Business Context

LendWise, a mid-size digital lender processing roughly 200K personal loan applications per quarter, wants a default-risk model that is accurate, stable, and explainable to credit analysts. The current model uses nearly every available column, causing leakage risk, unstable coefficients, and poor generalization.

Dataset

You are given an offline training dataset of historical approved loans and repayment outcomes. The task is to design a feature selection strategy and train a binary classifier that predicts whether a borrower will default within 12 months of origination.

• Rows: 120K loans over 24 months
• Target: default_12m (1 = defaulted within 12 months, 0 = did not default)
• Class balance: 14% positive, 86% negative
• Missing data: 12% missing in savings/income-related fields, 6% missing in bureau variables, some categorical levels are rare

Success Criteria

A strong solution should improve generalization over a model trained on all raw features, maintain interpretable drivers for risk review, and achieve ROC-AUC >= 0.80 with PR-AUC >= 0.42 on a held-out test set.

Constraints

• Credit analysts need a defensible explanation for why features were kept or removed
• Avoid target leakage and highly unstable correlated features
• Batch scoring must complete in under 5 minutes for 50K applications
• Retraining happens monthly, so the pipeline must be reproducible

Deliverables

1. Build a feature selection pipeline for mixed numerical and categorical data
2. Compare at least two approaches (e.g., model-based selection vs. regularized baseline)
3. Explain how you detect leakage, multicollinearity, and low-value features
4. Evaluate the final model on a held-out test set with concrete metrics
5. Provide the final selected feature set and rationale