Muscat Securities Market Index (MSM30) Prediction Using Single Layer LInear Counterpropagation (SLLIC) Neural Network Louay A. Husseien Al-Nuaimy * Department of computer Science Oman College of Management and Technology Barka, Sultanate of Oman Loay.alneimy@omancollege.edu.om Abstract The mankind has higher interest to find accurate tools for prediction especially at the financial sector. Many forecasting methods have been suggested through the history ranging from linear regression methods to nonlinear time series fitting approaches. Emerging countries market like security market in Sultanate of Oman has more importance in the international scenario according to its wide opportunities in investment sector. The MSM 30 index is used to control and benchmark the prices tendency for shares listed in Muscat security market. Predicting the future values of MSM 30 is very important for the investment sector in Oman. This research proposed the use of Single Layer LInear Counter propagation (SLLIC) neural network as a forecasting tool for MSM 30 index, and the resulted network are called SLLIC MSM 30 network. The performance of the SLLIC MSM 30 was tested and compared with other prediction models used with financial time series like linear regression and Radial Basis Function neural network. The test results shows that SLLIC MSM 30 network has good approximation and prediction capabilities for the MSM 30 index. Keywords Time series prediction, MSM30 prediction, Artificial Neural Networks, Single Layer Linear Counterpropagation network. I. INTRODUCTION The discrete values for an interest variable can be recorded and then it can be used to represent the observations sequence of time series related to that variable of interest. Time series are very important research subject for a large area of knowledge ranging from economy and meteorology to epidemiology and medicine. In economy the time series are used to represent data about industrial production, stock prices, and unemployment rate. In meteorology the time series are interesting to collect data about rain precipitation, wind velocity, and temperature [1]. The most important characteristic of the time series is that its observations are time-dependent, which means that the value collect in time t depends on the observation collected in time t-1 [2]. Achieving capital best return is the main goal of investing in securities. All categories of investors (either long term or short term) need to observe and then judge the performance of the market so they can take the appropriate decisions. The market index is an important tool to monitor and judge about an individual company performance or about all the market performance. The importance of the securities market indices come from its ability to determine the market situation yesterday, today, at future, or over any other chosen period of time. Also the market indices are benchmark which can be used to judge the performance of investment portfolio by the investors [3]. In 1992 the Muscat Security Market (MSM 30) index has been established according to 1990 as a base date. The number of companies which has been included within MSM30 increased overtime. Currently the number of companies included at MSM30 index is 30 companies drawn from different sectors of the market. The main aim of recording MSM 30 index as a time series is to determine the price tendency for the shares listed within MSM 30 in an objective approach. The investors in Muscat security market use MSM 30 as a benchmark measurement to scale their opinion about the path of investment [4]. Predicting or forecasting of future values for market indices are very important for the investment sector. Market index can be used as forecasting tool or yardstick to measure the future performance of the shares that compose the index. The market index extending an historical scene of stock market performance, provide investors more insight into their placing decisions. Market indexing can be used as a hand tool by investors who do not know in which stock they should put investments [5]. The prediction techniques which used to forecast the financial time series future values can be used to forecast the market index future values. The time series historical values are recorded as a base structure for future prediction. Most time series prediction researches depend on define the pattern or the structure of the recorded data to build a model that can be used to predicate the future series pattern and use it to estimate the future values of the series. Many models are used to formulate the financial time series; Box & Jenkins (ARIMA) is the most popular one of these linear models. Exponential autoregressive and smooth transition autoregressive as nonlinear models are also used to formulate the nonlinear time series [1]. * Louay A. Husseien Al-Nuamy is with Oman College of Management and Technology, Department of computer Science, Hay Asem, P.O. Box 680, 320 Barka, Sultanate of Oman (e-mail: Loay.alneimy@omancollege.edu.om). 978-1-4673-9584-7/16/$31.00 2016 IEEE
