Business Context

LendWise, a mid-size digital lender processing roughly 300K consumer loan applications per year, wants a default-risk model for underwriting. The current model uses nearly every available column, causing unstable performance, weak interpretability, and avoidable retraining issues when upstream fields change.

Dataset

You are given a historical loan application dataset and asked to design a feature selection strategy for a binary classification model predicting whether an approved loan will default within 12 months.

• Size: 240K loan applications, 45 candidate features
• Target: Binary — default within 12 months (1) vs no default (0)
• Class balance: 14% positive, 86% negative
• Missing data: 18% missing in savings_balance, 9% in employment fields, <3% elsewhere

Success Criteria

A good solution should improve generalization while reducing the feature set to a maintainable subset. Target performance is ROC-AUC e 0.80 and PR-AUC e 0.42 on a held-out test set, with no more than 20 selected features and a clear explanation of why they were kept or removed.

Constraints

• The underwriting team requires feature-level interpretability.
• Inference must complete in under 50 ms per application.
• Avoid features that are expensive to collect or likely to be unavailable at scoring time.
• The selected feature set should be stable enough for monthly retraining.

Deliverables

1. Define a feature selection methodology combining business rules and statistical/model-based methods.
2. Build a baseline model and a reduced-feature model.
3. Compare performance using cross-validation and a final holdout test set.
4. Explain how you handle missing values, correlated variables, and leakage risks.
5. Provide the final selected feature list and justify tradeoffs between accuracy and simplicity.