Predicting Indian Stock Market Using Artificial Neural Network Model. Abstract



Similar documents
FORECASTING MODEL FOR AUTOMOBILE SALES IN THAILAND

Bullwhip Effect Measure When Supply Chain Demand is Forecasting

Financial Data Mining Using Genetic Algorithms Technique: Application to KOSPI 200

Kyoung-jae Kim * and Ingoo Han. Abstract

PERFORMANCE COMPARISON OF TIME SERIES DATA USING PREDICTIVE DATA MINING TECHNIQUES

Ranking of mutually exclusive investment projects how cash flow differences can solve the ranking problem

A panel data approach for fashion sales forecasting

CHAPTER 22 ASSET BASED FINANCING: LEASE, HIRE PURCHASE AND PROJECT FINANCING

A Strategy for Trading the S&P 500 Futures Market

Modeling the Nigerian Inflation Rates Using Periodogram and Fourier Series Analysis

The Term Structure of Interest Rates

UNDERWRITING AND EXTRA RISKS IN LIFE INSURANCE Katarína Sakálová

Studies in sport sciences have addressed a wide

Capital Budgeting: a Tax Shields Mirage?

Managing Learning and Turnover in Employee Staffing*

1/22/2007 EECS 723 intro 2/3

A New Hybrid Network Traffic Prediction Method

4. Levered and Unlevered Cost of Capital. Tax Shield. Capital Structure

Combining Adaptive Filtering and IF Flows to Detect DDoS Attacks within a Router

Exchange Rates, Risk Premia, and Inflation Indexed Bond Yields. Richard Clarida Columbia University, NBER, and PIMCO. and

Why we use compounding and discounting approaches

3. Cost of equity. Cost of Debt. WACC.

Testing the Weak Form of Efficient Market Hypothesis: Empirical Evidence from Jordan

A formulation for measuring the bullwhip effect with spreadsheets Una formulación para medir el efecto bullwhip con hojas de cálculo

THE FOREIGN EXCHANGE EXPOSURE OF CHINESE BANKS

Research Article Dynamic Pricing of a Web Service in an Advance Selling Environment

FEBRUARY 2015 STOXX CALCULATION GUIDE

Reaction Rates. Example. Chemical Kinetics. Chemical Kinetics Chapter 12. Example Concentration Data. Page 1

Ranking Optimization with Constraints

Introduction to Statistical Analysis of Time Series Richard A. Davis Department of Statistics

APPLICATIONS OF GEOMETRIC

Chapter 4 Return and Risk

A Queuing Model of the N-design Multi-skill Call Center with Impatient Customers

Teaching Bond Valuation: A Differential Approach Demonstrating Duration and Convexity

An Approach for Measurement of the Fair Value of Insurance Contracts by Sam Gutterman, David Rogers, Larry Rubin, David Scheinerman

Chapter 8: Regression with Lagged Explanatory Variables

REVISTA INVESTIGACION OPERACIONAL VOL. 31, No.2, , 2010

The Norwegian Shareholder Tax Reconsidered

Improving Survivability through Traffic Engineering in MPLS Networks

Mechanical Vibrations Chapter 4

Determinants of Public and Private Investment An Empirical Study of Pakistan

When Two Anomalies meet: Post-Earnings-Announcement. Drift and Value-Glamour Anomaly

Predicting Stock Market Index Trading Signals Using Neural Networks

Modelling Time Series of Counts

Circularity and the Undervaluation of Privatised Companies

14 Protecting Private Information in Online Social Networks

Granger Causality Analysis in Irregular Time Series

COLLECTIVE RISK MODEL IN NON-LIFE INSURANCE

Distributed Containment Control with Multiple Dynamic Leaders for Double-Integrator Dynamics Using Only Position Measurements

Hilbert Transform Relations

Index arbitrage and the pricing relationship between Australian stock index futures and their underlying shares

Monitoring of Network Traffic based on Queuing Theory

APPLIED STATISTICS. Economic statistics

ACCOUNTING TURNOVER RATIOS AND CASH CONVERSION CYCLE

Data Analysis and Statistical Behaviors of Stock Market Fluctuations

Can Individual Investors Use Technical Trading Rules to Beat the Asian Markets?

ON THE RISK-NEUTRAL VALUATION OF LIFE INSURANCE CONTRACTS WITH NUMERICAL METHODS IN VIEW ABSTRACT KEYWORDS 1. INTRODUCTION

DBIQ USD Investment Grade Corporate Bond Interest Rate Hedged Index

Duration and Convexity ( ) 20 = Bond B has a maturity of 5 years and also has a required rate of return of 10%. Its price is $613.