Predicting the market index values is a complicated issue from the theoretical and practical views because the stock markets are influenced by many political and economic factors. According to those complications the classical econometric and statistical models which are used in forecasting the financial time series fails to efficiently treat uncertainty nature of market index time series [6]. Artificial Neural Networks (ANN) can be used as a fantastic powerful tool in prediction and classification, as the modern studies have been shown. The ANN powerfully used historical data that recorded to learn the pattern of the time series, generalize, and then predict the future values. Recent scientific researches demonstrated that ANN has the ability to approximate any time series as a continuous function that means the ANN is a powerful and successful tool for forecasting of financial time series [2]. In this article we apply a specialized type of neural networks in data prediction, called the Single Layer LInear Counter propagation (SLLIC) network, to predict the future values of the MSM 30 index as a time series. The SLLIC network is one of the most ANN that designed especially for time series prediction and function approximation [7]. The performance of SLLIC network which designed to predict MSM 30 will be compare with the performance of the traditional time series prediction methods, and with the other types of ANN to exam the SLLIC network abilities in forecasting MSM 30 future values. The used data series for MSM 30 in this article are yearly values from 1997 until 2002 and monthly from Jan 2003 till Dec 2014 that provided by annual statistical bulletin of Muscat security market [8]. II. MSM 30 INDEX In Sultanate of Oman there is one stock exchange called Muscat Securities Market (MSM). The Royal Decree number 53/88 in 21 June 1988 established MSM in order to control and regulate the Omani securities market. MSM effectively participated with other Omani organizations to build financial sector in the Sultanate. Two royal Decrees numbered 80/98 and 82/98 issued after ten years of great growth at MSM to restructure the market to get more perfect function [3]. Number that represents the relative change of a variable is called an index (e.g. period values, quantity, and prices). To get statistical indication in securities market especially about changes in entire level and direction of prices the securities market index are used. Securities index is a financial instrument to benchmark and measure the stocks, real estate, mutual funds, investment contracts, bonds, commodities, and money markets [5]. Due to importance of indices at the securities markets, Muscat securities market index has been established in 1992 and June 1990 has been chosen as a base date for the index [4]. To weight any securities index there are three major methodologies: price weighting, equal weighting, and value weighting. MSM 30 index used value weighting method to calculate its value [4]. To determine closing prices of shares traded MSM 30 use the volume weighted average price calculation method. The volume weighted average price method reflects the real situation about what happen at the market because it take trading volume in its account and do not take the minute to minute prices only. Representing the price movement for the listed shares in an objective manner and guide the investors in their investment processes are the main aims of MSM 30 index. To achieve its aims MSM 30 enforce three features; includes only freely available shares for trading, wider representation for smaller companies by using 10% capping value, and quarterly basis revision for free float and capping value. MSM 30 has three supplement indices: Industry index, Insurance & Services index, and Banks & Investment index that composed the MSM 30. The supplement indices importance comes from its ability to guide the investors to understand computation methodology, it provides clears selection criteria, and provides maintenance and review rules [3]. The base value of the MSM 30 index was 100 and the value has been updated in 2004 to be 1000 at first of June. According to that change the MSM 30 digits changed from three digits to four digits (e.g. 408.34 points become 4083.40). For the purpose of equities performance the 30 companies that compose of MSM 30 index must be include shares from all Muscat securities market sectors (Industry, Services & Insurance, and Banks & Investment). The selection of the MSM 30 components base on 45% liquidity, 40% market capitalization, and 15% profitability (company performance). The movement of the 30 underlying stocks constantly changes the value of the index. The prediction by the MSM 30 future values aims to elevate the awareness of investors and to achieve a transparency in higher level. III. SINGLE LAYER LINEAR COUNTER PROPAGATION (SLLIC) NETWORK Artificial Neural Networks (ANN) is a modeling system that used widely at the recent years in such cases with no prior knowledge about the actual system. In complex nonlinear dynamic systems the ANN are very useful modeling tool for such behavior [9]. The powerful performance of the ANN comes from its capability to cover and map the nonlinear past input values with output values of the real system in order to predict future values for that system. Many architectures, designs, and learning methods have been suggested by many authors of ANNs to model different types of real systems. In diversity applications type the ANN can be used successfully to solve predication problems. In financial sector ANN can be used to manipulate the forecasting problem for diversity types of time series, especially in stock prices movement prediction and securities market indices [9]. One of the most successful black-box modeling structures is the multilayer feed forward neural network. The multilayer feed forward neural net can be considered as an extension nonlinear structure for the linear dynamic systems [9]. The multilayer counter propagation feed forward neural net is one subtype of the multilayer feed forward sub networks. The experimental tests showed that multilayer counter propagation feed forward neural net produced more accurate results than the other architectures of the multilayer feed forward networks in time series prediction [7]. The original counter propagation
