Optimal neural network architecture for stock market forecasting

Predicting stocks accurately has always intrigued the market analysts. A possible forecast of stocks is done using trading parameters and Price/Earnings ratio. With the advances in Artificial Neural Networks, it has become possible to analyze a data set in temporal domain. The use of Time Series Forecasting empowers us to predict the value of an entity in the future based on the previously obtained outputs. The current best fit solution for stock forecasting produces a forecast result with 58% accuracy using feed-forward backpropagation neural network. In this paper, we have represented the data set containing financial stock price as a time series. This time series is forecasted by feeding it to a multi layer back propagation neural network. In real world scenario, stock prices are influenced by many non\ deterministic factors such as national & international economy and public confidence. This paper takes into account factors like Earnings Per Share (EPS) and public confidence and introduces an empirically defined neural network architecture of the form [m - m/2 - m/10 - 1] which gives an optimized structure for predicting the future value of a stock by extrapolating the near future value by the present value comparisons. The experimental results obtained after the training and testing of the financial data are very promising. This increase in accuracy of our financial prediction is due to the factors incorporated for forecasting which can give a clear binary classification for buying or withholding the stock in the current market scenario.