PERFORMANCE EVALUATION OF SEASONAL ARIMA-SVR AND SEASONAL ARIMAX-SVR HYBRID METHODS ON FORECASTING PADDY PRODUCTION

Abstract

This study explores advances in forecasting time series data by combining linear and non-linear models. Traditional methods such as ARIMA and its variant ARIMAX are effective for linear data but have limitations when dealing with non-linearity. Support Vector Regression (SVR), a non-linear method, complements these weaknesses. Hybrid models such as ARIMA-SVR and ARIMAX-SVR synergize ARIMA or ARIMAX for linear components and SVR for non-linear components, improving accuracy. The purpose of this study is to evaluate the performance of hybrid ARIMA-SVR and ARIMAX-SVR methods on Indonesian paddy production data. The data analyzed is national-level data per sub-round (i.e., three sub-rounds per year) from sub-round 1 (January-April) of 1992 to sub-round 3 (September-December) of 2024, obtained from the Indonesian Central Statistics Agency and the Indonesian Ministry of Agriculture.Forecasting accuracy is measured using Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). The results show that the best model is the Seasonal ARIMAX (1,1,1)(0,1,1)[3]-SVR ( 0.05) hybrid model, with the smallest RMSE and MAPE values of 0.304 and 1.473%. The addition of the harvested area variable and the ASF dummy improved the accuracy of the ARIMAX model prediction, while the application of SVR to ARIMAX residuals successfully captured previously undetected linear patterns. Based on these considerations, the Seasonal ARIMAX(1,1,1)(0,1,1)[3]-SVR ( 0.05) hybrid model was selected as the model with the best performance.