Financial Time-Series Forecasting based on a Neural Network with Weighted Fuzzy Membership Functions and the Takagi-Sugeno Fuzzy Model

Abstract

This paper proposes financial time-series forecasting using a feature selection method based on the non-overlap area distribution measurement method supported in a neural network with weighted fuzzy membership functions (NEWFM) and the Takagi-Sugeno (T-S) fuzzy model. The non-overlap area distribution measurement method selects the minimum number of features with the highest performance by removing the worst features one by one. This paper uses CPPn,m (current price position on day n: percentage of the difference between the price on day n and the moving average of the past m days' prices from day n-1) as a technical indicator. The performance result improves from 58.35% to 58.86% when CPPn,5 is added to a minimum number of features that are selected by the non-overlap area distribution measurement method. The T-S fuzzy model can provide the weighted average defuzzification method to represent a trend line such as a financial time series. This paper generally demonstrates fluctuations similar to a financial time series of the daily KOSPI's trend line by the defuzzification method.