Forecasting Short-Term KOSPI Time Series Based on NEWFM

Abstract

Fuzzy neural networks have been successfully applied to generate predictive rules for stock forecasting. This paper presents a methodology to forecast the daily Korea composite stock price index (KOSPI) by extracting fuzzy rules based on the neural network with weighted fuzzy membership functions (NEWFM) and the minimized number of input features using the distributed non-overlap area measurement method. NEWFM supports the KOSPI time series analysis based on the defuzzyfication of weighted average method which is the fuzzy model suggested by Takagi and Sugeno. NEWFM classifies upper and lower cases of next daypsilas KOSPI using the recent 32 days of CPPn,m (Current Price Position of day n : a percentage of the difference between the price of day n and the moving average of the past m days from day n-1) of KOSPI. In this paper, the Haar wavelet function is used as a mother wavelet. The most important four input features among CPPn,m and 38 numbers of wavelet transformed coefficients produced by the recent 32 days of CPPn,m are selected by the non-overlap area distribution measurement method. The total number of samples is 2928 trading days, from January 1989 to December 1998. About 80% of the data is used for training and 20% for testing. The result of classification rate is 59.0361%.