Business Context

LendWise, a mid-size digital lender processing ~120K personal loan applications per month, wants a default-risk model for pre-approval decisions. The current rule-based system drops rows with null values, reducing coverage and potentially biasing decisions.

Dataset

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

• Size: 240K loan applications, 31 input features
• Target: default_12m (1 = default, 0 = non-default)
• Class balance: 14% positive, 86% negative
• Missing data:
  • annual_income: 18% missing
  • credit_score: 11% missing
  • employment_status: 6% missing
  • debt_to_income: 9% missing
  • Several bureau fields are jointly missing for thin-file applicants

Success Criteria

A good solution should improve over row-dropping and simple mean imputation baselines, achieve ROC-AUC >= 0.78, and maintain stable performance across applicants with and without missing bureau data.

Constraints

• Batch scoring must complete in under 10 minutes for 120K applications
• The risk team needs feature-level explanations and a clear missing-data strategy
• No target leakage from post-origination fields

Deliverables

1. Build a classification pipeline that handles missing numerical and categorical values correctly
2. Compare at least two missing-data strategies and justify the final choice
3. Evaluate the model with appropriate classification metrics and subgroup checks
4. Explain how you would detect when missingness patterns change in production
5. Provide production-ready Python code for preprocessing, training, and evaluation