Machine learning-based credit scoring: A comparative analysis of logistic regression and random forest models

Abstract

This paper presents a controlled, head-to-head comparison of Logistic Regression (LR) and Random Forest (RF) for consumer credit scoring using the German Credit dataset. A consistent pipeline -label encoding, stratified 80/20 train-test split, and five-fold cross-validation -supports like-for-like evaluation on accuracy, precision, recall, F1, ROC-AUC, and confusion matrices. Results show competitive but distinct strengths: RF attains slightly higher accuracy (0.775 vs. 0.765) and notably higher precision (0.703 vs. 0.627), reflecting fewer false positives; LR achieves higher recall for the default class (0.533 vs. 0.433) and a marginally better ROC-AUC (0.79 vs. 0.78), indicating stronger discrimination at low-FP operating points. Confusion matrices corroborate these trade-offs (LR: TN=121, FP=19, FN=28, TP=32; RF: TN=129, FP=11, FN=34, TP=26). Feature analysis aligns with domain priors: credit amount, age, and duration dominate RF importance, while LR coefficients provide directionally transparent effects for audit. Practically, the findings support a hybrid deployment: RF for back-end risk ranking and early warning, LR for audit-facing decisions and regulatory reporting. Limitations include reliance on public datasets and a focus on discrimination over calibration and fairness. Future work should examine proprietary portfolios, cost-sensitive thresholds and drift monitoring, and integrate explainable AI to reconcile lift with governance.