Forecasting Stock Return Using Multiple Information Sources Based on Rules Extraction

Forecasting stock return has drawn great attentions of researchers from different disciplines although it is a very challenge problem. Time series analysis methods have been widely suggested in the literature to tackle the problem. However, they can not provide clear explanations for decision making in financial investment activities. In this paper, we present the use of rule extraction techniques to build intelligent and explanatory prediction systems for stock return. In addition to utilization of stock trading information, public Internet information sources are also introduced into the process of rule extraction to capture more features. Specifically, investor sentiment and stock recommendations which can be freely accessed from the Internet have been proved to become potential indicators to stock return. We incorporate multiple information sources to generate meaningful associate rules and reveal more insightful patterns from enormous history data. A novel forest-like structure is proposed to represent and store the rules. Experiment and empirical studies are done and results show that the prediction accurate rate is satisfactory and also provide further support for the work of Hribar and McInnis.