Fuzzy Models for Time Series Analysis: Towards Systematic Data Pre-processing

Time series forecasting is an all-pervasive task that affects almost all disciplines. Given that time varied phenomena almost invariably need pre-processing, it is important to develop a framework where such pre-processing is executed in a systematic and transparent manner. In this paper, we investigate the effect of data pre-processing on the forecast performance of subtractive clustering fuzzy model. Our work on benchmark data sets (US Census Board and US Federal Reserve data) shows that ad hoc application of pre-processing techniques is not optimal. We have used autocorrelation functions to understand both the behavior of time series and the effects of different preprocessing methods on prediction accuracy. Our results indicate that the use of autocorrelation functions to determine the suitability of different pre-processing methods is beneficial