Automatic structure identification of TSK fuzzy model for stock index forecasting

Abstract

Among financial time-series analysis tasks, stock index forecasting has been considered as one of the challenging and difficult tasks. Since an accurate stock index prediction enhances stock market returns, it is highly promising research area and has attracted particular attention. In this paper, to predict stock index with complex and nonlinear characteristics, an automatic structure identification (SI) method of TSK fuzzy model is proposed. Typically, SI procedures of fuzzy models consist of relevant input selection, fuzzy rule generation and parameter search space determination. In this study, mutual information is employed to select relevant input variables and fuzzy c-means (FCM) clustering algorithm is used to generate fuzzy if-then rules. In FCM clustering, the number of clusters should be fixed in advance. This paper uses performance criterion to determine the optimal number of clusters in FCM clustering. After deciding the optimal cluster number, fuzzy if-then rules are extracted and parameter search space boundaries are fixed. Finally, premise and consequent parameters are optimized by cooperative random learning particle swarm optimization proposed by Zhao et al. To confirm the effectiveness, the proposed TSK fuzzy modeling method and comparison methods are applied to Korea Composite Stock Price Index dataset. The experimental results show that the TSK fuzzy models with the proposed SI method outperform comparison methods.