Business Context

NorthStar Bank wants a credit risk model to predict whether a personal loan applicant will default within 12 months. The risk team needs a feature selection process that improves model performance, reduces overfitting, and keeps the final model explainable for regulatory review.

Dataset

• Size: 120K loan applications, 40 candidate features
• Target: Binary — defaulted within 12 months (1) vs repaid/current (0)
• Class balance: 11% positive (default), 89% negative
• Missing data: 12% missing in employment-related fields, 7% missing in bureau-derived variables

Success Criteria

A good solution should improve generalization over a no-selection baseline, achieve stable validation performance, and produce a reduced feature set that risk analysts can review. The final model should use no more than 15 features while maintaining strong recall on defaulters.

Constraints

• The model must be interpretable enough for model risk and compliance teams.
• Batch scoring should complete in under 5 minutes for 200K applications.
• The feature selection process must avoid leakage and be reproducible.

Deliverables

1. Build a baseline classification model using all candidate features.
2. Design and implement a feature selection process using at least two methods (for example: filter, embedded, or wrapper).
3. Compare model performance before and after feature selection using cross-validation and a held-out test set.
4. Explain why selected features were retained and why others were dropped.
5. Recommend a final feature set and model for production use.