← Korede Folarin · Read the case study A portfolio project · synthetic data
Sentinel Credit
Decision policy · held-out evidence

Where should the linebe drawn?

A model produces a probability; the business decides the cutoff. Move it and watch the trade-off: how many applications go to manual review, how many defaults are caught, and who bears the errors. Results are the champion's real held-out test outcomes, aggregated into score bands.

Review cutoff
Sent to review
Defaults caught
Precision
Bad rate, not reviewed

Workload against defaults caught

Navy: defaults caught. Grey: applications sent to review.

Defaulters missed, by credit-file depth

Navy: thin credit file. Grey: established credit file.

Error rates by segment at this cutoff

Missed = defaulters scored below the cutoff (false-negative rate). Wrongly reviewed = non-defaulters sent to review (false-positive rate). Vulnerability is a monitoring label, never a model input.

What this shows

This is why the cutoff is treated as policy, not a model setting: the trade-off between workload, losses and who carries the errors needs an owner. Segment-specific thresholds were considered and deliberately not built without that sign-off. The configured cutoff assumes a 15% review capacity, which is still a placeholder.