Credit risk explorer

German Credit data (1,000 applications, 1994). Everything on this page is precomputed by the repository's train and evaluate runs and read from results/explorer_data.json; nothing is fitted in the browser.

Research demo on small, old data. Not for real lending decisions.

Model comparison

Feature importance (SHAP, XGBoost)

Mean absolute SHAP value per original feature on the test split, in log-odds of bad credit. One-hot columns are summed back to their source feature.

Threshold and cost

An applicant is approved when the predicted probability of bad credit is below the threshold. Cost per applicant uses the test split's predicted probabilities and true labels for the model you pick.

Model

Threshold:

Cost of approving a bad applicant: (rejecting a good one costs 1)

model cost curveapprove everyone / reject everyoneselected threshold