Credit approvals determine which repayment outcomes a model can learn from

Selection bias in credit-model development arises when the borrowers with observed repayment outcomes differ systematically from the population the model is intended to assess. Reject inference is the attempt to account for missing repayment outcomes among applicants who were not granted the relevant credit.

Approval policy selects the observed outcomes

Take 10,000 loan applicants. A lender approves 4,000 and declines 6,000. After the specified performance period, it observes whether the 4,000 approved loans met its default definition.

If 80 approved loans default, the observed default rate is 2% among approved borrowers. That denominator cannot be expanded to all 10,000 applicants: the declined applications did not produce the same loan or its repayment history.

Recording declined applicants as nondefaults would replace missing outcomes with invented successful repayments. Recording all of them as defaults would invent the opposite result.

A model can perform well inside a narrow accepted population

The accepted borrowers were selected using the old score, income rules, fraud checks, overrides, and other underwriting conditions. A model trained on their outcomes learns from that selected population.

Strong ranking or calibration within accepted accounts does not establish equally strong performance among previously declined applicants. Their characteristics and the relationships between those characteristics and default can differ.

The missing outcome is also product-specific. A declined applicant’s later repayment on a different loan is additional information, but it does not directly reveal repayment under the amount, price, and terms originally declined.

Policy changes also change the evaluation sample

Suppose the lender lowers its approval threshold and begins lending to another 1,000 applicants. Those loans create observations in a region where the old dataset had less or no comparable experience.

A change in the overall default rate can then reflect the changed borrower mix, changed terms, economic conditions, model performance, or several of these together. Comparing the headline rate before and after the policy change cannot isolate the cause.

The evaluation needs the decision date, policy version, score version, approval path, loan terms, and a consistent outcome horizon. Cohorts that have not completed the observation period cannot supply the same evidence as mature cohorts.

Reject inference requires assumptions

Estimating outcomes for declined applicants does not make their missing counterfactual repayments observable. The inference depends on assumptions about selection and how observed behavior transfers to the unobserved population.

Those assumptions need explicit testing and sensitivity analysis. Approval-rate parity, calibration, and error rates also answer different fairness questions; none establishes all the others.

The second-order effect is a feedback relationship between policy and data. A lending policy determines who generates the outcomes used to develop the next model. Treating that dataset as an unselected sample of all applicants carries the old policy into the new model without identifying its influence.