Business Context

NorthBridge Finance issues unsecured personal loans to roughly 120,000 applicants per month. The risk team wants a production-ready model that predicts whether an applicant will default within 12 months so underwriting can reduce losses without rejecting too many good customers.

Dataset

You are given a historical loan-origination dataset built at application time only. Do not use post-loan repayment behavior as features.

• Rows: 420,000 funded loans from the last 24 months
• Target: default_12m = 1 if the borrower becomes 90+ days past due within 12 months, else 0
• Class balance: 11.4% positive, 88.6% negative
• Missing data: 18% missing in employment fields, 7% in bureau variables for thin-file applicants, <2% elsewhere

Success Criteria

A good solution should achieve strong ranking performance and support threshold-based underwriting decisions. Target ROC-AUC >= 0.84, PR-AUC >= 0.42, and recall >= 0.70 at precision >= 0.35 on a held-out out-of-time test set.

Constraints

• Inference must complete in <50 ms per application in an online API
• The risk team needs feature-level explanations for adverse action review
• Retraining can happen weekly; feature generation must be reproducible in batch and online
• The model must be robust to missing values and moderate class imbalance

Deliverables

1. Build a binary classification pipeline for default_12m
2. Compare at least one linear baseline and one tree-based ensemble
3. Explain feature engineering, leakage prevention, and validation design
4. Select an operating threshold for underwriting based on precision/recall tradeoffs
5. Report final test metrics and the top risk drivers