AI-Powered Credit Scoring vs. Traditional Scoring Models: A Comparative Study on Loan Default Prediction
0 viewsAbstract
This study compares the predictive performance of a traditional statistical credit scoring model against a modern artificial-intelligence-based model in forecasting consumer loan default. Because access to a real bank's confidential loan portfolio was not feasible for this project, a synthetic loan-applicant dataset (5,000 records) was engineered to mirror the statistical structure of real-world credit bureau data, including a realistic 22.8% default rate and ten borrower-level features such as income, debt-to-income ratio, credit utilization, and payment history. A logistic regression model, representing the rule-based scorecards traditionally used in retail banking, was benchmarked against a gradient boosting model, representing contemporary AI-driven scoring engines. Both models were trained on an identical 75/25 train-test split and evaluated using accuracy, precision, recall, F1-score, and AUC-ROC. The AI model outperformed the traditional model on every metric, most notably raising AUC-ROC from 0.814 to 0.936 and recall (the ability to correctly flag true defaulters) from 43.9% to 68.4%. These findings suggest that AI-based scoring can materially reduce a lender's exposure to undetected default risk, though questions of interpretability, fairness, and regulatory compliance remain important considerations for real-world adoption.
Keywords
Artificial Intelligence, Credit Scoring, Financial Risk Management, Gradient Boosting, Loan Default Prediction, Machine Learning