Consider a lender using a credit scoring model to decide which loan applicants to approve. Every applicant has a credit score that correlates with, but never perfectly predicts, whether they will actually default. The lender earns interest margin on loans that get repaid and absorbs a substantial loss on loans that default, the two outcomes pulling in opposite directions on every approval decision.
The threshold is the whole business lever
Lowering the approval threshold expands the pool of approved applicants, generating more interest revenue. But the newly-approved applicants sit lower on the score distribution, where the model itself predicts higher default risk, so a meaningful share of that additional volume will default and generate losses that can outweigh the extra revenue.
Why the safest threshold is not the best one
Raising the threshold far enough eliminates most default risk, but it also turns away a large share of applicants who would have repaid reliably and been profitable customers. An overly conservative threshold protects against losses at the direct cost of the revenue those declined-but-actually-safe applicants would have generated.
Modeling the trade-off explicitly
In practice, lenders model expected portfolio profit across a range of thresholds using historical default data, choosing the threshold that maximizes net expected profit rather than simply minimizing default rate or maximizing approval volume in isolation. Regulatory fairness constraints often add further limits on how the threshold, and any underlying model, can vary by protected characteristics.
Try it yourself
The AI Credit Scoring Lab simulates a thousand loan applicants with realistic score-versus-default correlation, letting you move the threshold directly and watch approval count, defaults, and net portfolio profit respond.
🧪 Try it yourself: the AI Credit Scoring Lab simulation lets you experiment with everything described above directly in your browser.