0t tep i to remo-e all the not a number ;NaN< and infinit% -alue from the data' $hi re6uire remo-ing the entire time tamp of the data containing uch -alue in an% of the feature' Ne0t& we normalized thi cleaned data uing the formula' >e run PCA anal%i on thi normalized data and calculate k4-ariation core for different -alue of A where A i number of feature' $he k4-ariation core return the -ariation in the reduced A4dimenion data againt the original normalized data calculated in the pre-iou tep' $he following MA$=A5 formula i ued to calculate k4-ariation core' =atent i the component -ariation in the data' "ms"m i MA$=A5 function to calculate cumulati-e um of the data' $he abo-e formula i -ector containing a man% element a the number of feature in the input data' >e elect a -alue of A that retain at leat LL'? M of the -ariation in the data' #n the ne0t tep& we elect thee i0 principal feature a input to our NARX neural network model' 2igure + 3 Preparing the input for the network #n phae two& we imulate -ariou different model of NARX network to determine the bet parameter for the model' $he NARX network ue the pat -alue of the actual time erie to be predicted and pat -alue of other input ;like currencie of other nation and technical indicator in our cae< to make prediction about the future -alue of the target erie' $here are two configuration in the NARX network a< erie 3 parallel b< parallel network hown in 2ig' ) 2ig' )'a 3 Serie Parallel NARX Network 2ig' )'b 3 Parallel NARX Network >e chooe Serie Parallel NARX network to train the network' 9owe-er& we ue Parallel NARX network to predict future tock price' >e ue the following initial configuration a hown in $able )' $able ) 3 Network Parameter A. 1etermining n"m9er of ne"rons >e initiall% tart with one neuron in the network' $hi parameter i indicated a hidden la%er ize in the $able )' >e iterati-el% increae the hidden la%er ize' "alidation et ratio i alo et to zero to enable the network to reach minimum gradient or mean 6uared error' =e-enberg 3 Mar6uartdt back propagation algorithm initialize the neuron with random -alue' 9owe-er& network ma% not alwa% train to the Network Parameter Value Input Delays 1:2; Feedback Delays 1:2; Hdden !aye" #$e 1 Data D%s&n 'and&( )"an 'at& 70 )est 'at& 30 *aldat&n 'at& 0 )"an Funct&n +e",&"(ance Funct&n -#. -a/(u( e""&" 1.00.0015 !e%enbe"1 2 -a"3ua"dt backp"&pa1at&n minimum gradient and leat mean 6uared error in +*** epoch' $hu& we take @* ample for each of thee model and ue the model which ha leat teting performance and leat earl% prediction performance' #n our model& we ue mean 6uared error ha performance parameter' A model with leat performance -alue i bet' $hi a-oid o-er training of the network' B. R"n time performance of Parallel training #n our final anal%i& we train the model parallel% and find the run4time performance impro-ement in training of the model' >e ue a local MA$=A5 pool to train the model in parallel' $hi allow u to cale the training phae to more proceor uing a cluter profile acro the compute node in local network' >e ued ) core #ntel i? BN*M clocked at )HH* M9z proceor with BD5 ram to run thi parallel model and clock the runtime performance of mode' "' O5S!R"A$#ON AN. R!S/=$S >e ran our model on +* different ecurit% namel% N$PC'NS& R!=#ANC!'NS& $CS'5O& #N2:'NS& MAR/$#'5O& >#PROS='5O& "#.!OCON'5O& $A$APO>!R'5O& #OC'NS& DOOD' Our reult were conitent acro all thee ecurit%' A mentioned in ection B& the firt tep i input data preparation phae for network which in-ol-e running PCA on raw data' Apart from ober-ing -er% low -ariance on +L feature of raw data& we alo ober-e that training the network with +L feature i highl% computationall% inteni-e tak' >e did run time anal%i b% training the network with +L feature from raw data againt training the network with onl% H principal component accounting LL'L?M -ariabilit% ;refer to $able +<' $hi clearl% indicate the importance of electing the right feature for training the network a hown in $able @ and 2ig' @' $able @ 3 Run time anal%i of raw data feature -eru PCA feature 2igure @ 3 Run time anal%i of +L raw feature -eru H PCA feature Our imulation how that hidden la%er ize of O to L i good for training NARX neural network for bet earl% prediction performance for tock price a ummarized below in $able B' $able B 3 .etermining Optimum Number of Neuron >e meaure the performance a a mean 6uared error' 9ence lower the performance of the network& better the future prediction of the network' #t i e-ident from 2ig' B that the network training performance keep decreaing a the number of neuron are increaed' 9owe-er& a threhold occur at neuron ize of N& after which tet performance increae ignificantl%' $hi indicate that the network i o-er trained and hence produce more error' $he ame pattern i ober-ed for the earl% prediction performance' Such pattern can be called a the elbow pattern and we chooe the lowet point in the elbow' Note that the chart i upide down' 2igure B 3 Number of Neuron "eru Performance ;MS!