Forecasting Direction of Stock Index Using Two Stage Hybridization of Machine Learning Models

Accurate forecasting of directional changes in stock prices is essential for trading decisions. This study attempts to predict the direction of movement for the next day for S&P BSE Sensex index. Experiments are done in two phases. The first stage provided the direction classification that is whether predicted movement is up or down based on six technical indicators distinctly. Indicators are calculated as per their definition by using daily trading data of open, high, low, close and volume. Stage two takes in the processed discretized data for predictions from stage one. The accuracy of both stages for each model is evaluated and compared. Experimental results show significant improvement of the combined model over single-pass model hence supporting the assumption that conversion from continuous to discrete form based on indicators filter more noise than relevant information and provides an effective mechanism of dimensionality reduction. Random forest provided the best accuracy which is closely followed by support vector and artificial neural network.