Business Context

NorthBridge Capital is a mid-sized consumer lender processing roughly 120,000 personal loan applications per month. The credit risk team wants a production-ready default prediction model to support underwriting decisions and expected loss forecasting.

Dataset

You are given an offline training dataset of historical funded loans. Each row represents one originated loan with borrower attributes captured at approval time and repayment outcomes observed over 12 months.

• Rows: 640,000 funded loans over 4 years
• Target: default_12m = 1 if the loan becomes 90+ days past due or charged off within 12 months
• Class balance: 8.7% positive, 91.3% negative
• Missing data: 18% missing in banking features, 6% missing in bureau variables, and some categorical levels drift over time

Success Criteria

A good solution should outperform a regularized logistic regression baseline and be suitable for underwriting review. Target ROC-AUC >= 0.82, PR-AUC >= 0.38, and recall >= 0.70 at precision >= 0.30 on the holdout period.

Constraints

• Predictions must be generated in < 50 ms per application in an online decisioning service.
• The model must remain auditable and reasonably interpretable for risk and compliance stakeholders.
• Training data is temporal, so leakage from future vintages is not allowed.

Deliverables

1. Build a binary classification model to predict 12-month default risk.
2. Explain model choice versus simpler financial modeling approaches such as scorecards or logistic regression.
3. Design preprocessing for missing values, mixed feature types, and temporal leakage prevention.
4. Evaluate the model with threshold-dependent and ranking metrics relevant to lending.
5. Describe how you would deploy, monitor, and retrain the model in production.