Business Context

LendWise, a digital consumer lending platform processing ~250K loan applications per month, wants a default-risk model to support instant underwriting decisions. You need to design a solution using XGBoost or LightGBM that predicts whether a newly issued loan will become 90+ days delinquent within 12 months.

Dataset

• Size: 420K historical loans, 47 features
• Target: Binary — defaulted within 12 months (1) vs non-default (0)
• Class balance: 11.4% positive, 88.6% negative
• Missing data: 18% missing in banking features, 6% in employment fields, and sparse missingness in bureau pulls

Success Criteria

A good solution should achieve ROC-AUC >= 0.84, PR-AUC >= 0.42, and improve approval-quality decisions over the current scorecard baseline while keeping false approvals within policy limits. The model should also provide feature-level explanations for adverse action review.

Constraints

• Online inference latency must stay below 50 ms per application
• The risk team needs reasonably interpretable outputs
• Retraining budget is limited to a weekly batch pipeline
• The solution must avoid target leakage from post-origination signals

Deliverables

1. Choose XGBoost or LightGBM and justify the choice for this underwriting problem.
2. Define preprocessing and feature engineering for mixed numeric/categorical data with missing values.
3. Propose a train/validation/test strategy that reflects production deployment.
4. Train and evaluate the model with appropriate metrics and threshold selection.
5. Explain how you would deploy, monitor, and refresh the model in production.