High-Order Fuzzy Time Series Forecasting by Using Membership Values Along with Data and Support Vector Machine

Abstract

In the past few years, non-stochastic fuzzy time series (FTS) models have drawn remarkable attention of researchers from different domains. Unlike traditional stochastic models, FTS models do not require any strict assumption on the characteristics of data to be modeled and are applicable to time series even with uncertainty. However, the effectiveness of FTS models largely depends on the determination of effective length of interval and modeling of fuzzy logical relationships (FLRs). Motivated by this, in this paper, we have developed a novel method using fuzzy c-means clustering to determine the unequal-length of interval. Additionally, for the first-time membership values are considered while modeling the FLRs using support vector machine (SVM). The order of the model is determined by analyzing the autocorrelation function and partial autocorrelation function of the time series. To measure the accuracy of the proposed model, ten different time series datasets are considered. Four recently developed fuzzy time series forecasting models and two popular crisp time series forecasting models using MLP and SVM are considered for comparative performance analysis. From the experimental result analysis, it is observed that the proposed model outperforms other alternatives and shows statistically better forecasting accuracy based on the popular Wilcoxon signed-rank test; and Friedman and Nemenyi hypothesis test.