Business Context

NorthStar Lending wants a credit risk model to predict whether a personal loan applicant will default within 12 months. The current model uses all available columns and is difficult to explain to risk analysts, so the team wants a feature selection approach that improves generalization while keeping the model interpretable.

Dataset

• Size: 120K loan applications, 46 candidate features
• Target: Binary — default within 12 months (1) vs no default (0)
• Class balance: Moderately imbalanced — 18% positive, 82% negative
• Missing data: 12% missing in banking features, 4% missing in employment fields, <1% elsewhere

Success Criteria

A good solution should improve out-of-sample performance over a baseline logistic regression using all features, while reducing the feature set to a smaller, explainable subset. Target performance is ROC-AUC >= 0.78 and PR-AUC >= 0.42 on the holdout set with no material calibration degradation.

Constraints

• Risk analysts must understand why features were kept or removed
• Inference must complete in under 20 ms per application
• The final feature set should be stable across retraining cycles
• The solution must avoid leakage from post-application variables

Deliverables

1. Propose and implement a feature selection pipeline for this classification problem
2. Compare at least two selection methods (for example: filter, embedded, or wrapper)
3. Train a final predictive model using the selected features
4. Report performance on validation and test sets with clear metrics
5. Explain which features were selected and why they are useful or risky
6. Describe how you would monitor feature drift and selection stability in production