Forecasting Stock Price in Tehran's Stock Market Using Evolutionary Strategies

This study aimed to forecast stock prices and daily stock returns of food products manufacturers, accepted in Tehran Stock Exchange, Using evolutionary strategies. The study was divided into two models. First, the time series of 14 variables related to price prediction, for a period of 5 years (2009 till 2013), was extracted to predict stock price. Then, the price for 19 statistical companies was calculated using evolutionary strategy and back propagation algorithm (Algorithm LM). In the used network, the sigmoid function and linear function were used in the middle layer. In order to find the best algorithm for three companies with low, medium, and high observations, their error square’s mean root and coefficient of determination were calculated. After finding the best model to predict stock prices, the RMSE value was matched with R2 and RMSE values of the accumulated moving average distribution (ARIMA).
In the second model, the (R / S) analysis of random and non-random time series was used to predict daily stock returns. The logarithmic functions relating to all companies were extracted. They along with six other variables were used as inputs in the forward multilayer neural network with back propagation error algorithm (Algorithm LM). In order to find the best algorithm used in the research, the quasi-Newton algorithm and LM algorithm were compared according to the RMSE value for all companies. The results of predicting stock price showed that forecasting stock prices is possible using neural networks, the best algorithm to predict the stock price is the 1-1-10-16 algorithm, and an increase in middle layer does not provide satisfactory results. The comparison of the RMSE value and coefficient of determination R2 in both neural network and ARIMA network showed the excellence of neural network. These results were obtained in predicting daily stock returns. First, neural networks have the ability to forecast daily stock returns with appropriate error. Second, the time series of daily stock returns of companies is not a random process and has memory. Third, LM algorithm is considerably better than quasi-Newton algorithm.