neural network has single input layer, single output layer, and multi hidden layers [11]. One of the best important improvements to the counter propagation neural network was done by merged of Kohonen unsupervised neural network to provide hybrid type of neural nets. Mixing between counter propagation and Kohonen neural networks resulted in a perfect architecture which is called as the Single Layer LInear Counter propagation (SLLIC) network [6]. In SLLIC net the training information (recorded values of the time series) are clustered into number of groups by using the Kohonen network. This division of the original time series into a number of clusters will reduce the complexity of prediction problem. For each cluster the SLLIC network construct a separated feed forward counter propagation network to approximate a function for it. The complete architecture of the Single Layer LInear Counter propagation (SLLIC) network, shown in Fig. 1, consist of a number of clusters, each of these clusters contains single processing element (feed forward network) that approximates a function for the subset of examples associated with that cluster [7]. The Learning algorithm for SLLIC network have been constructed based on the feed forward counter propagation network learning strategy, and it contains of two phases, the first phase for building the Kohonen network with M clusters, while the second phase use to build a feed forward counter propagation net for each Kohonen cluster. The two learning algorithm phases of SLLIC network can be listed as [7]: The meaning of abbreviations used at the algorithm: E: The absolute error of the current cluster, E0: The previous absolute error of the current cluster, Wj: The Jth weight of the processing element, Y : The actual output of the cluster, Y: The desired output, X: The input vector. a: The Learning rate (practical range for a is a = 0.1, 0.2,.., 0.9, 1.0). N: the number of associated examples with the current cluster net. K: counter for the examples associated with the cluster. Phase one {Kohonen Layer} Step 0: Initialized the cluster weight vectors randomly. Step 1: Repeat steps 2 to 7 until the Kohonen weight become stable. Step 2: For each training vector do steps 3, 4, and 5. Step 3: Set the input layer activation s to vector X. Step 4: Find the winner cluster j. Step 5: Vj = Vj + a*x *Vj Fig. 1. SLLIC network Structure Step 6: Reduce the learning rate a. Step 7: Go to step 1. Phase two {feed forward net} Step 0: For each cluster s network do the following steps: Step 1: Initialize the weight vector W randomly. Step 2: E = 0, K = 0 Step 3: K = K+1, X = X k, Y = Y k, Y = Wm Xm, E =Y Y, W = W + a * E*X/ Y, Y = Wm Xm, E= abs(y-y ). Step 4: if K < N then Go to step 3. Step 5: if E> E 0 then Go to step 2. IV. SLLIC NETWORK FOR THE MSM 30 The intense interchange between countries is the main trend of the globalization especially in stock markets. Emergent countries markets like Oman have more importance in the international scenario according to its wide opportunities. Modern stock markets like Oman market depend not only on corporations of finance, but mainly it depends on individual capitalization resources. Most of these individual investors are investing in a portfolio, and they hope to obtain big return with fewer risks [1]. To reach the goals of reducing risks and maximizing investment returns, the investors need to predict by the future values of the market indices. The neural networks are very well suited prediction tool to optimize the procedure of stock markets risk management. This study try to predict the MSM 30 future values by build a SLLIC network that mimics the formal structure of the MSM 30 as a time series. The MSM 30 time series that used
to build the MSM 30 SLLIC network contains 150 value of the MSM 30 index, yearly values from 1997 till 2002 and monthly from Jan 2003 until Dec 2014 that provided by annual statistical bulletin of Muscat security market. That data will be used as learning examples to teach the SLLIC network. Fig. 2 shows the actual values of the MSM 30 index. According to automatic structure determination feature of the SLLIC network model the 150 values of the MSM 30, leads to MSM 30 SLLIC network that needs Kohonen layer constructed with 18 clusters and a feed forward neural net for each cluster. Each feed forward network that associated with each cluster contains three layers (input, hidden, and output). Each feed forward net contains one neuron at the input layer and one at the output layer. The number of neurons at the hidden layer for each feed forward network are vary and chosen through trial and error strategy. Fig. 3 shows the MSM 30 index values produced by the MSM 30 SLLIC neural network. In order to compare the performance of the SLLIC MSM 30 network with the original MSM 30 time series values the Mean Absolute Error(MAR), Root Mean Square Error (RMSE), and correlation coefficient (R) statistical indicator are calculated and the results shown in table 1. The smallest values of MAR and RMSE are the best while R value nearest to one is the best. The MAR and RMSE represent the difference between the actual MSM 30 values and the estimated value by SLLIC MSM 30 network. The R value represents how much there is similarity between the MSM 30 and the SLLIC values, the value one for R represent the highest similarity. Relative absolute error and the Root relative squared error show the error percentage ratio of the SLLIC MSM 30 network relative to the original MSM 30 values deviation [10]. V. EVALUATION OF MSM 30 SLLIC NETWORK Performance measure of the proposed neural network for MSM 30 is very important procedure to appropriate evaluation of this work. In order to have appropriate evaluation, the results produced by SLLIC MSM 30 network, should be compared with the results produced by other approximation models used for predicting financial time series. For the purpose of this evaluation procedure two important methods in financial time series prediction are used; the regression model and Radial Basis Function (RBF) neural network. The GMDH Shell version 3.7.4 software for time series prediction was used to build the evaluation models (RBF network and the regression model). For the regression model table 2 shows the evaluation statistical indicators. And table 3 shows the same indicators for the RBF network. Fig. 4 shows the MSM 30 index values produced by the regression model. While Fig. 5 shows the MSM 30 index values produced by the RBF neural network. The evaluation of the SLLIC MSM 30 network performance is compared with the regression model and the RBF network in Fig. 6. Fig. 2. The MSM 30 index actual values Fig. 3. The MSM 30 SLLIC model TABLE I. EVALUATION CRTRIONS Mean Absolute Error(MAR) 294.7985 Root Mean Square Error (RMSE) 473.2541 Correlation Coefficient (R) 0.982282 Relative absolutee error 1.63809% Root relative square error 3.5178% TABLE II. REGRESSION MODEL EVALUATION Mean Absolute Error(MAR) 764.302 Root Mean Square Error (RMSE) 1076.03 Correlation Coefficient (R) 0.903807 Relative absolutee error 6.70287% Root relative square error 9.43671% TABLE III. RBF NETWORK EVALUATION Mean Absolute Error(MAR) 684.13 Root Mean Square Error (RMSE) 993.417 Correlation Coefficient (R) 0.919237 Relative absolutee error 5.99976% Root relative square error 8.71219%
Fig. 4. The regression model VI. CONCLUSIONS Using neural network approach to solve the problem of financial time series prediction is simpler especially for user with poor statistical and mathematical knowledge since the functional form of the relationship between the independent and dependent variables need to be known. From the test experiments (comparing SLLIC MSM 30 network with the regression model and RBF network), the suggested neural model has comparable and sometimes better performance than the other models due to its capability on clustering the original time series into number of feed forward neural models. The SLLIC MSM 30 network fits the real MSM 30 index time series in a natural approach with less error ratios and higher correlation coefficient as it can be conclude from Fig. 6 and the tables 1, 2, and 3. References Fig. 5. The RBF network model Fig. 6. SLLIC compared to the other models [1] Amorim Neto, M.C, Financial time series prediction using exogenous series and combined neural networks, Neural Networks, 2009. IJCNN 2009. International Joint Conference on Neural Networks IJCNN,, Atlanta GA, 2009, pp.149 156. [2] I. Beliaev, R. Kozma, Time series prediction using chaotic neural networks on the CATS benchmark, Neurocomputing, doi:10.1016/j.neucom.2006, 2007. [3] Frank K. Reilly and Keith C. Brown, Investment Analysis and Portfolio Management, 8th ed, SouthWesternCollege, 2005. [4] Oman Chamber of Commerce and Industry, "The Commercial Companies Law No. 4/1974 and its Amendments", http://om.mofcom.gov.cn/table/gsf.pdf, June 2008. [5] Asarkaya, A., Forecasting Volatility of Istanbul Stock Exchange. 4th International Conference Globalization and Higher Education in Economics and Business Administration (GEBA), 2010. [6] Anupam Tarsauliya, Shoureya Kant, Rahul Kala, Ritu Tiwari, Anupam Shukla, Analysis of Artificial Neural Network for Financial Time Series Forecasting, International Journal of Computer Applications (0975 8887), Volume 9 No.5, pp. 16-22, November 2010. [7] Ghwanmeh S., Al-Shalabi R., Kanaan G., and Al-Nuaimy L., Enhanced Neural Networks Model Based on a Single Layer Linear Counterpropagation for Prediction and Function Approximation, Information Technology Journal, Vol. 5, No. 1, 2006, pp. 45-50. [8] Muscut Securities Market, Annual Statistical Bulletin of Muscat Security Market, www.msm.gov.om, 2015. [9] A. Sheta and R. Hiary, Modeling lipase production process using artificial neural networks, in Multimedia Computing and Systems (ICMCS), 2012 International Conference on, May 2012, pp. 1158 1163. [10] Alaa F., Sara M., A Comparison between Regression, Artificial Neural Networks and Support Vector Machines for Predicting Stock Market Index, (IJARAI) International Journal of Advanced Research in Artificial Intelligence, Vol. 4, No.7, 2015, pp, 50-63. [11] Omer M., Farahat A., Learning Algorithm Effect On Multilayer Feed Forward Artificial Neural Network Performance In Image Coding, Journal of Engineering Science and Technology, Vol. 2, No. 2, 2007, pp.188-199.