Prasanna K · Credit risk and model validation
Risk ranks.
Probabilities drift.
Six mortgage vintages, a frozen model decision and two untouched evaluation cohorts. Inspect why a better score still does not justify operational approval.
Validation selected unchanged logistic and retrained boosting. The fixed .0001 simplicity tolerance recommends logistic. Display controls never refit or reselect a model.
Review allocation
| Group | Loans | Mean predicted | Observed |
|---|
Capture bounds assign unresolved outcomes adversarially. They are identification bounds, not confidence intervals or effects of reviewing borrowers. Calibration groups are descriptive.
Conditional loss scenario
Probability × outstanding balance × assumed severity. These research probabilities are miscalibrated; this identity is not an approved provision or expected realized loss estimate.
Serious delinquency is not a financial loss. Review effectiveness is unidentified. No avoided losses, net benefits or profitability are computed.
Keep the failure cases visible
At the primary 2020 test, observed events are 8.324 times logistic predictions and 6.155 times boosting predictions. A favorable paired Brier contrast does not correct that probability failure. The separate 2023 check cannot be used as a second selection set.
300,000 loans · 20,151,667 monthly records · 16 numeric and four categorical predictors · four unique full fits and two intercept updates · no parameter search. Raw rows, identifiers, individual predictions and fitted models are excluded from this explorer.
Confidence intervals resample loans and can understate common-shock dependence. Protected attributes are absent; no fairness claim is supported. Historical SQL cutoffs do not establish immutable historical-release availability.