Who Gets Misclassified? Age and Socioeconomic Disparities in Machine Learning Credit Default Prediction

Abstract

Machine learning models have increasingly displaced conventional statistical methods in consumer credit scoring, yet systematic evidence on whether their superior predictive performance comes at the cost of disparate treatment across demographic and socioeconomic groups remains limited. This study examines algorithmic bias and fairness in credit default prediction using the Give Me Some Credit dataset (n = 150,000 borrowers), comparing three models, Logistic Regression, Random Forest, and XGBoost, on both predictive performance and demographic fairness metrics. XGBoost achieves the highest AUROC (0.8541), followed by Random Forest (0.8347) and Logistic Regression (0.8021). However, fairness analysis across age groups reveals a Demographic Parity Difference of 0.3316 and an Equal Opportunity Difference of 0.3160 in the XGBoost model, with the 61+ age group exhibiting a False Negative Rate of 0.5285 compared to 0.2125 for the 18–30 group, indicating that older borrowers who will default are substantially less likely to be correctly identified. Socioeconomic disparity analysis across debt ratio quintiles identifies an Equal Opportunity Difference of 0.1419, with Q3 (middle debt ratio) borrowers exhibiting the highest FNR (0.3711). SHAP analysis confirms that delinquency history dominates model predictions, while age and monthly income, variables with protected characteristic implications, carry measurable feature importance. The findings demonstrate that the performance-fairness trade-off in credit scoring is not merely theoretical: the best-performing model on standard metrics simultaneously produces the largest disparities across demographic subgroups.