<![CDATA[Rainfall prediction is challenging due to the complex and nonlinear nature of meteorological data. Previous studies using XGBoost with feature selection have demonstrated superior performance compared to other models, but evaluations have focused solely on error metrics (RSME, SME, MAE). Recent research suggests that predictive models should be evaluated for generalization, stability, interpretability, and computational efficiency to ensure their reliability. To close this gap, this study uses 8,750 hourly records obtained from Open-Meteo with 81 engineered features to evaluate XGBoost under three scenarios: without feature selection, single feature selection (MI, Boruta, SHAP, mRMR, ReliefF), and hybrid feature selection. Our findings demonstrate that accuracy is not always increased by feature selection. It does, however, increase interpretability, decrease overfitting, and improve computational efficiency. SHAP provides the most reliable performance among single methods, achieving lower RMSE (0.72632) and improved stability. Hybrid feature selection produces the most balanced performance gap = 0.01325, and stable variance = 0.03315 while reducing feature complexity to 35 variables. This study theoretically shows the value of multidimensional evaluation that goes beyond error metrics. In practical terms, this study suggests a feature selection method for rainfall prediction systems that are effective, reliable, and simple to understand.]]>
Copyrights © 2026
