A Multimetric Evaluation of Ensemble and Neural Models for Imbalanced Credit Scoring on the German Credit Dataset

Abstract

Credit scoring models must balance global discrimination, minority-class detection, probability quality, and computational efficiency, particularly when default cases are less frequent than non-default cases. This study benchmarks Logistic Regression, Random Forest, XGBoost, and a Multilayer Perceptron using the German Credit Dataset, which contains 1,000 borrower records. All models were evaluated using a consistent preprocessing, partitioning, and imbalance-aware training protocol. Performance was assessed through ROC-AUC, PR-AUC, Balanced Accuracy, precision, recall, F1-score, Brier Score, and computational time. The results indicate moderate predictive performance across the evaluated models. XGBoost achieved the highest global discrimination with a ROC-AUC of 0.764, whereas the MLP obtained the highest PR-AUC of 0.378 and the lowest Brier Score of 0.123, indicating the best native probability quality among the evaluated models. No post-hoc probability calibration was applied. However, the MLP produced lower recall at the selected classification threshold. Logistic Regression required the least computation but recorded the lowest ROC-AUC and PR-AUC. Exploratory SHAP analysis was used to identify influential borrower attributes. These benchmark-specific findings demonstrate that model selection depends on the intended decision objective rather than architectural complexity alone.