Fashion & Beauty

Stocks selected using SOM and Genetic Algorithm based Backpropagation Neural Network gives better returns.

Stocks selected using SOM and Genetic Algorithm based Backpropagation Neural Network gives better returns.
of 5
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
  IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 2, March 2011ISSN (Online):   664   Abstract  Investment in stock market is one of the most popular type of investment. There are many conventional techniques beingused and these include technical and fundamental analysis.The main aim of every investor is to earn maximum possiblereturn on investments. The main issue with any approach isthe proper weighting of criteria to obtain a list of stocks that are suitable for investments. This paper proposes animproved method for stock picking using self-organizing mapsand genetic algorithm based backpropagation neuralnetworks. The stock selected using self-organizing maps and genetic algorithm based backpropagation neural networksoutperformed the BSE-30 Index by about  30.17 % based onone and half month of stock data. Key Words:    Neural Network, Stocks Classification , Technical Analysis,   Fundamental Analysis,   Self-OrganizingMap (SOM), Genetic algorithm based backpropagation neuralnetwork(GA-BPN). 1. Introduction Investments in stock market has long been an attraction in theminds of investors. The major forecasting method used infinancial area is either technical or fundamental. TechnicalAnalysis [1] provides a framework for studying investor  behavior, and generally focuses only on price and volumedata. Technical Analysis using this approach has short-terminvestment horizons, and access to only price and exchangedata. Fundamental analysis involves analysis of a company’s performance and profitability to determine its share price. Bystudying the overall economic conditions, the company’scompetition, and other factors, it is possible to determineexpected returns and the intrinsic value of shares. This type of analysis assumes that a share’s current   (and future) pricedepends on its intrinsic value and anticipated return oninvestment. As new information is released pertaining to thecompany’s status, the expected return on the company’sshares will change, which affects the stock price.   So the   advantages of fundamental analysis are its ability to predictchanges before they show up on the charts. Growth prospectsare related to the current economic environment. Due to thefact that stock markets are affected by many highlyinterrelated economic, political and even psychologicalfactors that interact with each other in a very complex fashion,it is very difficult to forecast the movement in the stock market. The purpose of this paper is to develop a method sothat investors get maximum returns in short period of time.Stocks have been selected by us on the bases of fundamentalanalysis criteria. These criteria are evaluated for each stock and compared in order to obtain a list of stocks that aresuitable for investments. Stocks are selected by applyingcommon criteria on the stocks listed on Indian National Stock Exchange (NSE).After selection of stocks using fundamentalanalysis, classification of selected stocks is done in to fixednumber of classes by Self-Organizing map. Each of the classis having its own properties; stocks having properties closer toa particular class get assigned to it. Among the classifiedstocks we then select stock for investments using geneticalgorithm based backpropagation neural network. 2. Stocks Classification Stocks are often classified based on the type of company it is,the company’s value, or in some cases the level of return thatis expected from the company. Some companies grow faster than others, while some have reached what they perceive astheir peak and don’t think they can handle more growth. Insome cases, management just might be content with the levelof business that they’ve achieved, thus stalling to make movesto gain further business. Before investing in a particular company, it is very important to get to know the company ona personal level and find out what the company’s goals andobjectives are for the short and long term. In order to prosper in the world of stock investing, a person must have a clear understanding of what they are doing, or they shouldn’t bedoing it at all. Stocks can be a very risky investment,depending on the level of knowledge held by the person(s) Stocks selected using SOM and Genetic Algorithm based Backpropagation NeuralNetwork gives better returns. Asif Ullah KhanAsst. Prof. Dept. of Computer Sc. & Engg.All Saints’ College of Technology, BhopalT. K. BandopadhyayaProfessor, Bansal Institute of Science and Technology,BhopalSudhir SharmaProfessor, RGPV,Bhopal  IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 2, March 2011ISSN (Online):   665 making the investment decisions. Below is a list of classifications which are generally known to us- GrowthStocks, Value Stocks, Large Cap Stocks, Mid Cap Stocks, andSmall Cap Stocks. Stocks are usually classified according totheir characteristics. Some are classified according to their