A Comparative Study of Decomposition Methods in Time Series Data Forecasting with Seasonal and Nonseasonal Variation

Abstract

     This research aimed to compare 5 decomposition methods: Regression Analysis, Ratio-to-Moving Average, Ratio-to-Trend, Theta and Census II by using 4 accuracy values of forecasting: MAPE, MSE, MAD and MRAE as criteria. The forecasting accuracy was investigated by using 3 tracking signals: e_t SUM_t / √t  and MTE_t. The studied data were monthly time series with 4 different trends and seasonal variation: Total Value of Exports, Consumer Price Index (food and beverages), Quantity of the Highest Electric Power Requirement, and Production of Manufactured Goods (compressor). Based on the study, the optimum forecasting method of these data was Theta except for Regression Analysis which was suitable for Quantity of Electricity Production with sizes of 5 years. From investigation of the forecasting accuracy, the optimum periods of Theta were 1, 2 and 5 months and the optimum period of Regression Analysis was 2 months.