Business Context
LendWise, a mid-size digital lender processing about 120K personal loan applications per month, wants a credit risk model to predict 90-day default before underwriting decisions are made. The current model performs well offline but degrades noticeably after deployment, and the risk team suspects overfitting.
Dataset
| Feature Group | Count | Examples |
|---|
| Applicant profile | 12 | age, employment_length, annual_income, housing_status |
| Credit bureau | 15 | fico_score, revolving_utilization, delinquency_count, inquiries_6m |
| Loan attributes | 8 | loan_amount, interest_rate, term_months, purpose |
| Behavioral / derived | 10 | debt_to_income, credit_age_months, utilization_trend_3m |
| | |
- Size: 240K historical applications, 45 features
- Target: Binary — defaulted within 90 days of origination (1) vs non-default (0)
- Class balance: 11.5% positive, 88.5% negative
- Missing data: 18% missing in employment_length, 9% in bureau variables for thin-file applicants, <2% elsewhere
Success Criteria
A strong solution should improve generalization on an unseen holdout set and clearly explain how overfitting is detected and prevented. Good enough means achieving ROC-AUC >= 0.80, PR-AUC >= 0.42, and keeping the train-test AUC gap below 0.03.
Constraints
- Predictions must be returned in <50 ms per application
- The risk team needs interpretable drivers for adverse action review
- Retraining is allowed monthly; feature generation must work in batch and online scoring
Deliverables
- Define overfitting in the context of this default prediction problem.
- Build a baseline model and a less-overfit model, then compare train/validation/test performance.
- Show at least three prevention methods (for example: regularization, feature selection, cross-validation, early stopping, or limiting model complexity).
- Recommend a final production-ready approach and justify the tradeoffs.
- Provide code to train, evaluate, and diagnose overfitting using appropriate metrics and plots/tables.
