Business Context

Northstar Bank uses a binary classification model to predict whether a new loan applicant will default within 12 months. The current notebook-based model works offline on sampled data, but the lending platform now needs a production-ready approach that can score 1.5M applications per month with stable latency, reproducible training, and measurable business impact.

Dataset

• Rows: 2.4M historical applications over 4 years
• Target: default_12m = 1 if the borrower defaults within 12 months of origination
• Class balance: 8.7% positive, 91.3% negative
• Missing data: ~12% missing in bureau fields, ~4% missing in self-reported income fields

Success Criteria

A good solution should improve ranking quality over the current logistic regression baseline, support batch and low-latency online inference, and provide enough transparency for risk and compliance review. Target online p95 inference latency is under 50 ms per request, with monthly retraining and full reproducibility.

Constraints

• Predictions must be explainable enough for model risk review
• Training must fit on a single medium-sized cloud instance
• Online serving must tolerate schema changes and missing values
• The model should degrade gracefully if one upstream feature source is delayed

Deliverables

1. Propose a production-scalable modeling approach and explain why it fits this dataset.
2. Build a training pipeline with preprocessing, validation, and threshold selection.
3. Define offline evaluation metrics and acceptance thresholds.
4. Describe how you would deploy, monitor, and retrain the model in production.
5. Identify key tradeoffs between accuracy, latency, interpretability, and operational complexity.