Abstract:
Time series forecasting is the important task, which is a starting point in many operational processes. The high forecasting accuracy will increase the efficiency of their operation. At present, the forecasting models which divide data into two components, linear and nonlinear components, called the traditional hybrid model is being popular. Many researchers claim that this hybrid approach is outperformance the individual models but some researchers indicated that this hybrid is not better or even worse.In this study, ARIMA is selected as linear model and ANN and SVM as nonlinear models to compare the accuracy between the ARIMAANN and their individual models, ARIMA and ANN and compare ARIMASVM with both ARIMA and SVM. By proposing another hybrid approach called combined (ARIMA+ANN+SVM) model and compare the accuracy to the ARIMAANN, ARIMASVM and their individual models. The experiment is done on 10 datasets which can be divided into 2 groups, the first 3 datasets is well-known datasets shown in some literature, the second 7 datasets is the real datasets in Thailand in many fields of study and different in characteristics. The result of this study shows that both ARIMAANN and ARIMASVM do not always outperform ARIMA or ANN and ARIMA or SVM in all datasets and sometimes even worse. However, the proposed combined (ARIMA+ANN+SVM) model outperforms the ARIMAANN and ARIMASVM in 9 datasets and better than the individual models in 8 datasets. For the rest 2 datasets, the accuracy is close to the best of individual models and none of datasets underperform the individual models. In conclusion, the combined (ARIMA+ANN+SVM) is a good alternative model for the important tasks that need a high forecasting accuracy.