Business Context

BrightLine Credit, a digital consumer lender processing ~250K applications per month, wants a model to predict 90-day loan default at origination. The goal is to improve approval quality without materially reducing funded volume.

Dataset

You are given an offline training dataset built from the last 18 months of applications. Each row represents one funded or declined application with features available at decision time.

• Size: 420K applications, 45 features
• Target: Whether the applicant defaults within 90 days of origination
• Class balance: 11.6% positive (default), 88.4% negative
• Missing data: ~9% missing in bureau features, ~6% missing in self-reported financial fields

Success Criteria

A solution is considered good enough if it improves ranking quality over the current scorecard and supports a policy that reduces default rate by at least 15% at a fixed approval rate. The model should also provide clear feature-level explanations for risk review.

Constraints

• Inference must complete in <50 ms per application in the online decision API
• The risk team requires interpretable drivers for adverse action review
• Retraining should be feasible on a monthly cadence using standard Python ML tooling

Deliverables

1. Build a supervised learning pipeline to predict 90-day default probability.
2. Compare at least one interpretable baseline and one stronger non-linear model.
3. Describe how you would handle missing values, categorical variables, and probability calibration.
4. Propose an evaluation framework aligned to lending decisions, not just raw accuracy.
5. Explain the main technical hurdles you expect in production and how you would mitigate them.