Credit-default model with fairness checks and a reason for every decision

Harbourline approves about 6,000 credit-card applications a month. Its current scorecard was built five years ago and the regulator has asked two pointed questions: can you show the model treats similar applicants similarly regardless of sex and age, and can you tell every declined customer the main reasons in plain language?

The chief risk officer wants a replacement model, but she is clear that accuracy alone will not get it approved. The model-risk committee will look at three things: how well it ranks risk, whether approval and error rates differ across groups by more than agreed limits, and whether the reasons given to customers are faithful to what the model actually did.

You will use the public UCI "Default of credit card clients" dataset (30,000 customers, 23 features including credit limit, payment history and bill amounts, plus sex, education, marriage and age). You will train an interpretable baseline and a gradient-boosted model, pick an approval cut-off for a target default rate of 6% among approved applicants, and then audit both models for group fairness. Sex and age must not be model inputs, but you will keep them aside for the audit.

The committee's limits: - approval rates between groups within a ratio of 0.8; - true-positive rates within 5 percentage points; - each decline must come with up to four reasons, each tied to a feature that genuinely lowered the score.

Done looks like a model pack: code, a validation report with AUC and calibration, a fairness report with the numbers against the limits and any mitigation you tried, a reason-code generator with tests, and a short memo to the committee.