SPEC model selection algorithm for ARCH models: an options pricing evaluation framework

A simple SSD-efficiency test

Department of Economics Working Paper 2011:6

NEURAL NETWORKS APPLIED TO STOCK MARKET FORECASTING: AN EMPIRICAL ANALYSIS

DBIQ Regulated Utilities Index

Hedging with Forwards and Futures

Hotel Room Demand Forecasting via Observed Reservation Information

Handbook on Residential Property Prices Indices (RPPIs)

Morningstar Investor Return

Estimating Non-Maturity Deposits

Principal components of stock market dynamics. Methodology and applications in brief (to be updated ) Andrei Bouzaev, bouzaev@ya.

A GLOSSARY OF MAIN TERMS

Hanna Putkuri. Housing loan rate margins in Finland

Experience and Innovation

Vector Autoregressions (VARs): Operational Perspectives

TACTICAL PLANNING OF THE OIL SUPPLY CHAIN: OPTIMIZATION UNDER UNCERTAINTY

Transforming the Net Present Value for a Comparable One

Market Liquidity and the Impacts of the Computerized Trading System: Evidence from the Stock Exchange of Thailand

MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR

TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS

A Heavy Traffic Approach to Modeling Large Life Insurance Portfolios

Nikkei Stock Average Volatility Index Real-time Version Index Guidebook

Stock Price Prediction Using the ARIMA Model

PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE

12. Spur Gear Design and selection. Standard proportions. Forces on spur gear teeth. Forces on spur gear teeth. Specifications for standard gear teeth

On Motion of Robot End-effector Using The Curvature Theory of Timelike Ruled Surfaces With Timelike Ruling

Task is a schedulable entity, i.e., a thread

IDENTIFICATION OF MARKET POWER IN BILATERAL OLIGOPOLY: THE BRAZILIAN WHOLESALE MARKET OF UHT MILK 1. Abstract

Performance Center Overview. Performance Center Overview 1

Derivative Securities: Lecture 7 Further applications of Black-Scholes and Arbitrage Pricing Theory. Sources: J. Hull Avellaneda and Laurence

DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS

Forecasting Sales: A Model and Some Evidence from the Retail Industry. Russell Lundholm Sarah McVay Taylor Randall

Appendix D Flexibility Factor/Margin of Choice Desktop Research

Individual Health Insurance April 30, 2008 Pages

Option Put-Call Parity Relations When the Underlying Security Pays Dividends

Transcription:

Predicig Idia Sock Marke Usig Arificial Neural Nework Model Absrac The sudy has aemped o predic he moveme of sock marke price (S&P CNX Nify) by usig ANN model. Seve years hisorical daa from 1 s Jauary 2008 o 29 h April 2014 are used o rai ad es he models.. Traiig ad esig is performed by usig Muli-Layer Percepro Nework archiecure. The forecasig performaces of he ANN model is accessed by usig back-propagaio eural ework of errors such as Mea Absolue Perceage Error (MAPE), Mea Absolue Error (MAE), ad Normalized Mea Square Error (NMSE).Furher he sudy has used Sig Correcess Perceage (SCP) o deermie he correcess i he predicio of he direcio of he sock marke price (S&P CNX Nify). The sudy fids ha Mea Absolue Perceage Error (MAPE) ad Mea Absolue Error (MAE) a 5-day lag wih sigmoid acivaio fucios are1.792 ad 0.889 ha gives highes accuracy predicios i he model. Furher, he sudy fids ha 82 perce daa aalysis is correcly prediced by he ANN model ad he predicive power of he ework model is more iflueced by he previous day closig price.this sudy is quie useful o ivesors, professioal raders ad regulaors for akig buy, sell ad hold decisio i he marke while makig ivesme decisio. I will also guide ha ANN is a effecive ool i forecasig he daily sock marke prices ad direcio. Keywords- Sock Marke, Arificial Neural Neworks (ANNs), MAPE, MAE, NMSE ad SCP JEL classificaios: G1, G17, C45 Iroducio:

Predicio of sock marke is a subjec of ieres o raders, marke makers, ivesors ad policy makers oday.sock marke is ypically cosidered o be a dyamic, o-liear, complicaed, oparameric, ad chaoic i aure. Hece predicio of sock marke reurs is a impora issue i fiace lieraure. However sock aalyss have bee usig some approaches for predicig sock marke reur. The radom walk heory saes ha sock price flucuaios are idepede of each oher ad price movemes do o follow ay paers or reds. Socks ake a radom ad upredicable pah. The supporer of radom walk believes ha i is impossible o ouperform he marke wihou assumig addiioal risk. As per radom walk heory log-erm buy-ad-hold sraegy is he bes ad idividuals are o required o aemp o ime he markes. O he oher had, he efficie marke heory saes ha i is impossible o bea he marke because whaever iformaio is available abou a sock o oe ivesor is available o everyoe. The price of a securiy is reflecig abou compay's prospecs ad he direcio of ecoomy o everyoe so idividual ivesor cao ouperform he marke. The weak-form efficiecy assers ha fuure sock price cao be prediced by aalysig pas sock price. Semi-srog-form efficiecy proclaims ha all publicly available iformaio is fully refleced i sock prices ad o excess reurs ca be eared by radig o ha iformaio. Srog-form efficiecy asses ha all iformaio is fully refleced i securiies prices. Eve isider iformaio is of o use. O he oher had, echical aalyss assume sock marke price moves i red ad hese reds ca be capured. So, fuure marke price ca be forecased by examiig pas sock prices. Techical aalyss use ools such as charig paers, echological idicaors, ad specialized echiques like Ellio Waves ad Fiboacci series ad believes ha here are recurrig paers i he marke behavior ha are coveioal. Bu mos of he echiques used by echical aalyss are lack of raioal explaaio for heir use. However echical aalysis is criicized due o is idiosycraic aure ad coradicory o Efficie Marke Hypohesis. Fudameal aalys cocered more wih compay raher ha acual sock. The aalys assumes ha share s curre ad fuure price depeds o is irisic value ad is expeced reur. The aalyss make heir decisios based o he pas performace of he compay, is earigs forecas, he paricular idusry secor ad he overall ecoomy. Fudameal raders do o rely o he marke movemes aloe. Fudameal raders believe ha he markes are reacig o eves i cerai ways ad fuure marke prices are forecased based o hese eves. Criics of he heory, however, oppose ha socks do maiai price reds over ime ad i is possible o ouperform he marke by carefully selecig ery ad exi pois for sock ivesmes.this heory saes ha socks always rade a heir fair value ad ispire ivesors o make profi by purchasig udervalued socks or sell socks a iflaed prices. Proagoiss of his heory debaes ha i is meaigless o use fudameal ad echical aalysis o predic reds ad search for udervalued socks. Markes are highly irraioal ad overreaced o he posiive earigs ews ha drives prices up oo high o good ews ad oo low o bad ews. Idividual ivesors ed o overreac or uderreac o ews. I is commoly believed ha ivesors i he marke ed o move wih he majoriy of ivesors seime. So whe marke is ehusiasic abou a sock, he marke s herd mealiy drives prices upward. The buyig volume eds o reach peak ad more people buy a higher price ha a he low ad vice-versa. This creaes a edecy o buy high ad sell low isead of he wiser opposie acio. Hece he iefficiecy of he markes ofe is caused by iefficiecy of ivesors ad raders. Some researchers arguig ha a class of irraioal ivesors has kow as ''oise raders'' move markes successfully. Noise raders are folks who ed o buy ad sell o bad iformaio.as he value of a sock rises or falls, people are iclied

o buy or sell ha sock. This i ur furher affecs he price of he sock, causig i o rise or fall chaoically. Hece i is believed ha he behavior of he sock marke is chaoic, irraioal ad, a imes, absolue iefficie. The profiabiliy of ivesig ad radig i he sock marke largely depeds o he predicabiliy. If he direcio of he marke is successfully prediced he ivesors may be beer guided ad ear good reur. If ay model predics he reds of he dyamic sock marke ad elimiaes uceraiies, i is o oly helpig o icrease he ivesmes i sock markes bu also helpig o he regulaors o ake correcive measures. Sice presece of o-lieariy i sock marke eables he raders o make excess profis, he prese sudy has bee uderake o ivesigae he same. Predicio of sock marke by usig o-liear model i.e. ANN model have o bee exesively researched i he Idia sock marke. Apparely his research would help o ivesors, ecoomiss, marke regulaors ad policy makers i udersadig he effeciveess of Idia sock marke o ake beer ivesme decisio ad devise beer risk maageme ools.i his coex, we have raised hree research quesios. The firs objecive is o predic he sock exchage marke idex values by usig ANN model. Secodly he sudy ries o deermie he predicio accuracy of he model by usig differe forecasig performace measure i.e. Mea Absolue Perceage Error (MAPE), Mea Absolue Error (MAE), Normalized Mea Square Error (NMSE).Thirdly he sudy ries o es he performace of he direcio of he prediced value of he closig of he ex day s CNX S&P Nify Idex by usig Sig Correcess Perceage (SCP) measure. Therefore, he prese work offers a value addiio o he exisig lieraure ad proves o be useful o he ivesors, researchers as well as o regulaors. The srucure of he paper is orgaized as follows: secio 2 discusses he lieraure review, he daa ad mehodology is preseed i secio 3, he empirical ivesigaio is repored i secio 4 ad secio 5 deals wih cocludig observaios. Why Neural Nework used? The direcio of fuure chage of he fiacial marke is really desirable oday. May sudies have esablished ha o-lieariy exis i fiacial marke daa series. Sice liear models are failed o udersad he daa paer ad relaioship bewee ipu ad oupu daa because of he complex, o-liear ad chaos aure of he sock marke, arificial eural eworks are used o derive meaig from complicaed or imprecise daa. They are used o exrac paers ad deec reds ha are oo complex o be oiced by eiher humas or oher compuer echiques.anns have he capaciy of performig oliear modelig wihou a priori kowledge abou he relaioship bewee ipu ad oupu variables. The eural ework is raied from a mass of hisorical daa ad ries o discover hidde depedecies o use hem for predicig io fuure. Neural Nework is a black box which akes some variables as a ipu, ad gives oher as a oupu ad able o lear somehig. I oher words, ANNs ca geeralize ad correcly gaher he usee par of he daa eve if he sample daa coai oisy iformaio. Eve if he uderlyig relaioships amog he daa are hard o describe, ANNs capure he delicae fucioal relaioships amog hem ad provides a pracical feasible way o solve real world problems. The Neural Nework model ca easily predic he direcio of sock marke reur chage whe compared o acual value ad eable o maximizig profis from ivesme radig. Thus ANNs are a more geeral ad flexible modelig ool for forecasig. From saisical poi of view, Neural Neworks are aalogous o oparameric, oliear, regressio model. Neural ework model is appropriae for capurig all he oliear vibra relaioships i he sock marke. So, Neural Nework suis beer ha oher

models i predicig he sock marke reurs. Today eural eworks are more preferred over he oher mehods for sock marke forecasig. Lieraure Review: Usig arificial eural eworks o predic sock markes has bee a acive research area durig las decade. Phua, e.al 2000) uses Neural Nework wih Geeic Algorihm o he Sigapore sock marke ad predics he marke direcio wih a accuracy of 81 perce. Qig Cao e.al. (2005) uses Arificial Neural Neworks o predic sock price moveme for firms raded o he Shaghai Sock Exchage ad compares he predicive power of uivariae ad mulivariae eural ework models. The sudy shows ha Neural Nework ouperform he liear models ad fids ha Neural Neworks are useful ool for sock price predicio i emergig markes like chia. Themozhi (2006) applied Neural Neworks o predic he daily reurs of he Bombay Sock Exchage (BSE) Sesex ad fids ha he predicive power of he ework model is iflueced by he previous day. The sudy shows ha saisfacory resuls could be achieved by applyig MLP o predic he BSE Sesex. Asif Ullah e.al. (2008) compared a Back propagaio Neural Nework ad Geeic Algorihm based back propagaio Neural Nework ad shows ha for sock rae predicio, Geeic Algorihm based back propagaio Neural Nework gives more accurae predicio. BirolYildiz e.al(2008) developed a efficie hree layer Neural Nework wih revised Back propagaio Algorihm o predic he direcio of Isabul Sock Exchage Naioal-100 Idices (ISE Naioal -100) ad fids ha he model forecas he direcio of ISE Naioal -100 o a accuracy of 74.51 perce.h. Al-Qaheri, e.al (2008), Bruce & Gavi (2009) Mira (2009) Alay ad Sama (2005) forecasig sock marke volailiy usig Neural Neworks ad fids ha ANNs are beer ha he resuls deermied by liear ad logical regressio models.zhao e al. (2009) predics he sock price usig BPNN by cosiderig a sigle closig price as he ime series vecor. The auhors uses wo seps forecas approach. Firs, use Gray correlaio aalysis o choose he se of variable which ca describe he characerisics of he sae of he sock marke from a umber of echical idicaors. The classify he sae of sock marke by he Self-orgaizig feaure map (SOFM) ework. Ad based o his classificaio, BPNN is used for predicio. The resul shows ha he predicive accuracy of SOFM-BP model was beer ha ha of he radiioal BPNN model. Leadro ad Ballii(2010)aalyze Neural Neworks for fiacial ime series forecasig o predic fuure reds of Norh America, Europea, ad Brazilia sock markes ad fids ha ANNs do ideed have he capabiliy o forecas he sock markes ad, if properly raied, robusess ca be improved, depedig o he ework srucure. S. M. Alhaj Ali, e.al. (2011) uilizes arificial eural ework i he modelig of sock marke exchage prices. The resuls fids ha he use of ANN provides fas covergece; high precisio, ad srog forecasig abiliy for real sock prices which i ur will brig high reur ad reduce poeial loss o sock brokers. Tripahy ( 2011) forecasig he ex day s close value of Sock price by usig ANN Model ad ARIMA Model i Idia Sock Marke. The sudy fids ha ANN model is very useful for predicig sock marke price ha he ARIMA models i Idia. Pee.,A.I. e al., (2012)predic he Nigeria sock marke by usig Arificial Neural Nework ad fids ha arificial eural ework ca be used o predic fuure sock prices. Abbas (2012) predics Tehra sock price usig Arificial Neural Nework for aual daa from 2000 o 2008. The sudy shows ha esimaio ad predicios of Tehra sock price wih Arificial Neural Nework is possible ad gives sroger resuls. Daa ad Mehodology:

The daa employed i he sudy cosiss of daily closig prices of S&P CNX Nify Idex. The required ime series daily closig prices of idices have bee colleced for a period of 7 years from 1 s Jauary 2008 o 29 h April 2014 from www.seidia.com. For forecasig, feed-forward wih back-propagaio eural ework archiecure has bee used. These mehods ormally provide bes resuls as compared o oher ANN archiecures. Selecio of ipu variable for he eural ework model is a criical facor for he performace of he eural ework because i coais impora iformaio abou he complex oliear srucures of he daa. The criicaliy i selecig he ipu variables lies i selecig he umber of ipu variables ad he lag bewee each. Wih less lag bewee ipu, he correlaio bewee he lagged variable icreases which may resul i a over-fiig pheomeo. O he oher had, wih icrease i he lag bewee each ipu variable, he eural ework may lose ou esseial iformaio of ipu variables, resulig i uder-learig. To hadle wih his dilemma of over fiig or uder-learig ad selec a opimal srucure, he sudy has cosidered various lagged srucure (muliple lag ipu variable wih differe lag bewee each) ad es he performace of he eural ework o a rial ad error basis. The ipu variables seleced for his model are he lagged observaio of he closig prices of Nify Idex. As eural eworks are paer recogizers, he qualiy of he daa uses i he sudy largely iflueces he accuracy of he resul. The ipu daa uses i he eural ework model faciliae de-redig of he daa so as o faciliae proper ework learig process. I order o improve he performace of he ework a oliear scalig mehod is adoped. The closig price of Nify idex is scaled wih liear scalig fucios. Furher, if more ha wo hidde layers are ake, here is a chace ha ANN model migh creae over-fiig i.e. he i-samplig error may become very low ad ou-sample error may go up. So he umber of hidde layers is ake i our sudy is 2. The umber of euros i oupu layer is equal o he umber of oupus required. I our sudy, here is oly oe oupu which is he esimaed Nify Reurs. Hece, he umber of oupu layers is 1.Afer he eural ework model is cosruced; raiig of he eural ework is he ex esseial sep of he forecasig model. The resul of he raiig process of he ework depeds o he algorihm used for he purpose.the predicio is carried ou by usig differe acivaio fucios i.e. Sigmoid ad Ta Hyperbolic fucios. The predicios sysem predics he ex day s value usig he above closig price Nify daa. I eural ework, daa are ormally ormalized io rage of (0, 1) or (-1, 1) accordig o he acivaio fucios of euros. I our paper he value of Nify is ormalized io he rage of (0, 1) ad eworks are raied ad esed usig he back propagaio algorihm. Though ideal predicio i sock marke is difficul oe, everheless eural ework is used o predic reasoably good predicio i umber of cases. I is advisable o cosider boh shor erm (oe day lag) ad log-erm (muli-lag) predicios o ge healhier resuls. I oe lag predicio, he ex day value is expeced o be based o acual pas day value bu o he oher had a few prediced values are also used o predic fuures values. Arificial Neural Nework Model: The basic srucure i he huma brai is called euro. All euros have four basic compoes, which are dedries, soma, axo ad syapses (coecios). A euro coais dedries. Through dedries euro is able o receive sigals from oher cells. The axo carries ougoig sigals from he cell. Where he ougoig axo mees he dedrie of aoher euro, a coecio is made i he form of a elecrochemical device called a syapse. Basically, a

biological euro receives ipus from oher sources, combies hem i some way, performs a geerally oliear operaio o he resul, ad he oupu he fial resul. Arificial euros are much simpler ha he biological euro. A Arificial Neural eworks process iformaio i a similar way he huma brai does. Every ANN cosiss of a se of euros ad a se of coecios bewee hem. The coecio bewee he euros is aribued by a weigh. The iformaio se o he euro ad muliplied by correspodig weighs is added ogeher ad used as a parameer wihi a acivaio fucio. If hese sigals are sufficie, he euro becomes acivaed. Each euro akes a umber of ipus ad afer processig wih a acivaio fucio called rasfer fucio yields a disic oupu. The arificial euro oupu a value based o ipus. The mos commoly used rasfer fucios iclude, he hard limi, he pure liear, he sigmoid ad he a sigmoid fucio. Fig 1 shows a graphical represeaio of a arificial euro. Nework srucure wih ipus (x1, x2,..xi) are beig coeced o euro j wih weighs (w1j, w2j,...w ij) o each coecio are show i Figure 1, The euro sums all he sigals i receives he each sigal is muliplied by is associaed weighs o he coecio. This oupu (hj) is he passed hrough a rasfer (acivaio) fucio, g (h), ha is ormally o-liear o give he fial oupu Oj. The mos commoly used fucio is he sigmoid (logisic fucio) because of is easily differeiable properies which is very coveie whe he backpropagaio algorihm is applied. The overall ipu o he euro is calculaed by a a w x i i 0 i

