Ensemble Method Based on Two Level ARIMAX-FFNN for Rainfall Forecasting in Indonesia

One of relatively new modern methods for time series forecasting is ensemble forecasting that employs averaging or stacking from the results of several methods. This paper focuses on the development of ensemble ARIMAX-FFNN-Hybrid for rainfall forecasting by using averaging and stacking method. Three data about dasarian rainfall in Indonesia, i. e. Karangsuko, Kalipare and Gondanglegi area, are used as case study. Root mean of squares errors and Symmetric mean absolute percentage errors in testing datasets are used for evaluating the forecast accuracy. The results of ensemble ARIMA-FFNN-Hybrid are compared to two classical statistical method, i. e. individual ARIMA and ARIMAX, and five modern statistical methods, namely individual FFNN, individual Hybrid ARIMAX-FFNN, ensemble ARIMAX, ensemble FFNN, and ensemble Hybrid. The results show that ensemble hybrid yields more accurate forecast in Karangsuko area than other methods, whereas in Gondanglegi and Kalipare area show that individual hybrid and FFNN, respectively, is the best method. Additionally, this conclusion in line with the results of M3 competition, i. e. modern methods or complex methods do not necessarily produce more accurate forecast than simpler one.