Deep Learning vs. Traditional Machine Learning for Consumer Credit Default Prediction: A Comparative Study

Credit default prediction is fundamental to risk management in consumer lending. This study compares logistic regression (LR), a multilayer perceptron (MLP), and XGBoost for credit default prediction using 30,000 synthetic consumer lending records modeled after LendingClub data (34.0% default rate, 22 features). Stratified 5-fold cross-validation reveals that LR achieves the highest AUC-ROC (0.726 ± 0.006), marginally exceeding MLP (0.723 ± 0.006) and XGBoost (0.705 ± 0.007). Bootstrap DeLong tests confirm all pairwise AUC differences are significant (p < 0.001), and the Friedman test validates the overall ranking (χ² = 10.0, p = 0.007). However, XGBoost achieves the highest F1-score (0.548) and recall (0.584), detecting substantially more defaults at the cost of lower precision. LR achieves the best calibration (Brier score = 0.192). These results illustrate a metric-dependent trade-off: LR ranks best by discrimination (AUC) and calibration, while XGBoost ranks best by default detection rate. The findings caution practitioners against equating model complexity with predictive value and argue for metric-driven model selection in regulated lending.