Business Context

NorthPeak Lending, a digital consumer lender processing about 120K loan applications per month, wants a default-risk model for instant underwriting. The current model performs well offline but degrades noticeably in production, suggesting overfitting to historical training data.

Dataset

You are given a supervised binary classification dataset for predicting whether a loan will default within 12 months of origination.

• Size: 240K historical applications, 40 modeled features
• Target: default_12m — 1 if the loan defaulted within 12 months, else 0
• Class balance: 14% positive, 86% negative
• Missing data: 6-18% missing in income verification and employment fields; sparse missingness in credit bureau features

Success Criteria

A good solution should improve generalization on unseen data, not just training accuracy. Target test AUC-ROC >= 0.78, test F1 >= 0.45, and keep the train-test AUC gap below 0.04 after applying regularization and model selection controls.

Constraints

• Predictions must be returned in <50 ms per application
• Risk and compliance teams require reasonable interpretability
• Retraining happens monthly; feature engineering should remain maintainable
• Avoid leakage from post-origination variables

Deliverables

1. Explain the main causes of overfitting in this dataset and training setup.
2. Build a baseline model and at least one regularized model.
3. Compare train vs validation/test performance and show how regularization changes the gap.
4. Use cross-validation to tune regularization strength.
5. Recommend a production-ready approach and justify the tradeoffs.