Reject inference is necessary when acceptance rates are low (<20%), but all methods introduce bias. The best defense is to design experiments that accept a random sample of borderline applicants to create unbiased data.
: The primary math tool used to find default risk.
Considered by many to be the "bible of credit scoring," Credit Scoring and Its Applications credit scoring and its applications by l c thomas hot
Before Thomas, credit scoring was mostly (should we lend at application?). Thomas championed behavioral scoring , which uses a borrower’s transaction and payment history over time to predict future risk.
Beyond simple approval, L.C. Thomas explored the ultimate goal of lending: profit. In his influential follow-up, Thomas shifts the lens from individual risk assessment to portfolio management. He argues that lenders should move beyond models of individual credit risk to models that assess the risk of entire portfolios of consumer loans. This approach influences operating decisions in consumer lending, moving the goalpost from "avoiding bad debt" to "maximizing overall profitability" and capital efficiency. Reject inference is necessary when acceptance rates are
: How to manage existing customers by adjusting limits or marketing efforts.
This addresses the initial decision of whether to grant credit to a new applicant. It predicts the probability that a future loan will default based on historical data. Considered by many to be the "bible of
The core value of Thomas’s work lies in its systematic exploration of the used to convert raw consumer data into predictive metrics. Rather than relying on simple rules of thumb, the text outlines robust mathematical frameworks. Logistic Regression and Discriminant Analysis
At its simplest, a credit score is a statistical number that represents the likelihood a borrower will fail to repay a debt as agreed. L.C. Thomas emphasizes that a score is never a judgment of character but a probabilistic forecast based on historical data.
While machine learning has expanded into ensemble methods like Random Forests or CatBoost, Thomas, Edelman, and Crook highlight logistic regression as the industry standard due to its unmatched transparency, regulatory compliance, and interpretability. 3. Reject Inference
“How does this existing customer behave over time?”