Forecasting with multivariate fuzzy time series: A statistical approach

Abstract

In recent times, forecasting using fuzzy time series (FTS) models is a very popular research topic.In multivariate fuzzy time series forecasting models, for a given sequence, similar sequences are extracted by explicit matching of rules. In these models, index value of fuzzy set with the maximum membership value is used for rule matching. In this case, two situations can arise: (1) more than one similar sequence can be found in the training set and (2) no matching rule is found. In order to get an accurate forecast, an efficient defuzzification technique should be adopted. To address this problem, a new multivariate forecasting model is presented based on the Bayesian approach. The proposed model is applied to standard datasets: (1) Taiwan Stock Exchange (1997–2003) and (2) US Civilian Unemployment. Forecast accuracy is measured using root mean square error (RMSE). Rank test is also performed. Proposed idea shows better result than some of the popular fuzzy time series forecasting models.