<![CDATA[The plantation subsector is a cornerstone of the national economy, yet its productivity is increasingly volatile due to climate change. Predicting production yields remains challenging as traditional models often fail to capture complex nonlinear temporal dependencies and seasonal cycles. This study aims to improve the prediction accuracy of five major plantation commodities, namely palm oil, rubber, coffee, tea, and sugarcane, by optimizing the Category Boosting (CatBoost) algorithm. The analysis uses monthly data from 2009 to 2024, combining official production and land statistics from the Central Bureau of Statistics (BPS) with national temperature and rainfall records from the Meteorology, Climatology, and Geophysics Agency (BMKG) to ensure transparency. Unlike standard approaches that rely on default parameters and random data splitting, this research applies a rigorous optimization pipeline. Random Search is used for hyperparameter tuning, supported by lag features to capture short term dynamics and sinusoidal transformations to represent seasonal cycles. A Walk Forward Validation technique with an expanding window is employed to prevent look ahead bias and ensure realistic evaluation. The optimized model significantly outperforms the baseline. Sugarcane (R² 0.95) and Coffee (R² 0.97) show excellent accuracy, while Palm Oil improves markedly (R² 0.80) as more historical patterns are learned. Rubber and Tea remain difficult to predict, indicating insufficient explanatory features rather than model limitations. The study concludes that combining hyperparameter optimization with temporal feature engineering enables CatBoost to effectively model agricultural time series data and provides a solid foundation for strategic production planning.]]>
