Automated Neural-ware System for Stock Market Prediction

Transcription

1 Proceedings of the 2004 IEEE Conference on Cybernetics and Intelligent Systems Singapore, 1-3 December, 2004 Automated Neural-ware System for Stock Market Prediction Arosha. Senanayake School of Enginecring,, Monash University, Malaysia. Abstract- This article uses neural networks in forecasting stock market prices. With their ability to discover patterns in nonlinear and chaotic systems, neural networks offer the ability to predict market directions more accuratcly than current techniques such as technical analysis, fundamental analysis, and regression comparcd with neural network performance. Proposed intclligcnt stock market prediction system is based on the Quantitative and Qualitativc factors. Three feedforward neural models can be used to analyze these factors. Input data to the neural network proposed are quantitative factors. Input data to the neural network proposcd for qualitative factors can be factors related to the political cffcct considered. Third neural network consists of decision integration in which input data will be the outputs of above-mentioned neural networks. This facilitates to make right decision whether stock market is influenced by quantitative or qualitative factors. B. lnrelligent Stock Mcrrkst Prediction factors Related factors collection for the stock market environment. factors 1. INTRODUCTION A. Introduction fur Neural- Ware System The Stock market is always one of the most popular investments due to its high profit. However, higher profit tends to higher risk too. Thus, various research works intended to develop models in order to provide the investors an optimum prediction. Among the traditional research, time series analysis techniques and multiple regression models were used. Recently due to the computational speed, Artificial NeuraI Networks (ANN) has been also used in this area. Through various models have been proposed, they only concentrated quantitative factors. However, in developing countries, like Sri Lanka, sometime non-quantitative factors are more important than qualitative factors. Therefore, proposed intelligent stock market prediction system intends the inclusion of both factors. Therefore, proposed intelligent stock market prediction system intends the inclusion of both factors such that right decision Intelligent stock market prediction is based on the systems integration. Figure 1 represents the block diagram corresponding to the proposed System [ 11. Network (Prediction 3) Artificial Neural Network (Prediction 2) Figure 1 Structure of Stock Market Predictian- Factors Collection In order to make right decision, collecting the effective information regarding the predicted object is crucial. The assumption that makes, is collected factors are good enough to support the prediction model. Quantitative factors The factors that can be considered as qualitative factors are volume, average volume, rate of volume, index in open, index in close, index fluctuation, rate of index changc, financial buying, financial Selling, remaining quota with financing, remaining quota with stock. Turn over rate, other related factors in the Stock market. These are the inputs to the ANN $ IEEE 1166

