Business Context

LendWise, a mid-size digital lending platform processing about 120K personal loan applications per quarter, wants a model to predict whether an approved borrower will default within 12 months. The risk team needs a simple, explainable model to support underwriting decisions and reduce charge-offs without materially lowering approval volume.

Dataset

The training data contains historical funded loans from the last 3 years.

• Rows: 240K funded loans, 33 input features
• Target: default_12m — whether the borrower becomes 90+ days delinquent within 12 months
• Class balance: 11.5% default, 88.5% non-default
• Missing data: ~8% missing in employment and income verification fields; ~3% missing in revolving utilization

Success Criteria

A good solution should achieve ROC-AUC >= 0.78, recall >= 0.60 at precision >= 0.35, and provide coefficients or feature effects that the credit policy team can explain to auditors.

Constraints

• The model must be interpretable enough for regulated lending review.
• Batch scoring should complete in under 10 minutes for 150K applications.
• Retraining should be feasible monthly with limited ML ops support.

Deliverables

1. Build a binary classification model to predict 12-month default.
2. Explain why the chosen algorithm fits this business problem.
3. Describe preprocessing for missing values, categorical variables, and skewed numeric features.
4. Evaluate the model using threshold-free and threshold-based metrics.
5. Recommend a decision threshold for underwriting based on business tradeoffs.