A dimension reduction method for stock-price prediction using multiple predictors

Stock-price prediction has been the focus of extensive studies. Historical price values have been proven to have power to predict future prices. At the same time, different economic variables have also been used in the literature to predict stock-price values with high accuracy. In this work, we develop a general method for stock-price prediction using multiple predictors. First, we use multichannel cross-correlation coefficient as a measure for selecting the most correlated set of variables for each stock. We then construct the temporally local covariance matrix of the data and use this as the basis for a dimension-reduction method for prediction. This method involves resolving the predictive data (predictors) onto a principal subspace and from there producing a prediction that is consistent with the resolved data. Our method is easily implemented and can accommodate an arbitrary number of predictors. We investigate the optimal number of predictors based on two performance metrics: mean squared error of the prediction and the directional change statistic. We illustrate our results based on historical daily price data for 50 companies.