A STUDY OF THE HYBRID MODEL PERFORMANCE FOR TIME SERIES FORECASTING

Abstract

Automotive export is an important industry that contributes to Thailand economy. The current models to forecast the export quantity include Moving Average (MV), Holt-Winters (HW), Autoregressive Integrated Moving Average (ARIMA), and Artificial Neural Networks (ANNs). However, time series data often contain both linear and nonlinear patterns which the current forecasting models cannot provide much accuracy. In this study, a hybrid model is proposed to forecast automotive export quantity. The hybrid model combines the unique strength of ARIMA and ANN which is good for modeling linear and nonlinear behavior data. The comparison of ARIMA, ANN, Additive ARIMA-ANN and Multiplicative ARIMA-ANN is presented. Performance of the hybrid models is measured using the mean absolute deviation, the mean square error and mean absolute percentage error. The results indicate that the performance of the hybrid models is better in term of forecast accuracy than the other compared models.