growth potential in the long run and the others as per their current valuations. Similarly, stocks can also be classifiedaccording to their market capitalization. The classificationsare not rigid and no rules are laid down anywhere for their classification. We classified stocks by taking in account theShareholding Pattern, P/E Ratio, Dividend Yield, Price/Book Value Ratio, Return on Net worth (RONW), Annual growthin Sales, Annual growth in Reported Profit After Tax, Returnon Capital Employed (ROCE) and Adjusted Profit After TaxMargin (APATM) with Self-Organizing Map. 3. Stock Market Index A stock market index is a method of measuring a stock market as a whole. Stock market indexes may be classed inmany ways. A broad-base index represents the performanceof a whole stock market — and by proxy, reflects investor sentiment on the state of the economy. The most regularlyquoted market indexes are broad-base indexes comprised of the stocks of large companies listed on a nation's largest stock exchanges, such as the American Dow Jones IndustrialAverage and S&P 500 Index, the British FTSE 100, theFrench CAC 40, the German DAX, the Japanese Nikkei 225,the Indian Sensex and the Hong Kong Hang Seng Index.   Movements of the index should represent the returns obtained by "typical" portfolios in the country. Ups and downs in theindex reflect the changing expectations of the stock marketabout future dividends of country's corporate sector. When theindex goes up, it is because the stock market thinks that the prospective dividends in the future will be better than previously thought. When prospects of dividends in the future become pessimistic, the index drops. 3.1 Composition of Stock Market Index The most important type of market index is the broad-marketindex, consisting of the large, liquid stocks of the country. Inmost countries, a single major index dominates benchmarking, index funds, index derivatives and researchapplications. In addition, more specialised indices often findinteresting applications. In India, we have seen situationswhere a dedicated industry fund uses an industry index as a benchmark. In India, where clear categories of ownershipgroups exist, it becomes interesting to examine the performance of classes of companies sorted by ownershipgroup. We compared BSE-30 SENSEX with the stock selected using SOM and GA-BPN. We choose BSE-30SENSEX for comparison because   SENSEX is regarded to bethe pulse of the Indian stock market. As the oldest index inthe country, it provides the time series data over a fairly long period of time (From 1979 onwards). Small wonder, theSENSEX has over the years become one of the most prominent brands in the country. SENSEX is calculated usingthe "Free-float Market Capitalization" methodology. As per this methodology, the level of index at any point of timereflects the free-float market value of 30 component stocksrelative to a base period. The market capitalization of acompany is determined by multiplying the price of its stock  by the number of shares issued by the company. This marketcapitalization is further multiplied by the free-float factor    todetermine the free-float   market capitalization. The base periodof SENSEX is 1978-79 and the base value is 100 index points. This is often indicated by the notation 1978-79=100.The calculation of SENSEX involves dividing the Free-floatmarket capitalization of 30 companies in the Index by anumber called the Index Divisor. The Divisor is the only link to the srcinal base period value of the SENSEX. It keeps theIndex comparable over time and is the adjustment point for allIndex adjustments arising out of corporate actions,replacement of scrips etc. During market hours, prices of theindex scrips, at which latest trades are executed, are used bythe trading system to calculate SENSEX every 15 secondsand disseminated in real time. Table 1: List of companies of BSE-30 index 4. Application of Neural Networks in Stocks 4.1 Overview The ability of neural networks to discover nonlinear relationships [3] in input data makes them ideal for modelingnonlinear dynamic systems such as the stock market. Neuralnetworks, with their remarkable ability to derive meaningfrom complicated or imprecise data, can be used to extract patterns and detect trends that are too complex to be noticed by either humans or other computer techniques. A neuralnetwork method can enhance an investor's forecasting ability[4]. Neural networks are also gaining popularity in forecastingmarket variables [5]. A trained neural network    can be thoughtof as an expert in the category of information it has beengiven to analyze. This expert can then be used to provideBSE-30 A.C.C., JAIPRAKASH ASSOCIATS, BHARTI TELEVENTURES,BHEL, CIPLA LTD, DLF LTD, GRASIM IND, GUJARATAMBUJA CEMENT, HDFC, HDFC BANK, HINDALCO,HINDUSTAN LEVER, ICICI BANK, INFOSYSTECHNOLOGIES, ITC LTD., LARSEN & TOUBRO,MAHINDRA & MAHINDRA, MARUTI UDYOG, NATIONALTHERMAL POWER, ONGC, RANBAXY LAB., RELIANCE,RELIANCE COMMUNICATIONS, RELIANCE ENERGY,SATYAM COMPUTER, STATE BANK OF INDIA, TATACONSULTANCY, TATA MOTORS, TATA STEEL, WIPROLTD.  IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 2, March 2011ISSN (Online):   666  projections given new situations of interest and answer "whatif" questions. Traditionally forecasting research and practicehad been dominated by statistical methods but results wereinsufficient in prediction accuracy [6]. Monica et al’s work [7] supported the potential of NNs for forecasting and prediction. Asif Ullah Khan et al. [8] used the back  propagation neural networks with different number of hiddenlayers to analyze the prediction of the buy/sell. Neuralnetworks using back propagation algorithms having onehidden layer give more accurate results in comparison to two,three, four and five hidden layers. 4.2 Kohonen self-organizing map Self-organizing maps (SOM) belong to a general class of neural network methods, which are nonlinear regressiontechniques that can be applied to find relationships betweeninputs and outputs or organize data so as to disclose so far unknown patterns or structures. It is an excellent tool inexploratory phase of data mining [9]. It is widely used inapplication to the analysis of financial information [10]. Theresults of the study indicate that self-organizing maps can befeasible tools for classification of large amounts of financialdata [11]. The Self-Organizing Map, SOM, has established its position as a widely applied tool in data-analysis andvisualization of high-dimensional data. Within other statisticalmethods the SOM has no close counterpart, and thus it provides a complementary view to the data. The SOM is,however, the most widely used method in this category, because it provides some notable advantages over thealternatives. These include, ease of use, especially for inexperienced users, and very intuitive display of the data projected on to a regular two-dimensional slab, as on a sheetof a paper. The main potential of the SOM is in exploratorydata analysis, which differs from standard statistical dataanalysis in that there are no presumed set of hypotheses thatare validated in the analysis. Instead, the hypotheses aregenerated from the data in the data-driven exploratory    phaseand validated in the confirmatory    phase. There are some problems where the exploratory phase may be sufficientalone, such as visualization of data without more quantitativestatistical inference upon it. In practical data analysis problems the most common task is to search for dependencies between variables. In such a problem, SOM can be used for getting insight to the data and for the initial search of potentialdependencies. In general the findings need to be validatedwith more classical methods, in order to assess the confidenceof the conclusions and to reject those that are not statisticallysignificant. In this contribution we discuss the use of the SOMin searching for dependencies in the data. First we normalizethe selected parameters and then we initialize the SOMnetwork. We then train SOM to give the maximum likelihoodestimate, so that we can associate a particular stock with a particular node in the classification layer. The self-organizingnetworks assume a topological structure among the cluster units [2]. There are m cluster units, arranged in a one or twodimensional array: the input signals are n-dimensional. Fig. 1shows architecture of self-organizing network (SOM), whichconsists of input layer, and Kohonen or clustering layer. Fig.1: Architecture of Kohonen self-organizingmap The shadowed units in the Fig. 1 are processing units. SOMnetwork may cluster the data into N number of classes. Whena self-organizing network is used, an input vector is presentedat each step. These vectors constitute the “environment” of the network. Each new input produces an adaptation of the parameters. If such modifications are correctly controlled, thenetwork can build a kind of internal representation of theenvironment. Fig. 2: A one-dimensional lattice of computingunits. The n-dimensional weight vectors w 1  , w 2  , …,w m   are used for the computation. The objective of the clustering for each unitis to learn the specialized pattern present on different regionsof input space as shown in Fig. 2. When an input from such aregion is fed into the network, the corresponding unit shouldcompute the maximum excitation. SOM may distinctly reducemisclassification errors [12]. Kohonen’s learning algorithm isused to guarantee that this effect is achieved. A Kohonen unitcomputes the Euclidian distance between an input  x and itsweight vector  w . The complete description of Kohonenlearning algorithm can be found in [2] and [3]. 