< Ne0t& we anal%ze the train4tet ratio optimum for training the NARX network' Our anal%i how that ne0t da% prediction of the tock price i bet with more number of training e0ample' $able ? how that a the number of training e0ample are increaed& the earl% prediction performance of the network decreae' 9ence& tock price forecat are better' Neurons Train Test Early Predict 1 37.205 54.147 42.291 2 40.9179 41.411 41.0659 3 37.9771 47.6813 40.8903 4 36.602 45.3633 39.2321 5 36.0442 45.4374 38.8641 6 38.7358 37.1853 38.2704 7 34.0347 46.3367 37.7278 8 36.7988 38.4801 37.3035 9 34.0765 45.5366 37.5169 10 23.0856 68.7214 38.7343 Neurons Run Time wit!out PC"# Run Time wit! $ PC" %eatures 1 9.31 5.44 5 82.98 19.8 10 297.14 57.13 20 1500.38 206.51 $herefore& we ugget to adapt the network with the latet tock price data after firt training of network' $he time to ad(ut the weight i much le than retraining the whole network' $able ? 3 $rain $et Ratio "eru $rain $et Performance #n our earlier anal%i& we did not et the -alidation et' 9owe-er& -alidation et can ignificantl% impro-e the performance;decreae me< of the network' "alidation et help control o-er training of the network' #n the final part of our anal%i& we tr% to determine the parameter for parallelim in NARX neural network' >e ued inbuilt parallel tool bo0 of MA$=A5 to imulate our network' Although& neural network training i embarraingl% parallel& we ober-ed that the parallelim in MA$=A5 implementation of the neural network onl% lightl% depend on input ize of the network a hown in $able H' #t i independent of the number of neuron in the network' $he run time of the -ariou batch ;input izeIneuron< in the table i minimum of the thirt% imulation for each batch to minimize the randomization in-ol-ed in the training of the network' Pre-iou work in 7N8 7L87+*8 ugget that DP/ baed parallelim for network training cale better than CP/ baed parallelim' $he author in 7L8 ha-e hown a heterogeneou implementation to achie-e parallelim that outperform all other method' $able H 3 Runtime of erie and parallel implementation of NARX network "#' CONC=/S#ON AN. 2/$/R! >ORAS #n thi paper& we anal%zed and tried to find the optimum -alue of number of neuron and tet4train4-alidation et for the NARX neural network' Our anal%i ugget that for ingle tock price prediction& network hould be trained with O4L neuron in the hidden la%er' >e alo tudied the effect of feature ize b% training our network with original feature ;N< and A ;A P N< principal component containing at leat LL'L?M of the -ariation in feature' #t i worth mentioning that the time re6uired to train network with A principal component & in thi cae A Q H& i much le than training the network with original data feature' >e alo ober-e that the parallel implementation of the NARX neural network on CP/ doe not impro-e the run time performance ignificantl%' 9ence& we need to parallelize the training algorithm on DP/ or heterogeneou en-ironment' $hu& we ha-e been able to implement an optimized NARX network for tock price prediction' #n future& we plan to implement a Neural Network baed polic% claification %tem for a market egment in order to impro-e the tock price prediction' $he network etting and parameter will be baed on thi tud%' ACANO>=!.DM!N$ $he author would like to thank .r A'N' Ahan& Profeor& "#$ /ni-erit% Chennai Campu and M $e% Mathew& Ph. Scholar& "#$ /ni-erit% Chennai Campu' >e would alo like to thank .r Andrew ND& Stanford /ni-erit%& for offering a MOOC Coure on Machine =earning at Courera' R!2!R!NC!S 7+8 !' !' Peter& Chao and order in the capital marketE A new -iew of c%cle& price& and market -olatilit%& John >ile% R Son #nc' ;+LL+<' 7)8 $' Plummer& 2orecating 2inancial MarketE A $echnical Anal%i and the .%namic of Price& New :ork ;+LL+<' 7@8 .%na Seng& 2undamental Anal%i and the Prediction of !arning& #J5M ;)*+)<' 7B8 Cinhg49ue Cheng& 2undamental Anal%i of Stock $rading S%tem uing Claification $echni6ue ;)**O<' 7?8 .' R' Coulon& $he #ntelligent #n-etor1 Duide to Profiting from Stock Market #nefficiencie& Probu Publi' Co' ;+LNO<' 7H8 Narendra& Aumpati S' and Aannan Parthaarath%& K=earning Automata Approach to 9ierarchical Multiob(ecti-e Anal%i&K +EEE %ransactions on *ystems, Man and y9ernetics& "ol' )*& No' +& Januar%I2ebruar% +LL+& pp' )H@3)O)' 7O8 cite %ahoo 7N8 /do SeiffertE Artificial neural network on mai-el% parallel computer hardware' Neurocomputing ?OE +@?4+?* ;)**B< 7L8 9onghoon Jang& An(in Park& Aeechul JungE Neural Network #mplementation /ing C/.A and OpenMP' .#C$A )**NE +??4+H+ 7+*8 Rahedur M' Rahman& Ruppa A' $hulairam& Parimala $hulairamanE Performance Anal%i of Se6uential and Parallel Neural Network Algorithm for Stock Price 2orecating' #JD9PC @;+<E B?4HN ;)*++< 7++8 5rown& Stephen J'& >illiam N' Doetzmann& and Alok Aumar' K$he .ow theor%E >illiam Peter 9amiltonC track record reconidered'K $he Journal of 2inance?@'B ;+LLN<E +@++4+@@@' 7+)8 Aenneth =e-enberg ;+LBB<' KA Method for the Solution of Certain Non4=inear Problem in =eat S6uareK' Suarterl% of Applied Mathematic )E +HB3+HN' Train&Test Ratio Test 'er%ormance Early Predict Per%ormance 50450 27.1 13.6 60440 27.2 10.9 70430 23.3 7 90410 16.8 1.6842 (ata Set Si)e Serial Run Time Parallel Run Time Ratio O** +@L'++?O ++@'LNOO +')) +*** +L*'NL@N +?+')+?@ +')H +@** )@N'B))O +OL'+NO +'@@ +H** )HB')BON +LN'@@N@ +'@@