Credit scoring

Credit scoring produces a numeric score, or a categorical risk grade, that ranks a borrower's relative creditworthiness, typically derived from payment history, outstanding debt, length of credit history, and the type and number of accounts held. Widely used consumer scores like FICO or VantageScore in the United States range across a fixed scale, with higher scores indicating lower expected credit risk.

Traditional credit scores are built with logistic regression or similar models trained on historical repayment behavior; newer alternative-data and machine-learning scoring models add data such as cash-flow patterns or rental payments, especially for borrowers with thin credit files. Credit scoring differs from probability of default estimation in that a score is usually a relative ranking tool tuned for ordering borrowers by risk, while a PD model produces a calibrated probability meant to be used directly in loss and pricing calculations, though the two are closely related and often built from similar underlying data.

Lenders use credit scoring to approve or decline applications, set interest rates and credit limits, and monitor existing accounts for rising risk. Model performance is commonly assessed with discrimination measures like ROC AUC or the gini coefficient. A well-documented pitfall is proxy discrimination, where variables correlated with protected characteristics such as race or age influence the score even when those characteristics are excluded outright, which is why credit-scoring models in regulated markets face fair-lending review.