Business Context

NovaCredit, a mid-market consumer lending fintech, wants to improve its pre-approval underwriting model for personal loans. The current rules-based system approves too many risky applicants and rejects some creditworthy ones, so the team needs a machine learning model that predicts 90-day default risk at application time.

Dataset

You are given a historical application dataset built from loan originations over the last 24 months.

• Rows: 420K loan applications, 46 usable features after basic cleaning
• Target: default_90d — whether the borrower missed payments severely enough to be classified as default within 90 days of origination
• Class balance: 8.7% positive class
• Missing data: ~12% missing in bureau fields, 6% in self-reported income-related fields, and some categorical unknowns from third-party integrations

Success Criteria

A good solution should improve risk ranking enough to support underwriting decisions and achieve:

• ROC-AUC >= 0.84
• PR-AUC >= 0.42
• Recall >= 70% at precision >= 35% for the high-risk review queue
• Calibrated probabilities suitable for threshold-based approval policies

Constraints

• Inference must complete in <100 ms per application in an online API.
• Risk and compliance teams require global and local explainability.
• Retraining should be feasible monthly on commodity cloud infrastructure.
• Avoid leakage from post-origination signals.

Deliverables

1. Build a binary classification pipeline for default_90d.
2. Explain model choice, feature engineering, and leakage prevention.
3. Compare at least one interpretable baseline with a stronger nonlinear model.
4. Tune the decision threshold for underwriting or manual review.
5. Report offline metrics and discuss deployment tradeoffs.