CREDITVINTAGE / RESEARCH CONTROL PANEL

Know what was known.

SYNTHETIC EXAMPLE

Application-time features, 180-day outcome maturity and chronological train / calibration / test windows. The default demo uses generated loans and makes no claim about a real lender.

Out-of-time loans480 2024 Jul–Dec vintage
Observed 90+ DPD22.3% six-report horizon
Calibrated Brier0.158 constant baseline 0.174
Review capture23.4% at 10% fixed capacity

Vintage drift

Observed default rate and predicted mean by origination month.

2024-07
21.2%mean PD 18.6% · n=80
2024-08
21.2%mean PD 17.7% · n=80
2024-09
18.8%mean PD 16.8% · n=80
2024-10
30.0%mean PD 22.0% · n=80
2024-11
20.0%mean PD 20.6% · n=80
2024-12
22.5%mean PD 17.0% · n=80

How to read this

ROC-AUC 0.735; average precision 0.407. The vintage-block bootstrap 95% ROC-AUC interval is 0.653–0.801; only six test months are available.

Raw logistic Brier 0.155; calibration increased it to 0.158 in this run. Calibration is reported as a measured step, not assumed to improve every sample.

Scenario expected loss $663,835.47 assumes an unchanged LGD of 45% and origination principal as exposure. It is not a measured recovery loss, accounting provision, or regulatory capital estimate.

Training outcomes are mature before calibration applications; calibration outcomes are mature before test applications. Test labels never enter fitting or threshold selection.

Risk buckets

Bucket 1 has the highest predicted probability. Rates are descriptive and may be noisy.

BucketLoansMean PD 90+ DPDScenario EL
14854.8%52.1%$179,984.69
24827.6%31.2%$91,583.87
34820.5%39.6%$86,049.15
44818.1%33.3%$64,520.30
54814.3%25.0%$56,051.37
64814.3%16.7%$47,018.59
74814.3%6.2%$51,640.55
84811.2%8.3%$42,989.03
9489.4%0.0%$33,582.43
10483.3%10.4%$10,415.48