Business Context

LendWise, a mid-size digital lender processing about 120K consumer loan applications per month, wants a default-risk model that is accurate, explainable, and stable in production. The current challenge is not only training a classifier, but selecting a feature set that improves generalization while remaining interpretable for risk and compliance teams.

Dataset

You are given a historical loan application dataset for binary classification: predict whether an approved loan will default within 12 months.

• Size: 240K applications, 44 candidate features
• Target: default_12m (1 = defaulted within 12 months, 0 = repaid/no default)
• Class balance: 14% positive, 86% negative
• Missing data: 18% missing in bureau variables for thin-file applicants, 6% missing in income-related fields

Success Criteria

A good solution should improve validation performance over a simple all-features logistic regression baseline while reducing unnecessary or redundant features. The final feature set should remain explainable enough for model risk review.

Constraints

• Prefer interpretable models or interpretable feature selection outputs
• Batch scoring only; latency is not strict, but retraining should be practical monthly
• Avoid leakage from post-application or post-origination variables

Deliverables

1. Propose and implement a feature selection strategy for this dataset.
2. Compare at least two approaches (for example: filter, embedded, or wrapper methods).
3. Train a baseline and final model using the selected features.
4. Report validation and test metrics, and explain why features were kept or removed.
5. Call out risks such as leakage, multicollinearity, instability across folds, and fairness concerns.