An innovative use of historical data for neural network based stock prediction

Abstract

Using artificial neural network Is a common approach for the stock time series prediction problem. Unlike variety of researches that focus on selecting different indicators, network training, network architecture, etc., we are focusing on the selection of appropriate time points from the time sequence to serve as the input of the neural network prediction system for dimensionality reduction. We propose to select the time points based on data point importance using perceptually important point identification process. The empirical result shows that the proposed method generally outperformed the traditional method using uniform time delay.