A modular Takagi-Sugeno-Kang (TSK) system based on a modified hybrid soft clustering for stock selection

This study presents a new hybrid intelligent system with ensemble learning for stock selection using the fundamental information of companies. The system uses the selected financial ratios of each company as the input variables and ranks the candidate stocks. Due to the different characteristics of the companies from different activity sectors, modular system for stock selection may show a better performance in comparison with an individual system. Here, a hybrid soft clustering algorithm is proposed to eliminate the noise and partition the input data set into more homogeneous overlapped subsets. The proposed clustering algorithm benefits from the strengths of the fuzzy, possibilistic and rough clustering to develop a modular system. An individual Takagi-Sugeno-Kang (TSK) system is extracted from each subset using an artificial neural network and genetic algorithm. To integrate the outputs of the individual TSK systems, a new weighted ensemble strategy is proposed. The performance of the proposed system is evaluated among 150 companies listed on Tehran Stock Exchange (TSE) regarding information coefficient, classification accuracy and appreciation in stock price. The experimental results show that the proposed modular TSK system significantly outperforms the single TSK system as well as the other ensemble models using different decomposition and combination strategies.