Business Context

NorthBridge Bank issues unsecured personal loans to roughly 1.2M applicants per year across the U.S. The credit risk team wants a model that predicts whether an applicant will become 90+ days delinquent within 12 months of origination so underwriting can reduce losses without materially hurting approval volume.

Dataset

• Size: 420K originated loans over 4 years, 52 features
• Target: Binary label indicating 90+ DPD within 12 months of origination
• Class balance: 8.7% positive (default), 91.3% negative
• Missing data: 22% missing in employment_length, 11% in bureau variables for thin-file applicants, <3% elsewhere

Success Criteria

A solution is considered good enough if it improves ranking quality over the current scorecard and achieves ROC-AUC >= 0.80, PR-AUC >= 0.32, and recall >= 70% at precision >= 25% on an out-of-time test set. The model should also provide stable, regulator-friendly explanations.

Constraints

• Inference must complete in <50 ms per application in the online underwriting API.
• The risk team requires reason codes or interpretable feature contributions.
• Training data is temporal; leakage from post-origination signals is not allowed.
• The model should be retrained quarterly and monitored for drift by applicant segment.

Deliverables

1. Build a binary classification model for 12-month default risk.
2. Explain feature engineering, missing-data handling, and class imbalance strategy.
3. Use a time-aware validation scheme and justify it.
4. Choose an operating threshold aligned to underwriting tradeoffs.
5. Provide evaluation metrics, model interpretation, and deployment considerations.