5. Genetics Algorithm 5.1 An overview A genetic algorithm is an iterative procedure maintaining a population of structures that are candidate solutions tospecific domain challenges [13]. During each temporalincrement (called a generation), the structures in the current population are rated for their effectiveness as domain  IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 2, March 2011ISSN (Online):   667 solutions, and on the basis of these evaluations, a new population of candidate solutions is formed using specificgenetic operators such as reproduction, crossover, andmutation. Genetic Algorithms (GAs) are search algorithms based on the mechanics of the natural selection process(biological evolution). The most basic concept is that thestrong tend to adapt and survive while the weak tend to dieout. That is, optimization is based on evolution, and the"Survival of the fittest" concept. GAs has the ability to createan initial population of feasible solutions, and then recombinethem in a way to guide their search to only the most promising areas of the state space. Each feasible solution isencoded as a chromosome (string) also called a genotype, andeach chromosome is given a measure of fitness via a fitness(evaluation or objective) function. The fitness of achromosome determines its ability to survive and produceoffspring. A finite population of chromosomes is maintained.GAs use probabilistic rules to evolve a population   from onegeneration to the next. The generations of the new solutionsare developed by genetic recombination   operators Biased Reproduction : selecting the fittest to reproduce Crossover : combining parent chromosomes to producechildren chromosomes Mutation : altering some genes in a chromosome. Crossover combines the "fittest" chromosomes and passes superior genesto the next generation. Mutation ensures the entire state-space will be searched, (given enough time) and can leadthe population out of a local minima.Determining the size of the population is a crucial factor.Choosing a population size too small increases the risk of converging prematurely to a local minimum, since the population does not have enough genetic material tosufficiently cover the problem space. A larger population hasa greater chance of finding the global optimum at the expenseof more CPU time. The population size remains constant fromgeneration to generation.Fitness Function drives the Population toward better solutionsand is the most important part of the algorithm. 5.2. Genetic Algorithm Based BPN NetworkTraining Step1: Randomly generate an initial population of, say, Pstrings of length d: S(0)={ s 1 , ..., s P }  .Step 2: Compute the fitness score f  ( s k  ) of each individualstring s k  of the current population S(t).Step 3: Generate an intermediate population [termed mating pool] by applying the selection operator.Step 4: Generate S(t+1) by applying recombination operators(crossover and mutation) to the intermediate population.Step 5: t:=t+1 and continue with Step 2 until some stoppingcriterion applies [in this case designate the best-so-far individual as the result of the GA]. The first step generates aninitial population S(0), i.e. S(0)={ s 1 , ..., s P }  .In GA eachmember of S(0) is a string of length d that corresponds to the problem coding. S(0) is usually generated randomly, becauseit is not known a priori where the globally optimal strings in    are likely to be found. From this initial population,subsequent populations S(1), ..., S(t), ... will be computed byemploying the three genetic operators of selection(reproduction), crossover and mutation. After calculating therelative fitness for all the strings in the current populationS(t) (Step 2), selection is carried out and then strings in thecurrent population are copied (i.e. duplicated) and placed in   the intermediate population proportional to their fitnessrelative to other individuals in the population. After selectionhas been carried out the construction of the intermediate population is completed. Then crossover and mutation areapplied to the intermediate population to create the next population S(t+1) (Step 4). Crossover and mutation provide ameans of generating new sample points in  While partially preserving   distribution of strings across hyper planes whichare observed in the intermediate population. Crossover is arecombination mechanism to   explore new regions in    The   two new strings, called offspring, are formed by the juxtaposition of the first part of one parent and the last part of the other parent. Continue with Step 2 until some stoppingcriterion applies to find final population. Then final weightsare determined from it for backpropagation algorithms. 5.3. Genetic Algorithm Based BackpropagationNeural Network Organizations Hybrid approach offer strong advantage over either rule based or unaided neural network approaches [14]. Geneticalgorithm based back propagation neural network offer goodgeneralization abilities although it may be difficult todetermine the optimal network configuration and network  parameters. Genetic algorithms based backpropagation neuralnetworks gives higher prediction accuracy in comparison of  backpropagation neural networks [15].The architecture of GA based neural network used is asfollows:1). Input layer with 2 nodes2). One hidden layer with 2 nodes3). Output layer with one node. 6. Experimental Results The system has been developed and tested on Windows XPoperating system .We have used Visual Basic and MicrosoftAccess as front end and back end tool.   