Optimal Selection of Multivariate Fuzzy Time Series Models to Non-stationary Series Data Forecasting

This paper links testing of non-stationary time series features to the selection of fuzzy model for time series prediction. The data for model test are obtained from AREMOS, Taiwan. Empirical results show that fuzzy time series models have different performance patterns in predicting non-stationary time series. Data with a clear time trend, such as consumption, exports or other macroeconomic data, are best predicted with a Heuristic model. For data with significant oscillations, financial indices, such as a futures index, a Markov model performs best. A Two-factor model produces the most accurate predictions for series that need to be differenced twice to become stationary. The results provide the fuzzy models a beneficial reference for an effective use of time series prediction.