<![CDATA[Accurate credit scoring is critical for financial stability, yet modern machine learning models often operate as "black boxes," creating a significant challenge for regulatory compliance and business trust. This research addresses this problem by developing a comprehensive, transparent, and robust framework for credit risk prediction. The primary objective was to build a high-performance model that is not only accurate but also fully interpretable and stable over time. The research method involved using a public loan dataset (2007-2014) with 466,285 loan records to train a Light Gradient Boosting Machine (LightGBM). To ensure real-world applicability and prevent temporal data leakage, a rigorous time-aware validation strategy was employed, splitting the data into a historical training set (2007-2013) and a future out-of-time (OOT) test set (2014). The SHapley Additive exPlanations (SHAP) framework was integrated to provide clear explanations for every prediction. The main research results demonstrate the model's strong predictive capability, achieving an AUC-ROC of 0.711 and a KS-statistic of 0.309 on the OOT dataset. The model's stability was confirmed with a low Population Stability Index (0.0685). Furthermore, SHAP analysis revealed that predictions were driven by financially intuitive factors like interest rate and annual income. The study concludes by presenting a complete workflow that successfully bridges the gap between high-performance modeling and the practical need for transparent, business-aligned risk management tools.]]>