Where xi represes he ipus o he euro ad wi represes he weighs of he euro. I sigifies ha he higher he weigh of he coecio more ifluece o he euro. I Neural Nework bias plays a impora role. Someimes he hreshold is called a bias value. The hreshold is a real umber ha is subraced from he weighed sum of he ipu values.so we ake he bias as a ipu wih value 1 ad is correspodig weigh is he sum of he average of he oher ipu weighs.( where w0 = θ ad x0 = -1) To ormalize his sum io a sadard rage, fucios called hreshold fucios, (sigmoid fucios beig he mos widely preferred oe) are used. Threshold Fucio is a simple fucio ha compares he summed ipus o a cosa, ad depedig o he resul, may reur a -1, 0, or 1. The acivaio fucio f is ypically of sigmoid form ad may be a logisic fucio or hyperbolic age A sigmoid fucio is defied Logisic fucio f 1 1 e a a Hyperbolic fucio f e a a a a e a e e The Neuro Oupu (y): The arificial euro compues is oupu accordig o he equaio show below. This is he resul of applyig he acivaio fucio o he weighed sum of is ipus, less he hreshold. This value ca be discree or real depedig o he acivaio fucio used.the; he oupu of he euro is defied by y f i 1 w i x i f i 0 w i x i θis called he bias of he euro, sice θ he bias is cosidered as a ipu, x0 = 1 ad he associaed weigh w0 = θ. The quaiies Xi ad Wi deoes he ipus ad weighs respecively. I is almos always he case ha a euro oupu a value bewee [0, 1] or [-1, 1] The ework is composed of a large umber of highly iercoeced processig elemes called euroes. These euros are disribued i few hierarchical layers. Mos of he eural eworks are hree layered: ipu, middle or hidde ad oupu. Arificial euros are arraged i hese layers. Feed-forward eworks were firs sudied by Rosebla (1961). Geerally here is o daa processig is happeig a ipu layer. The ipu layer akes he ipus ha feed io he ework ad passes o he middle layer. There are more ha oe middle layers/hidde layers. Hidde layers are he followed by a oupu layer. Arificial euros reside i hidde layers. I middle (hidde) layer, all complexiy resides ad compuaios are made here. The resuls are achieved i oupu layers. I feed forward eworks all coecios are uidirecioal. The layers of arificial eural ework are show i fig-2.

Fig 2 Layers of arificial eural ework Machie learig approach is based o he priciple of learig from raiig ad experiece so i is appealig for arificial ielligece model. ANNs are well suied for machie learig where coecio weighs adjused o improve he performace of a ework. The closig price of Nify idex used as ipu of a radiioal ANN for he raiig ad predicio process. I raiig module, i ivolves Mulilayer (MLP), radom samplig of daa se ad he mai core of his module is Back propagaio Neural Nework raiig. MLP is a feed forward Neural Nework raied wih error Back propagaio which is a wo-pass weighlearig algorihm o adjus he weighs i-bewee odes o reduce error of he oupu. The archiecure of MLP is cosisig of Ipu layer, hidde layer ad oupu layer. Durig raiig, back propagaio is he process of back propagaig errors hrough he sysem from he oupu layer owards he ipu layer. I is ecessary o use back propagaio sice hidde uis have o raiig arge value ha ca be used. So hey mus be raied based o errors from previous layers. The oupu layer is he oly layer which has a arge value for which o compare. As he errors are back propagaed hrough he odes, he coecio weighs are chaged. Traiig occurs uil he errors i he weighs are sufficiely small o be acceped. I has observed from some sudies ha he sigmoid fucio works bes whe learig is approached owards average behavior ad hyperbolic age fucio works bes whe learig deviaio from he average. Performace Measureme: To measure he performace of he eural ework model, he sudy used Mea Absolue Perceage Error (MAPE), Mea Absolue Error (MAE), Normalized Mea Square Error (NMSE) ad Sig Correcess Perceage (SCP). These ess are used for evaluaig he predicio accuracy of he model. MAPE compue he mea error value.mape measure he residual errors. This ormally gives a idea of he differece bewee he acual ad prediced

value. The lower he MAPE values he beer he model. The followig equaio shows he process o calculae he MAPE MAPE 100 1 o o p Where N is he oal umber of es daa, O is acual sock price o day ad P is forecas sock price o day The mea absolue error is a average of he absolue errors. The loger MAE meas higher bias level ad less accurae forecas o predic prices. Bu however MAE is a suiable model o predic sock marke flucuaios. Smaller values of hese measures shows more correcly prediced oupus. MAE 1 o p Where N oal umber of es daa, O isacual sock price o day ad P is forecas sock price o day Normalized Mea Square Error is used for evaluaig he predicio accuracy of he model. The lower he value beer is he model. NMSE 1 1 o ( o p p ) 2 2 Where Ois he acual sock price of he daa series, P is he prediced value for he same days closig price ad p is he mea of acual value of sock price. The correcess i predicabiliy of he direcio of he moveme of he idex is preferred over he correcess of he magiude of he idex. Predicio of he direcio of he sock marke is more impora ha he value of he idex. If he accuracy of he direcio of he predicio of he idex is high ad reliable, he loss i a porfolio reurs ca be miimized. I order o es he performace of he direcio of he prediced value of he closig of he ex day s Nify Idex, Sig Correcess Perceage (SCP) is used. SCP for he sample period uder sudy gives he perceage of correcess i he predicio of he direcio of he idex. The direcio of chage is calculaed by subracig oday s price from he forecas price ad deermiig he sig (posiive or egaive) of he resul. The perceage of correc direcio of chage forecass is correspodig o he perceage of profiable rades eabled by he ANN model.

