OBSERVED UCI DATA / FIVE FROZEN SPLITS
30,000 card accounts.
Every split retained.
23,364 non-events and 6,636 native default labels. Each seed uses 18,000 training, 6,000 calibration and 6,000 test records. Scaling and categories fit only on training; isotonic calibration uses a separate partition and a fixed six-decimal score input grid. No tuning or best-seed selection.
The source describes one historical Taiwan cohort and April–September 2005 repayment, bill and payment history. These random customer splits cannot establish future-period performance or reconstruct original information availability.
Test metrics: mean and observed range
| Method | ROC-AUC | Average precision | Brier ↓ |
|---|---|---|---|
| constant | 0.500 (0.500–0.500) | 0.221 (0.221–0.221) | 0.172 (0.172–0.172) |
| logistic | 0.771 (0.755–0.782) | 0.543 (0.527–0.562) | 0.136 (0.133–0.140) |
| isotonic | 0.770 (0.754–0.781) | 0.527 (0.512–0.543) | 0.136 (0.133–0.139) |
| hist gradient boosting | 0.781 (0.766–0.793) | 0.554 (0.537–0.573) | 0.135 (0.132–0.138) |
Ranges describe five overlapping random splits, and are not confidence intervals. The native subsequent-default target differs from the separate synthetic 180-day 90+ DPD experiment.
Compare the same review capacity
Every method receives 300, 600 or 1,200 test records at 5%, 10% or 20% capacity. Scores rank descending; original source ID resolves ties. Inspect all 60 method/seed/budget rows in policy metrics. The 10% queues omit outcome labels. This measures historical ranking, without an observed intervention or loss saving.
Run from the installed package
creditvintage card-benchmark --out outputs/card-credit
creditvintage verify-card outputs/card-credit
The package includes the frozen attributed table. Run and verification need no network. Verification refits every model and checks source bytes, split membership, probabilities, review ranking, summaries, page and evidence ZIP.
Download complete evidence · Protocol and all results · 20 model/split rows · 150,000 assignments · 30,000 held-out rows / 120,000 model scores · Frozen source table · Source and conversion notice · File digests
Validation boundaries
- One historical cohort; random customer splits do not validate future periods
- No publication/revision timestamps: information availability remains unvalidated
- Native subsequent-default target differs from the synthetic 180-day 90+ DPD label
- Overlapping repeated splits: min/max are descriptive, not confidence intervals
- ID is excluded from fitting; demographic exclusion does not demonstrate fairness
- No observed recoveries, intervention effects, live deployment or monetary savings
Source: Yeh (2009), Default of Credit Card Clients, UCI Machine Learning Repository, CC BY 4.0. Original values and row order retained; the Excel table was converted to integer CSV and compressed. IDs and four demographic fields are excluded from model fitting. Native repayment-status codes, including -2 and 0, are retained as categories.