Simulation data wassourced from Indian National Stock Exchange (NSE).Fromthe 2007 compendium of Top 500 Companies in India weselected 100 companies as per their Shareholding Pattern, P/ERatio, Dividend Yield, Price/Book Value Ratio, Return on Net Worth (RONW), Annual growth in Sales, Annual growth  IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 2, March 2011ISSN (Online):   668 in Reported Profit After Tax, Return on Capital Employed(ROCE) and Adjusted Profit After Tax Margin (APATM),with these inputs SOM divides them into different classes. Asthe SOM are more relevant to the problem where stocks of different companies are to be compared on some common parameters and arranges in the form of different classes. Outof these classes we compared stocks belonging to the bestclass with genetic algorithm based backpropagation neuralnetworks.Normalisation is a key part of data pre-processingfor neural networks and should enable more accurate predictable rates. Normalised data is used for traininggenetic algorithm based backpropagation neural network. Wenormalize inputs so that input values lies between 0 and 1. Input attributes should be carefully selected to keep thedimensionality of input vectors relatively small [16]. As weknow close rate and volume are primary quantitative factorsfor individual equities and from quantitative factors the keyqualitative factor of the market sentiment can be derived. Sowe used close rate and volume of stocks as our input ingenetic algorithm based backpropagation neural network andnext stock rate as our target for training networks. GA-BPNis trained on data set of classified stocks for the years of 1-Aug- 2005 to 30-Jul- 2007 after training, testing is done ondata set of 31-Jul- 2007 to 31-Oct- 2007.Classified stocks arecompared using GA-BPN. Selected stock using GA-BPN isthen compared with BSE-30 index for the period 12/11/2007to 01/01/2008 i.e near about one and half month data.. Wehave found that our selected stock gives 30.17% more returnsin comparison to BSE-30 Index as shown in fig. 3. Fig. 3: Comparison chart between BSE-30 Indexand stock selected using SOM and GA-BPN. 7. Conclusion This paper compares the performances of the stock selectedusing self-organizing maps and genetic algorithm based backpropagation neural network with BSE-30 Index. Theresult shows that the performance of stock belonging to the best class among the classes generated by self-organizingmaps and best prediction accuracy on test data using geneticalgorithm based backpropagation neural network givesmaximum return on investment. Stock selected using SOMand GA-BPN gives 30.17% more returns in comparison toBSE-30 Index References [1] Mizuno, H., Kosaka, M., Yajima, H. and Komoda N.,“Application of Neural Network to Technical Analysis of Stock Market Prediction”, Studies in Informatic and Control , 1998, Vol.7, No.3, pp.111-120.[2] Haykin, Simon, “Neural Networks: A ComprehensiveFoundation”,  Macmillian College Publishing Company , New York,1994.[3] Phillip D. Wasserman, Van Nostrand, "Neural Computing:Theory and Practice", Van Nostrand Reinhold  , New York, 1989.[4] Youngohc yoon and George swales, “Predicting stock price performance: a neural network approach”,  IEEE publishing , 1991.[5] Shaikh A. Hamid, “Primer on using neural networks for forecasting market variables”, in  proceedings of the conference at school of business , Southern New Hampshire university, 2004.[6] Ramon Lawrence, “Using Neural Networks to Forecast Stock Market Prices”, Course Project  , University of Manitoba Dec. 12,1997.[7] Monica Adya and Fred Collopy, “How Effective are Neural Networks at Forecasting and Prediction? A Review and Evaluation”,  Journal of Forecasting, 1998.[8] Asif Ullah Khan et al., “Stock Rate Prediction Using Back Propagation Algorithm: Analyzing the prediction accuracy withdifferent number of hidden layers”, Glow gift  , Bhopal, 2005.[9] Juha Vesanto and Esa Alhoniemi, “Clustering of the Self-Organizing Map”,  IEEE Transactions on Neural Networks , Vol. 11, No. 3, May 2000.[10] Serrano, C., “Self Organizing Neural Networks for FinancialDiagnosis",  Decision Support Systems Elsevier Science , 1996, Vol17, July, pp. 227-238.[11] Tomas Eklund, “Assesing the feasibility of self organizing mapsfor data mining financial information”,  ECIS, June 6–8, 2002,Gdansk, Poland.[12] Egidijus Merkevicius, Gintautas Garsva, “Forecasting of creditclasses with the self organizing maps”,  Informacines Technologies(ISSN 1392 – 124X) Ir Valdymas, 2004, Nr.4(33).[13] D. E. Goldberg, “Genetic Algorithms in Search, Optimizationand Machine Learning.” New York: Addison-Wesley, 1989.[14] K. Bergerson and D. Wunsch, “A commodity trading model based on a neural network- expert system hybrid”,  IJCNN-91-Seattle International Joint Conference, Volume I  , Issue 8-14 Jul1991, Page(s): 289 – 293.   [15] Asif Ullah Khan et al., “ Comparisons of Stock Rates PredictionAccuracy using Different Technical Indicators with Backpropagation Neural Network and Genetic Algorithm Based Backpropagation Neural Network”, pp. 575-580, 978-0-7695-3267-7/08 $25.00 © 2008 IEEE DOI 10.1109/ICETET.2008.59. [16] H. White, “Economic prediction using neural networks: Thecase of IBM daily stock returns”, in  Neural Networks in Finance and  Investing , chapter18, pages 315–328, 1993.
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!