Business Context

NorthStar Bank uses a supervised learning model to predict whether a personal loan applicant will default within 12 months. The current model performs well on training data but degrades noticeably after deployment, so the team wants a more robust approach that explicitly addresses overfitting.

Dataset

You are given a historical loan dataset used for binary classification.

• Size: 120K loan applications, 32 input features
• Target: Binary — defaulted within 12 months (1) vs paid as agreed (0)
• Class balance: 18% positive, 82% negative
• Missing data: 8% missing in employment_length, 5% in annual_income, <2% elsewhere

Success Criteria

A good solution should improve generalization, not just training accuracy. Target performance is ROC-AUC >= 0.82 on a held-out test set, with a train-test AUC gap <= 0.03 and stable cross-validation results.

Constraints

• The risk team requires reasonable interpretability
• Batch scoring must complete in under 5 minutes for 50K applications
• The solution should avoid leakage and be easy to retrain monthly

Deliverables

1. Build a baseline model and show evidence of overfitting
2. Propose and implement at least two overfitting prevention techniques
3. Compare a simpler regularized model against a higher-capacity model
4. Use cross-validation and a true holdout test set
5. Explain which approach you would ship and why