SCP 100 1 D Where D=1 if (O - P) (O - P )>0 or D=0, oherwise I his case, he model wih he higher SCP more accuraely predics he markes movemes. Empirical Aalysis: The forecasig ool ANN has bee used o forecass he Nify Idex value. To obai forecased value, he daily closig price is used as a ipu o his model. This model he auomaically accusomed wih he raiig daases ad makes forecas, give he curre day s sock prices. ANN models wih differe ework parameers are creaed, raied ad esed for each series is preseed i able-1. For he Nify Idex, he sudy has ru eural ework simulaios for 10 differe cases based o daa ipu (differe lags) ad he acivaio fucio (Hyperbolic age ad Sigmoid). N Sice he behaviour of sock marke is more radom, he sudy has cosidered 5 day lag daa as ipu for he eural ework so ha learig ca be maximum ad higher precisio ca be achieved. IFTable-1 Predicio Performaces of ANN model Sl No Case Acivaio Fucio MAPE MAE NMSE (%) Sig Correcess Perceage Sl No 2.7002 1.315 0.767 71.33 Hyperbolic Tage 1 Oe Day 2 Lag Sigmoid 2.636 1.248 0.338 82.25 3 4 5 6 Two Day Lag Three Day Lag Hyperbolic Tage 2.836 1.379 2.237 68.03 Sigmoid 2.393 1.181 0.798 74.25 Hyperbolic Tage 2.414 1.174 1.589 71.90 Sigmoid 2.0001 0.961 0.494 72.72

Hyperbolic Tage 2.883 1.435 3.386 60.60 7 8 Four Day Lag Sigmoid 1.830 0.903 1.132 72.91 Hyperbolic Tage 2.153 1.071 2.351 67.48 9 Five Day 10 Lag Sigmoid 1.792 0.889 1.734 69.54 The followig fig-1ad 2 shows he MAPE.MAE, NMSE ad SCP calculaed for he forecas of Nify Idex value by usig four forecasig echiques. Figure 1: Predicio Performaces of ANN model (Hyperbolic Tage fucio) Figure 2: Predicio Performaces of ANN model (Sigmoid fucio) 100 ANN Predicio 80 60 40 82.25 74.25 72.72 72.91 69.54 MAPE MAE NMSE 20 SCP 0

1 44 87 130 173 216 259 302 345 388 431 474 517 560 603 646 689 732 775 818 861 904 947 990 1033 1076 1119 1162 1205 1248 1291 1334 1377 1420 1463 1506 1549 I is observed form able-1 ha error is miimised a 5-day lag for predicio. For boh MAPE ad MAE ework models, he highes accuracies predicios are obaied wih 5 day lag period for he sigmoid acivaio fucio are 1.792 ad 0.889 respecively.. So 5-day lag value shows beer predicios for forecasig he ex day s close value of Nify Idex. The able-1 also exhibis ha he miimum value of ormalised error is occurrig whe daa for 1 day lag is used alog wih a sigmoid acivaio fucio. Sig correcess perceage (SCP) has bee used o correcly predic he direcio of moveme of Nify Idex. The SCP is maximum for 1 day lag ad sigmoid acivaio fucio. Thus i is observed ha 1 day lag is producig he bes prediced resul for he Nify Idex over he 7 year daa period. This shows ha learig for he ework is highes uder 1 day lag daa which is opimum o predic he fuure values of he Nify Idex price. The aalysis idicaes ha 82 perce daa aalysis is correcly prediced by he ANN model.so i is recommed for usig he sigmoid acivaio fucio whe predicig he behaviour of a sock marke price. Fig-1 is showig he graphical represeaio of prediced ad acual sock marke price of ANN based sock marke price predicio where back propagaio algorihm is used. The resuls achieved i boh he cases are equally accurae. Fig-2 Acual ad Prediced Nify Sock Idex values for 2 hidde layers 15000 10000 5000 0 1-Day lag Nify Acual Nify Prediced Fig- 3Acual ad Prediced Nify Sock Idex values for 2 hidde layers

The fig-3 ad 4 depics he predicio of Nify sock Idex value ad i is observed ha ANN based sysems perform quie saisfacory. I is also observed ha feed forward ework usig Back propagaio is reasoable for sock marke price predicio. Coclusio: This sudy has bee uderake wih he objecive of fidig he bes model for he predicio of Idia sock marke (Nify Idex) price wih Arificial Neural Nework. The sudy has used hisorical prices of he idex value for predicio. I is also foud from he aalysis ha he prediced oupu is very close o he acual daa. Amog ANN models, SCP performace measures is foud o be more appropriae for he predicio. The ANN model accuraely predic he direcio of moveme wih predicio rae of 82 perce i he daa aalysis which is quie good effecs. Accurae predicio of sock marke price moveme is highly sigifica opic for ivesors oday. I will also help o raders, ivesors ad professioals i heir decisio makig process of buy, sell or hold posiio. Sice sock markes plays a impora role i ecoomy, ANN model ca help o ivesors, professioal raders o esimae he sock marke price ad apply he radig sraegy ha will help hem o miimise risk ad maximise profi i he sock marke.sock marke are highly volaile ad are affeced by may ierrelaed ecoomic ad poliical facors across he globe, so he prese sudy will provide some impora ipu o he policy makers o ake measures by usig ANN model o predic he sock marke. Oe limiaios of he sudy is ha we have used 1571 daa for raiig purpose. Hece ANN model ca give far beer resuls if more daa ca be used durig raiig. The secod limiaio of sudy is ha we have oly used he hisoric prices of he Idex values for predicio. Oher macro-ecoomic facors ad oher ieraioal sock marke daa as ipu variables ca also be used as ipu i order o improve he accuracy of he model Refereces: Abbas Vahedi,(2012), The Predicig Sock Price usig Arificial Neural Nework, Joural of Basic ad Applied Scieific Research, Vol. 2, Issue 3,pp.2325-2328 Asif Ullah Kha, Badopadhyaya T.K. ad Sudhir Sharma (2008), Geeic Algorihm Based Back propagaio Neural Nework Performs beer ha Back propagaio Neural Nework i Sock Raes Predicio, Ieraioal Joural of compuer sciece& ework securiy, Vol. 8, No. 7 pp. 162-166 Bruce Vasoe & Gavi Fiie, (2009), A empirical mehodology for developig sock marke radig sysems usig arificial eural eworks, A Ieraioal Joural of Exper Sysems wih applicaios, Vol. 36, Issue 3, pp.6668-6680 Birol Yildiz, Abdullah Yalama, ad Mei Cosku (2008), Forecasig he Isabul Sock Exchage Naioal 100 Idex Usig a Arificial Neural Nework World Academy of. Sciece, Egieerig ad Techology, Vol.46, pp.35-41 H. Al-Qaheri, A. E. Haassaie, ad A.Abraham, Discoverig sock price prediciorules usig rough ses, Neural Nework World, Vol. 18, pp.181-198

Mira, S. K., [2009] Opimal Combiaio of Tradig rules usig Neural Neworks Ieraioal Busiess Research,Vol2, No 1, pp 86-99. M. Majumder ad A. Hussia, MD (2010) Forecasig of Idia Sock Marke Idex Usig Arificial Neural Nework, seidia.com, pp. 1-21 Phua, P.K.H. Mig, D., Li, W. (2000), Neural Nework wih Geeic Algorihms For Socks Predicio, Fifh Coferece of he Associaio of Asia-Pacific Operaios Research Socieies, 5h - 7h July, Sigapore Peer Adebayo Idowu, Chris Osakwe, Aderoke Ahoia Kayode, Emmauel Roimi Adaguodo (2012), Predicio of Sock Marke i Nigeria Usig Arificial Neural Nework, Ieraioal Joural of Iellige Sysems ad Applicaios, Vol.4, No.11, pp.68-74 Leadro S. Maciel, RosagelaBallii(2010), Neural Neworks Applied o sock marke Forecasig: A Empirical Aalysis, Joural of he Brazilia Neural Nework Sociey, Vol. 8, Issue 1, pp. 3-22 M. Themozhi, (2006) Forecasig Sock Idex Reurs Usig Neural Neworks, dbr.shr.org, Vol. 7, No. 2, pp. 59-69 Quig Cao, Kary B.Leggio ad Marc J.Schiederjas (2005), A compariso bewee Fama ad Frech's model ad Arificial Neural Neworks i Predicig he Chiese Sock Marke,Compuers & Operaios Research, Vol. 32, Issue 10, pp. 2499 2512 Rosebla F. (1961), Priciples of Neuro dyamics: percepros ad he heory of brai Mechaisms. Spara Press, Washigo. S. M. Alhaj Ali, A. A. Abu Hammadb, M. S. Samhouria, ad A. Al-Ghadoora(2011) Modelig Sock Marke Exchage Prices Usig Arificial Neural Nework: A Sudy of Amma Sock Exchage Jorda Joural of Mechaical ad Idusrial Egieerig, Volume 5, No. 5, pp. 439 446 Tripahy.Nalii( 2011), A Compariso of Arificial Neural Nework (ANN) Model & Auo Regressive Iegraed Movig Average (ARIMA) Model For Forecasig Idia Sock Marke Ieraioal Joural of Coemporary Busiess Sudies, Vol. 2, No 3, pp.2156-7506