Business Context

NovaBank, a digital lender processing ~120K personal loan applications per month, wants a default-risk model for pre-approval decisions. The hiring manager is less interested in raw modeling than in whether you can clearly justify why you chose a specific library and algorithm for a real production problem.

Dataset

You are given a historical loan dataset built from application-time features only.

• The dataset contains both numerical and categorical variables.
• Some categorical fields have moderate cardinality.
• Missingness is concentrated in employment and bureau-derived fields.
• The target is whether the applicant defaults within 12 months of origination.

Success Criteria

A strong solution should achieve a holdout ROC-AUC of at least 0.78 and PR-AUC above the default-rate baseline, while remaining explainable enough for risk and compliance review. Your explanation of why you chose the library and algorithm should be specific, evidence-based, and tied to the dataset and production constraints.

Constraints

• Inference latency must stay under 50 ms per application.
• The model must support reason-code style explanations.
• Retraining will happen monthly.
• The first production version should use the Python ecosystem already adopted by the risk team.

Deliverables

1. Train a baseline and a final classification model for 12-month default prediction.
2. Explain why your chosen library and algorithm are appropriate versus at least one alternative.
3. Show preprocessing, validation, and threshold selection.
4. Report evaluation metrics on a held-out test set.
5. Provide a short production-oriented justification covering interpretability, latency, and maintainability.