01 / THE PRACTICAL QUESTION
A decision, before a model.
A lender can rank applicants correctly while systematically overstating their default probabilities. Those are different failures with different consequences.
Compare a simple credit model with boosting, checking ranking and probability accuracy separately.
02 / DATA & COMPARISON
The idea in plain language.
Discrimination asks whether defaulters tend to receive higher scores. Calibration asks whether, among applications assigned a similar probability, defaults occur at roughly that rate.
- AUC
- A ranking measure: 0.5 represents chance ordering; 1 represents perfect separation.
Data. Synthetic labeled credit observations with separate held-out evaluation.
Baseline. Logistic regression with held-out calibration checks.
03 / THE EXPERIMENT
What the saved experiment shows.
On 1,000 held-out synthetic applications, logistic regression records AUC 0.882, versus 0.861 for boosting. Extra complexity does not improve this ranking comparison.
Inspect the supporting resultEvidence record: research-validation.json#numerical_checks/03
Explore the related lab04 / RESULTS & LIMITATIONS
Evidence with its boundaries attached.
The related lab is a cross-project demonstration. Read this investigation’s evidence and limits before transferring its conclusions.
Inspect numerical checks and validation records
Loading validation evidence…
Synthetic performance does not establish fairness or real-world lending suitability.
05 / REPRODUCE
Reproduce and challenge the result.
Code, configuration, and reproduction
The project contains its implementation, configuration, tests, and walkthrough. Download the lab configuration to record the exact parameters used in an interactive run.
Project code and walkthrough