This study develops a governance-aligned framework for probability of default modeling under class imbalance, combining domain-informed feature engineering, resampling, ensemble learning and explainable artificial intelligence. Using real mortgage data, multiple configurations are evaluated under strict cross-validation to ensure unbiased performance estimation. The results show that economically grounded feature transformations substantially improve discrimination while maintaining transparency, whereas resampling provides only limited additional benefits once features are engineered. Ensemble architectures, in particular LightGBM and stacking, deliver strong discriminatory performance without synthetic data generation. Shapley additive explanations (SHAP) analysis confirms the economic plausibility of model drivers, while bootstrap diagnostics demonstrate robust generalization. Beyond predictive accuracy, the framework links methodological choices to supervisory priorities, showing that feature engineering can in some cases outweigh oversampling for boosting models and that explainability supports reliable model assessment.