Business Context

LendWise, a mid-size digital lender processing ~300K consumer loan applications per year, wants a simpler and more stable default-risk model. The current model uses every available field, which has increased training time, reduced interpretability, and introduced noisy predictors.

Dataset

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

• Rows: 120K approved loans from the last 24 months
• Target: default_12m (1 = defaulted within 12 months, 0 = repaid/current)
• Class balance: 18% positive, 82% negative
• Missing data: ~9% missing in employment and income-related fields; ~4% missing in bureau variables

Success Criteria

A good solution should reduce the feature set meaningfully while maintaining predictive quality. Target performance is ROC-AUC = 0.78 and F1 = 0.50 on the holdout set, with no more than a 2-point ROC-AUC drop versus a full-feature baseline.

Constraints

• The risk team needs a feature selection approach that is explainable and reproducible.
• Training must complete within 30 minutes on a standard CPU machine.
• The final feature set should be stable enough for monthly retraining and model governance reviews.

Deliverables

1. Build a baseline model using all candidate features.
2. Implement at least two feature selection methods (for example: filter, embedded, or wrapper methods).
3. Compare selected feature sets using cross-validated performance and holdout metrics.
4. Recommend a final feature set and explain why it is appropriate for production.
5. Identify risks such as leakage, correlated variables, and instability across retraining windows.