Evolutionary fuzzification of RIPPER for regression: Case study of stock prediction

Abstract

This paper presents a novel approach for extracting the knowledge base (KB) of a Mamdani fuzzy rule based system (FRBS) for stock market prediction. The KB, which is the most important component of the FRBS, has two main components: rule base (RB) and data base (DB). In the proposed model, the RB is learned using repeated incremental pruning to produce error reduction (RIPPER), which is state-of-the-art in classification rule induction. The RIPPER algorithm is applied to classification problems. In order to extend it for regression problems, unsupervised discretization of the target attribute and supervised discretization of continuous input attributes are used. Next, the classification rules obtained from RIPPER are fuzzified and the initial Mamdani FRBS is formed. The DB is tuned using a Genetic Algorithm (GA). The proposed model in this paper is the first model utilizing directly capabilities and benefits of rule-based classification systems in regression. The accuracy of the proposed model is tested in the context of stock market prediction, which is a complex and difficult area in regression problems. The proposed model is implemented using several indices from different stock markets such as the Taiwan Stock Exchange Index (TSE), Tehran Price Index (TEPIX). Other indices including the Industry Index, Top 50 Companies Index and Financial Group Index from Tehran Stock Exchange are also considered. Furthermore, the daily stock prices of multiple large companies such as Apple, DELL, IBM, British Airlines and Ryanair Airlines are incorporated. As shown by the mean absolute percentage error (MAPE) and non-parametric statistical tests, the proposed model offers superior performance compared to other models.