Combining artificial neural networks and statistics for stock-market forecasting

Abstract

We have developed a stock-market forecasting system based on artificial neural networks. The system has been trained with the Standard & Poor 500 composite indexes of past twenty years. Meanwhile, the system produces the forecasts and adjusts itself by comparing its forecasts with the actual indexes. Since most of stock-market forecasting systems are based on some kind of statistical models, we have also implemented a statistical system based on Box-Jenkins ARIMA(p,d,q) model of time series. We compare the performance of these systems. It shows that the artificial neural network's forecasting is generally superior to time series but it occasionally produces some very wild forecasting values. We then developed a transfer function model to forecast based on the indexes and the forecasts by the artificial neural networks.