AI‐Driven Credit‐Risk Stratification for Banking: A Leakage‐Controlled Ordinal Benchmark of Logistic Regression, Random Forest, XGBoost, and Neural Networks

Retail-bank credit decisioning still relies on binary good/bad classifiers, and published results on the field's three canonical public corpora, German Credit, Taiwan Default, and Lending Club, are largely incommensurable because they differ in preprocessing placement, resampling design, and leakage control. This study contributes a leakage-controlled ordinal benchmark: multinomial logistic regression, random forest, XGBoost, and a neural network are compared on a three-tier (Low/Medium/High) formulation of all three corpora under one identical, fully reproducible R protocol in which the split precedes all fitting, scaling and SMOTE are estimated inside cross-validation folds, and tier-defining columns are excluded from the predictors. Under this protocol the best model is corpus-dependent: logistic regression leads on the small, weakly correlated German corpus (High-tier AUC 0.738), while tree ensembles lead on the larger Taiwan and Lending Club corpora (accuracy 0.86 and 0.70). Re-admitting the excluded columns lifts the ensembles to a perfect 1.000 on Taiwan, quantifying how much apparent performance uncontrolled designs can manufacture and explaining the wide spread of published results. The study also documents a multi-class probability-decoding fault that silently drives a trained model to chance, and maps predicted tiers to graduated lending actions for deployment.