Analisis Komparatif Akurasi dan Stabilitas Model Peramalan Harga Beras Harian Nasional

Abstract

Daily forecasting of national rice prices is important for supporting data-driven food price monitoring systems. This study evaluates statistical, machine learning, and deep learning approaches for forecasting Indonesia’s national rice prices over the 2017–2025 period. Missing values in the daily data were handled using the Last Observation Carried Forward (LOCF) method, limited to gaps of up to seven days, which successfully filled all missing observations. The completed time series was then transformed into a supervised learning format using a seven-day sliding window selected through temporal analysis and validation-based sensitivity testing. The dataset was divided chronologically, with the final 90 days reserved as the test set. The evaluated models included Seasonal Naive, XGBoost, Support Vector Regression optimized using Particle Swarm Optimization (SVR-PSO), and Temporal Convolutional Network (TCN). Model performance was evaluated using MAE, RMSE, MAPE, and MSE, while SVR-PSO and TCN were each run 10 times to assess performance stability. SVR-PSO achieved the best performance, with an MAE of 16.830 ± 2.722, an RMSE of 22.189 ± 2.216, a MAPE of 0.1068% ± 0.0173, and an MSE of 497.270 ± 107.992. These findings indicate that SVR-PSO produces relatively accurate and stable short-term forecasts and has the potential to support the development of data-driven national food price monitoring and forecasting systems.