Business Context

LendWise, a mid-size digital lender processing about 250K personal loan applications per year, wants a predictive model to estimate whether an applicant will default within 12 months of origination. The model will support underwriting decisions and risk-based pricing, so it must be accurate, explainable, and stable in production.

Dataset

You are given a historical loan-origination dataset with one row per funded loan.

• Size: 120K funded loans, 38 input features
• Target: default_12m — whether the borrower defaulted within 12 months
• Class balance: 11% positive (default), 89% negative
• Missing data: 8% missing in employment-related fields, 5% in revolving utilization, and up to 18% in some bureau-derived attributes

Success Criteria

A good solution should achieve strong ranking performance and usable recall for the risk team. Target performance is ROC-AUC >= 0.82, PR-AUC >= 0.42, and recall >= 0.70 at precision >= 0.35 on a held-out test set.

Constraints

• Predictions must be generated in <50 ms per application in an online underwriting API.
• The model must be explainable enough for risk review and adverse-action reason reporting.
• Training can run offline daily or weekly; inference cost should remain low.

Deliverables

1. Build an end-to-end supervised classification pipeline from raw data to evaluated model.
2. Explain your preprocessing, feature engineering, model selection, and validation strategy.
3. Compare at least one interpretable baseline with a stronger non-linear model.
4. Select an operating threshold based on business tradeoffs, not accuracy alone.
5. Describe how you would deploy, monitor, and retrain the model in production.