Building the Forecasting Model for Time Series Based on the Improved Fuzzy Relationship for Variation of Data

Abstract

Forecasting for time series has always been of interest to statisticians and data scientists because it offers a lot of benefits in reality. This study proposes the fuzzy time series model which can both interpolate historical data, and forecast effectively for the future with the important contributions. First, we build the universal set based on the percentage of the original data variation, and divide it to clusters with the suitable number by the developed automatic algorithm. Next, the new fuzzy relationship between each element in series and the obtained clusters is established. The bigger the variation is, the more the clusters are divided. Finally, combining the two above improvements, we propose the new principle to forecast for the future. The experiments on many well-known data sets, including 3003 series of M3-competition data show that the proposed model has shown the outstanding advantage in comparing to the existing ones. Because the proposed model is established by the Matlab procedure, it can apply effectively for real series.