2 of quantitativc factors. The output to the samc specifies stock tendency performance [l], [2]. Qualitative factors These factors must be collected from the society in consideration. Depcnding on the factors, it is necessary to formulate a qucstionnairc. Answers to the questionnaire will get from the experts and determine minimum, maximum and mean values of factors consider. In this way, sevcral different questionnaires will bc formulatcd and forced to learn the proposed ANN. Therefore, the output of this ANN is the effcct of qualitative factors to the stock market. Decision integration From the above-proposed two ANNs can be obtained gencral stock market tendency and specid factors effect, which are based on the quantitativc and qualitative models rcspectivcly. The overall results can detcrmine the integration of these outputs together with time effect through a third ANN as shown in figure I. ' II INTRODUCTION FOR COLOMBO STOCK EXCHANGE A. The Colombo Stock Exchange (CSE) The Colombo Stock exchangc was established in It is the ccntral, and only authorized market for shares listed for public transaction. These securities are mainly the sharcs of public Iisted companies. There is also a Iimitcd amount of trading in treasury bills and convcrtible debentures. The CSE is not mushroom dcvelopment of a market economy. It traces its history more than hundred years to 1896, when the Colombo sharc market was established under the administration of the Colombo share broker's association, Iatcr called thc Colombo brokers association. The CSE is now situated at World Trade Center, Colombo 01. Like any other exchange the CSE does not itself buy, sell or set price of the share traded on it's trading floor. The prices of these are strictly determined by supply and demand. It provides all facilities for convenient bidding by brokers, similar to a public auction, publishes a hand book of listed companies and othcr material usehl to the investor. The CSE's Central Depository System (CDS) is a hliy computerized system of recording and processing all share transactions. The CDS has a comprehensive database of shares listed on the CSE, and a market information system making it one of the most modem integrated systems of any stock exchange worldwide, which assists in efficient posttrade operation [I]. B. Share price indices Share price index is an important datum to know about current situation of the CSE. The CSE calculates share price index for each and every sector called sectorial index.' It also calculates the All Share Price Index (ASPI) and Milanka Price Index, which are main share indices. All Share Price Index (ASPI) rcflects the share price fluctuations of all companies (n = 240) in the stock market. The base year for ASPI is 1985 and base index was set at 100 points. The CSE computcs ASPl using the following formula. i= 1 Wherc Pli - Prescnt price of the i ' company. Q,, -Total issued quantity of Shares of the i"company Poi - Basc year price of a sharc of the i th company Qui - Base year quantity of a share of the, ith company The CSE introduced the Milanka Price Index (MPI) with effect from 4th January Thc base index was set at IO00 points as at 3Ist December The MPl replaced thc Sensitive Price Index (SPI). The introduction of the MPl was timely for several rcasons. Some of the companies included in the SPI did not continue to mcct the evaluation criteria since the SPI did not continue to meet the evaluation criteria since the SPI was last revised in August Unlike the SPI, thc MPI is not a blue chip index and the criteria taken in to its constructions are size and Iiquidity. The MPI is comprised of 25 companics representing 7 sectors. The index. represents over 10% of thc listed companies and accounts for over 50% of market capitalization of the CSE. C. Qirantitatire Faclors in CSE The factors that can be considered as qualitative factors in CES are Turnover, Trades, Shares Traded, Companies Traded, Companies Listed and Market Capitalization. These are the inputs to the ANN of quantitative factors. Thc outputs to the System are ASPl and SSPI. D. Qualitative Factors in CSE The Share Price lndiccs mainly depends on the following Qualitative factors. 1, Socio Economic conditions of the country. 2. Political situations. 3. Performance of listed companies. 4. Activities of foreign investors. 5. International scenario. Fluctuations in share price of the companies are mainly depending on the above factors. These factors are discussed briefly as follows. Economic growth of countsy plays an important role in CSE. Insufficient infrastructure facilities have been onc of the factors that have affected the competitiveness of Sri Lankan products and constrained faster economic expansion in Sri Lanka. The conflict in Northeast has curtailed the development of key sectors due to resource constraints. Inflation (depreciation of money) and weathcr conditions, which determine the Socio Economic Conditions, do have a direct impact on price fluctuations of shares. 1167

3 The poiicy of ruling govemmcnt makes a drastic change in Stock Market activity. Privatization of the sectors fiscal policy (public revenue), monetary policy and economy policy of the government play an important role in Stock Market activities. The pcrfomance of listed companies depends on many aspects such as values of shares traded, frequency of shares traded, and volume of shares traded, Market Capitalization, movement in sector indices etc. Foreign investor interest in the Colombo Stock Market has been inextricably linked to the Indian Market. The indirect investments of foreign investors through Colombo Stock Exchange make substantial price movements of the shares. The Colombo Stock Exchange is very much dependent on international sccnario because it is a member of the FIBV (International Federation of Stock Exchanges). For example, the Colombo stock markct commenced the year 1998 amidst the aftcrmath of the East Asian Contagion, and plunged almost immediately into a second dilemma precipitated by the nuclear explosions by India and Pakistan. As because of those reasons All Share Price index declined by 15% and the Sensitivc Pricc Index by 14% in In 2001 October also ASP1 and SPI are affccted because of attack of American Buildings. i I 111 DESIGN ANN FOR STOCK MARKET PKEDlCTlON The specification and design of an ANN application should aim to produce the best system and overall performance. Much of the work is done in the initial data preprocessing and feature extraction.as possible to reduce the task of the network. A number of different types of ANN can be used in the same application [ 11, [2], 131. ANN design involves five main tasks: 1) Data collection 2) Selecting Inputs and Outputs 3) Raw data preprocessing 4) Selection of an ANN type and topology 5) ANN training, testing and validation Initially the problem is specified and an ANN model is chosen. Next suitable data. are collected and preprocessed before the features are extracted to make the input'output vector pair sets. These vector sets are then used to train and test the ANN. If the results are satisfactory the ANN is finally validated as shown in figure 2. A. Data collection All necessary data's are collected from Colombo Stock Exchange. They gave daily based data for ALSH, SSPI, Daily Turnover, Number of Shares Traded, Number of Trades, Market Capitalization, Tumover ratio and Sectorial indices. Also gave monthly based data for Domestic and Foreign Total Turnover, Domestic and Foreign Trades, Domestic and Foreign shares traded, Companies Traded, Companies Listed, All Share Price Index, Sensitive / Milanka (since 1999), Market Capitalization, Market PER, Dividend Yield and Market PBV. Figure 2 ANN Design Proccdurc for Stock market prediction B. Selecting Inputs and Outputs The ability of neural nctworks to discover nonlinear relationships in input data makes them ideal for modeling nonlinear dynamic systems such as the stock market. Various neural network configurations have been developed to model the stock market. Often these networks use raw data and derived data from technical and fundamental analysis. One of the most important factors in constructing a neural network is deciding on what the network will leam. The challenge is determining which indicators and input data will be used, and gathering enough training data ta train the system appropriately. The input data consider here is daily change. A neural network must be trained with large amount of pattern. Each example pattern consist set of inputs and outpurs. These pattems are collected from Colombo Stock Exchange. Inputs from CSE Number of Shares Traded (Xli), Number of Trades (X2i), Market Capitalization (X3 i) Outputs from CSE _j All Share Price Index (Y li), Milanka Price Index (Y2,) C. Raw dofa preprocessing Quantitative factors arc modified for the Stock market Prediction. New Inputs and outputs are formulated by following steps. I168

