<![CDATA[Accurate demand forecasting is critical in medical relief supply chains where prediction errors can lead to stockouts, delayed response, or inefficient allocation of limited resources. While machine learning (ML) approaches have demonstrated superior predictive capabilities compared to traditional statistical methods and existing research often treats ML as a homogeneous category and rarely conducts systematic benchmarking within specific algorithm families. Furthermore, many studies rely on default model configurations that limiting the reproducibility and failing to fully assess its robustness under volatile demand conditions common in humanitarian logistics. This study addresses these gaps by systematically evaluating multiple tree-based machine learning algorithms for medical relief supply demand forecasting under a structured framework. The research integrates GridSearchCV as hyperparameter optimization, repeated K-fold cross-validation, and statistical significance testing to ensure fair comparison and robustness assessment. The findings indicate that advanced gradient boosting models outperform single-tree and simpler ensemble approaches in terms of predictive accuracy and stability. CatBoost consistently achieved the lowest prediction errors, the narrowest residual dispersion, and the most stable cross-validation performance. Although statistically comparable to other advanced boosting frameworks, CatBoost demonstrated superior robustness during volatile demand conditions and demand surges. These results provide both methodological and practical contributions by establishing a benchmarking framework and identifying a stable forecasting model that suitable for operational deployment in AI-driven medical relief inventory system.]]>
