A study on leading machine learning techniques for high order fuzzy time series forecasting

Fuzzy time series forecasting (FTSF) methods avoid the basic assumptions of traditional time series forecasting (TSF) methods. The FTSF methods consist of four stages namely determination of effective length of interval, fuzzification of crisp time series data, modeling of fuzzy logical relationships (FLRs) and defuzzification. All the four stages play a vital role in achieving better forecasting accuracy. This paper addresses two key issues such as modeling FLRs and determination of effective length of interval. Three leading machine learning (ML) techniques, namely deep belief network (DBN), long short-term memory (LSTM) and support vector machine (SVM) are first time used for modeling the FLRs. Additionally, a modified average-based method is proposed to estimate the effective length of interval. The proposed FTSF-DBN, FTSF-LSTM and FTSF-SVM methods are being compared with three papers from the literature along with four crisp TSF methods using multilayer perceptron (MLP), LSTM, DBN and SVM. A total of fourteen time series datasets (Sun Spot, Lynx, Mumps and 11 TAIEX time series datasets i.e. 2000-2010) are considered for comparative performance analysis. Results revealed the statistical superiority of FTSF-SVM method and proposed improved average-based method based on the popular Friedman and Nemenyi hypothesis test. It is also observed that the proposed FTSF methods provide statistical superior performance than their crisp TSF counterparts.