4 Srtpl Calculate the different between Today s values and Yesterday s values. dxl, XI, - XI,., dx2, X2, - X2,-1 dx3, X3, - x3,1 dy 1, = Y 1 I - Y 1 dy2, = Y2, - Y2i-1 step2 Inputs Assign new values to the inputs of the training data. These values in the rangc of -1.0 to 1.0 are called normalizcd values of input. Spanncd and normalize the inputs for each output. dyi-1 dxi New Input Range b 0.5 t i b I.O Case 1 dy*, => -t dxi => - dyi.l/dxi New Weight Range O- 0.9 I Case 2 dy,, => + dxi => + dykltdxi New Weight Range Case 3 dycl => - dq => - dyblldxi New Weight Range Case 4 dyll => - dxi => + dyi-ltdxi New Weight Range o Oulputs Assign new values to the output of the training data. It is important that the target values (desired response) be chosen within the range of sigmoid activation function. These values in the range of -I.O to 1.O are called normalized valuc of output. New Value = dyi/maximum value of last ten values - dyi I max(dyi, dyi.1, dyi-2,... dyi-28, dyi-29, dyi-30) Thc patterns arc divided into two Training Set and Testing set. Comparatively training set is larger than thc testing set. Training set is used to train the network. Testing sct is used test the performance of the network. Patterni + (Input li, Input2i. Input3i. Input.?,, InputSi, Inputd;,Output I,,Output2i) 1V SELECTION OF AN ANN TYPE ANI1 TOPOLOGY A. ArriJiciof Ncirml Network Type The most common network architecturc used is the backpropagation network. Howcver, stock market prediction networks have also been implemcnted using genctic algorithms, recurrent nctworks, and modular networks. Backpropagation networks are the most commonly used network becausc they offer good generalization abilities and are relatively straightforward to implement. Although it may be difticutt to determine the optimal nctwork configuration and network parameters, these networks offer very good performance when trained appropriately [I], [Z], [3]. B. Artificial Neural Network Topology To find the optimal number of neurons in the hidden layer, network architecturre is changcd for the each training and observe the mean squarc error on the network. For the anatizysis (6-8-2) and (6-6-2) architecturs are seleted. Both are trained with 1750 patterns and 0.2 learing rate. Training outputs are shown bellow. When Comparing two ouputs, MSE of (8-8-2) architecture is higer than the that of architecure(8-6-2). Second one is more suitable for training because it?s minimize error with minimum number of ncurons (Figure 3). Smallest number of hidden neurons that increase the performance of the network. So optimal number of hidden neuron is six. Arcbitecture(6-6-2) Input layer: 6 Neurons Output layer: 2 Neurons Hidden layer: one hidden layer with 6 Neurons Number of Training patterns: I750 pattern Training Algorithm: Std. Backpropagation Number of epochs: Learning Rate:

5 Cycles (Table II) Input layer: 6 Neurons Output layer: 2 Neurons Hidden layer: one hidden layer with 6 Neurons Number of Training pattcms: 1750 pattern Training Algorithm: Std. Backpropagation Number of epochs: I00000 Learning Rate: 0.2 TABLE II CONVERGENCE OF ANN FOR CYCLES SSE I MSE I SSe/o-units I 1 epoch 1 Figure 3. ANN training for the topology I. I4538 O.2986G I To find the optimal learning-rate paramctcr (q), nchvork architecture is trained with various learning-rate paramctcr. It varies from 0 to 1.0. Network architecture (6-6-2) is trained with 1750 pattcms. Leaming rate parameter starts from 0.2 and each time it increased by 0.2 up to 1.0. MSE of three networks is observed, training network with learning-rate 0.2 is converged morc than nchvork with learning-rate 0.4, as well as network with learning-rate 0.8. Also network with learningratc 0.4 is converged little bit more than network with learning-rate 0.8. Learning rate 0.2 is more suitable for the network. Increasing the teaming-rate lead to converge the network quickly but is not stable (Table 1) Learning rate parameter Input layer: 6 Neurons Output layer: 2 Neurons Hidden layer: one hidden layer with 6 Neurons Number of Training patterns: 1750 pattern Training A Igorithm: Std. Backpropagation. Number of epochs: Learning Rate: 0.2 TABLE I CONVERGENCE OF ANN FOR LEARNlNG RATE 0.2 V. ANN TRAINING AND VALIDATION A neural network must be trained on large amount of input data. Training a network involves presenting input pattems in a way so that the system minimizes its error and improves its performance. The training algorithm may vary depending on the network architecture, but the most common training algorithm used when designing financial neural networks is the backpropagation algorithm. Training occurs until the errors in the weights are sufficiently small to be accepted. The major problem in training a neural network is deciding when to stop training. Since the ability to generalize is fundamental for these networks to predict future stock prices, overtraining is a serious problem. Overtraining occurs when the system memorizes pattems and thus looses the ability to generalize. It is an important factor in these prediction systems as their primary use is to predict (or generalize) on input data that it has never seen. Overtraining can occur by having too many hidden nodes or training for too many time periods (epochs). However, overtraining can be prevented by performing test and train procedures or crossvalidation. The test and train procedure involves training the network on most of the pattems (usually around 90%) and then testing the network on thc remaining pattems. The network s performance on the test set is a good indication of its ability to generalize and handle data it has not been trained on. If the performance on the test set is poor, the network configuration or learning parameters can be changed. The network is then retrained until its performance is satisfactory. Cross-validation is similar to test and train except the input data is divided into k sets. The system is trained on k-1 sets and tested on the rcmaining set k times (using a different test 1170

6 set each time). The amount of training data is also important. The system was trained on 1750 patterns and tcsted on 100 pattems. Each pattcm corresponded to the input vafucs on a particular day. VI. CONCLUSIONS A network s pcrformance is often measured on how well the systcm predicts market direction. A backpropagation network which use Number of Trades (Xlj), Number of Shares Traded (X2,) and Markct Capitalization (X3J as inputs. Then modify the inputs and give it to the trained network. The output produced by the system should be rcnormalizing into ASP1 and MPI. These predicted valucs are comparing with real values and calculate the error. Mean squarcs error between the desired and network response for the testing sets is calculated. By using this error, network performancc will be observed. Thc mean squares error should on average get progressively smaller as the network learns. At that point the evaluation data set should be run through the network and the mean squares error checked to see that it is acceptably low and that the network accuracy is acceptably high. If so, then the network design has been compteted and the network can bc put into service. If not, it may be necessary to alter the number of hidden nodes and repeat thc process. If it is stiil not acceptable it may be necessary to improve thc preprocessing and feature extraction from the raw data or increase the training data set size. REFERENCES [ 11 Senanayake, Arosha, Efficient Neural Stock Market Prcdiction System, Proceeding of International Conference on Information Technology, Colombo, Sri Lanka, [2] Thawomwong, S., D. Enke, and C. Dagli Neural Networks as a Decision Maker for Stock Trading: A Technical Analysis Approach, Joumnl of Smart Engineering Systems Design, Vol. 5 (2003): [3] Thawomwong, S., and D. Enke, Forecasting Stock Retums with Artificial Neural Networks, Chapter 3 in Neural Networks in Business Forecasting, edited by Peter Zhang (2003):