A lender needs to know whether a credit model ranks borrowers by default risk before choosing an approval cutoff. Area under the receiver operating characteristic curve (ROC AUC) gives one measure of that ranking ability. The ROC curve plots the true positive rate—defaults caught—against the false positive rate—good borrowers flagged—across possible cutoffs; AUC is the area beneath it. section 4.4.1
AUC can be read as the probability that a randomly chosen borrower who defaults receives a lower credit score than a randomly chosen borrower who does not. Unlike accuracy at a fixed cutoff, it lets teams compare models’ ranking power before applying business rules. In consumer lending, 0.70 is described as a general baseline for a workable model, while advanced or specialized models may aim above 0.80. section 4.4.1
Use AUC alongside checks that answer different questions:
- Compare separation of good and bad borrowers with AUC or K-S. section 4.4
- Monitor shifts in the applicant population with PSI; AUC is not a drift measure. section 4.4 section 4.4.1
- For highly imbalanced fraud data, consider area under the precision-recall curve (AUC-PR): ROC AUC can give an overly optimistic view. section 8.1.1
A cutoff is a separate decision
Suppose a lender is considering rejecting applicants below a credit score of 650. AUC assesses ranking across cutoffs, not just results at 650. It also does not weigh the financial difference between approving a borrower who defaults and rejecting one who would have repaid. Choosing the cutoff therefore requires a separate assessment of business costs. section 4.4.1
