SYNTHETIC RESEARCH

Same review budget.
Different risk policies.

Five frozen methods score the same 480 mature applications. Training, calibration and test periods are separated by outcome-maturation gaps. Each policy receives exactly the same global review capacity; ties use application ID.

Capacity comparison

Method	Budget fraction	Reviewed	Events captured	False positives	Event capture
business_rule	0.05	24	9	15	8.4%
business_rule	0.1	48	17	31	15.9%
business_rule	0.2	96	37	59	34.6%
raw_logistic	0.05	24	12	12	11.2%
raw_logistic	0.1	48	25	23	23.4%
raw_logistic	0.2	96	41	55	38.3%
calibrated_logistic	0.05	24	12	12	11.2%
calibrated_logistic	0.1	48	25	23	23.4%
calibrated_logistic	0.2	96	40	56	37.4%
woe_logistic	0.05	24	10	14	9.3%
woe_logistic	0.1	48	22	26	20.6%
woe_logistic	0.2	96	39	57	36.4%
hist_gradient_boosting	0.05	24	10	14	9.3%
hist_gradient_boosting	0.1	48	18	30	16.8%
hist_gradient_boosting	0.2	96	39	57	36.4%

Follow the evidence

Scores and ranks · Portfolio, month and segment metrics · Label-free raw-logistic explanations · Protocol, WOE scorecard and paired intervals · Artifact hashes

The fixed business rule is a ranking score, not a default probability. WOE boundaries and event rates are learned only from training applications. Raw-logistic contributions add to its log odds relative to training means; they do not explain calibrated probabilities or establish causality. Higher scorecard points imply lower fitted risk; a 20-point increase doubles fitted good/bad odds.

Paired intervals resample complete origination months using the same draw for all methods. Only six held-out months are available; inspect the intervals and group results before interpreting a small difference. Run creditvintage verify-benchmark DIRECTORY to independently recompute capacities, group outcomes and log-odds explanations.

Same review budget.Different risk policies.

Capacity comparison

Follow the evidence

Same review budget.
Different risk policies.