Business Context

LendWise, a mid-size digital lender processing ~120K personal loan applications per month, wants a binary classifier to predict whether an approved applicant will default within 12 months. The risk team currently uses a Logistic Regression scorecard, but recent analysis suggests nonlinear interactions between credit behavior, income stability, and prior delinquencies are being missed.

Dataset

• Size: 240K historical applications, 42 features
• Target: default_12m — 1 if the borrower defaults within 12 months, else 0
• Class balance: 11.5% positive, 88.5% negative
• Missing data: ~8% missing in banking features, ~3% in bureau history for thin-file applicants

Success Criteria

A solution is considered good enough if it improves minority-class detection over Logistic Regression while maintaining operational precision: ROC-AUC >= 0.82, PR-AUC >= 0.42, and recall >= 0.70 at precision >= 0.45 on a holdout test set.

Constraints

• Batch scoring only; inference must score 120K applications in under 10 minutes.
• The risk team needs feature importance and clear justification for choosing Random Forest over Logistic Regression.
• Retraining should be feasible monthly without specialized GPU infrastructure.

Deliverables

1. Build and compare a Logistic Regression baseline and a Random Forest classifier.
2. Explain why Random Forest is or is not the better production choice for this dataset.
3. Handle missing values, mixed feature types, and class imbalance appropriately.
4. Report threshold-free and threshold-based metrics on validation and test sets.
5. Summarize key tradeoffs: interpretability, calibration, latency, and maintenance.