Forecasting Philippine Rice Prices: Comparison of Traditional Time Series and Machine Learning Models

Abstract

The Philippine rice market is characterized by high volatility and a critical impact on national food security, necessitating the use of accurate forecasting tools for price monitoring and risk mitigation. The primary objective of this study was to determine the optimal forecasting methodology for the nominal wholesale price of regular-milled rice. The analysis utilized 430 monthly observations (January 1990 – October 2025), partitioned into an 80% training set (344 data points) and a 20% out-of-sample test set (86 data points). Eight time series models including traditional methods: Seasonal Naïve, ETS, ARIMA, TBATS, Theta, and Prophet, and machine learning algorithms: Random Forest and XGBoost were evaluated. Performance was assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), supplemented by Ljung–Box and Lilliefors tests for residual diagnostics. Evaluation revealed that the Random Forest model achieved the best predictive accuracy, confirming the superior capability of non-linear models to capture the volatile patterns present in the wholesale price data. Residual diagnostics indicated a fundamental trade-off between structural adequacy and predictive accuracy. The final projection forecasts that wholesale prices will stabilize within the Php 34.00/kg to Php 35.50/kg range through 2027. The general price stability predicted for the next two years suggests policy focus may prioritize long-term supply-side improvements rather than short-term demand controls, unless external shocks occur.