Answers are grounded in this paper and cite their sources.
A credit model outputs a probability that an applicant will turn bad. The lending decision needs a cut-off, and many systems default to 0.5 or re-weight the classes to make the data look balanced. I prove three results. First, if accepting a bad applicant costs c_FN and rejecting a good one costs c_FP, the cost-minimising rule rejects when the calibrated probability of bad is at least c_FP/(c_FP + c_FN). Second, the extra cost of any other cut-off equals (c_FP + c_FN) times the expected distance of the affected applicants' probabilities from the optimal cut-off, which is quadratic in the error for small errors. Third, re-weighting the bad class by a factor w multiplies the model's odds by w, so the optimal cut-off on the re-weighted scale is wτ/(wτ + 1 − τ). On the German credit data (1,000 applicants, 300 bad, donor cost ratio 5 to 1) with 20 repeats of 5-fold cross-validation, a logistic model at cut-off 0.5 costs 0.912 per applicant, worse than rejecting everyone (0.700). The theoretical cut-off of 1/6 costs 0.537, within 0.0016 of the best cut-off found by search, a saving of 0.375 per applicant (bootstrap 95% interval 0.272–0.480). A class-balanced model used with 1/6 costs 0.564; with the corrected cut-off of 0.318 it costs 0.530.
No takes yet. Share an insight, caveat, or question.
Manoj Kumar Aguru (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: