Comparative Evaluation of Forecasting Models for Short-Term Natural Rubber Prices Using a Web-Based Monitoring Platform

Abstract

Thailand is one of the world’s largest producers of natural rubber, a key agricultural export. The price of natural rubber is highly volatile in the short term, driven by factors such as weather, global demand, and spillovers from energy markets. Reliable price forecasts, even for short horizons, can assist farmers, cooperatives, and local government agencies in designing better procurement and support policies. This study presents an offline empirical benchmark comparing three commonly used forecasting methods, Prophet, Support Vector Regression (SVR), and Autoregressive Integrated Moving Average (ARIMA), for short-term daily natural rubber prices using data from a web-based monitoring and forecasting dashboard of rubber prices. The dataset covered the period from 1 January 2016 to 31 August 2026. The data were divided chronologically into a training set from 1 January 2016 to 31 December 2024 and a held-out test set from 1 January 2025 to 31 August 2026, respectively. Each model was trained on a rolling historical window and tested on a held-out forecast horizon using three complementary error metrics: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and coefficient of determination ($R^2$). The best performance was obtained by the Prophet model (MAE = 1.10 baht/kg, MAPE = 3.20\%, $R^2$ = 0.898), followed by SVR (MAE = 1.35, MAPE = 3.80\%, $R^2$ = 0.875) and ARIMA (MAE = 1.45, MAPE = 4.10\%, $R^2$ = 0.852). These findings indicate that, within this offline empirical benchmark, Prophet's additive decomposition of trend and seasonal components offers a good trade-off between interpretability and accuracy for short-horizon commodity price forecasting, compared with ARIMA and SVR. The results provide operational guidance for selecting lightweight, deployable forecasting models for agricultural price-monitoring systems.