Answers are grounded in this paper and cite their sources.
A credit cut-off does not only refuse people who would have defaulted. At any stated risk appetite it also refuses a measurable number of applicants who, in the event, repaid. This study counts them, prices the alternatives, and then tests a remedy against a pre-registered pass line it fails. On the UCI credit card portfolio (Taiwan, 2005; 30,000 accounts; held-out test set of 9,000), a gradient boosting scorecard (AUC 0.7810, 95 percent bootstrap CI 0.7693 to 0.7919) is cut at the point where the approved book's observed default rate equals the portfolio's performing-book rate of 12.00 percent. At that cut-off (PD 0.2293; 6,375 approved, 2,625 rejected), 1,399 of the 2,625 rejected applicants (53.3 percent, CI 48.2 to 77.6 percent) did not default: the creditworthy rejects. A model restricted to behavioural features, the thin-file situation of a newcomer with no bureau history, scores AUC 0.7687 against the full model's 0.7810, a gap of 0.0123 that confirms the pre-registered hypothesis that behavioural evidence alone carries most of the signal. A fairness audit at the same cut-off finds approval-rate gaps of 3.6 points by sex, 12.6 by age band, and 29.2 by education, with group sizes printed beside every rate. The proposed remedy, a Second-Look Recovery Policy that re-scores rejects with the thin-file model and approves those whose behavioural PD matches the typical approved applicant's, was pre-registered to recover at least 15 percent of creditworthy rejects at a cost of at most 1.00 percentage point of book default rate. It recovered 1.2 percent (17 of 1,399) at a cost of 0.02 points, and is therefore REFUTED on its recovery bound: the strict band is nearly empty because the two models' scores are correlated. The method transports to a second population (UCI Statlog German Credit: logistic regression AUC 0.8022, gradient boosting 0.7848, on 300 test accounts with wide intervals). The contribution is the accounting discipline: every claim in this paper was registered before it was run, every interval prices its whole procedure, and the failed remedy is reported as the finding.
No takes yet. Share an insight, caveat, or question.
Jesun Ahmad Ushno (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: