Rough-set rule induction to build fuzzy time series model in forecasting stock price

Abstract

The investment research of stock is an hot issue because it will bring significant returns for investor. The advantages of fuzzy time series can solve the forecast problem in data with linguistic expression. In order to improve fuzzy time-series models, this study direct replaced fuzzy logic relationships with rule-based method to extract fuzzy forecast rules from time series observations. Therefore this paper proposed a rule induction based fuzzy time series model to forecast stock index. In summary, there are four refinements to improve accuracy of forecast: (1) use Miller's magic seven (plus or minus two) to determine the lengths of linguistic intervals, (2) utilize LEM2 algorithm to generate forecast rules, (3) defuzzify the forecast interval based on the generated rules, and (4) use adaptive expectation model to strengthen forecasting performance. For evaluating the proposed model, the practically collected TAIEX stock price from 1998 to 2006 years are used as experimental dataset, and Chen's model, Yu's model, stepwise regression based on ANFIS, and stepwise regression based on support vector regression are compared with the proposed model in RMSE (root mean square error) and profits criteria. The results express that proposed method holds good performance in accuracy.