Business Context

LendWise, a mid-size digital lender processing ~120K personal loan applications per month, wants to improve its default-risk model before the next underwriting policy update. The current model is a regularized logistic regression, but performance has plateaued and the credit risk team wants a structured hyperparameter tuning approach that improves ranking quality without sacrificing interpretability.

Dataset

• Size: 240K historical applications, 38 features
• Target: Binary — default within 12 months (1) vs no default (0)
• Class balance: 14.6% positive, 85.4% negative
• Missing data: 9% missing in banking features, 3% missing in employment fields, categorical unknowns present

Success Criteria

A successful solution should improve model quality over the current baseline and produce a repeatable tuning workflow. Good enough means achieving ROC-AUC >= 0.80, PR-AUC >= 0.42, and F1 >= 0.48 at a threshold chosen on validation data, while keeping the final model explainable to risk analysts.

Constraints

• The final model must support explanation at the feature level for compliance reviews.
• Batch scoring latency must stay under 10 minutes for 120K applications.
• Training should run on a single standard CPU machine.
• The team prefers scikit-learn-compatible tooling and monthly retraining.

Deliverables

1. Build a baseline classification pipeline for default prediction.
2. Tune key hyperparameters using cross-validation and justify the search space.
3. Compare baseline vs tuned model on a held-out test set.
4. Report concrete evaluation metrics and the selected decision threshold.
5. Summarize which hyperparameters mattered most and the tradeoffs of the tuned configuration.