credit-risk-scorer

Credit risk scorer

Scores an applicant for the German Credit dataset, where the positive class is a bad credit risk. Promoted on expected loss rather than on accuracy.

Metrics

Measured on 200 held-out applicants at a decision threshold of 0.45.

The cost matrix is the dataset's own: a missed default costs 5 times a rejected good applicant. Accuracy is reported because it is how a classifier is usually described, and on this problem it is the least useful of the three: the threshold that maximises it is not the one that minimises loss.

Provenance

Held-out data digest ba349129c0b0c828, which the platform's gate compares before it compares any score. A candidate measured on other data is refused as unjudged rather than promoted.

The artifact is a pickle and executes on load. The library versions it was written under are recorded in release.json, and the platform refuses to serve it under versions that could reconstruct the estimator differently.

Trained by Credit-Risk-Assessment and operated by ML-Platform.