Business Context

Microsoft Store offers device financing for small-business customers in several markets. The credit risk team needs a model to predict whether an application will become 90+ days delinquent within 12 months, and they want a clear justification for why a particular model should be chosen over simpler or more complex alternatives.

Dataset

You are given a historical underwriting dataset built from Microsoft Dynamics 365, internal payment history, and third-party bureau attributes.

• Rows: 240K applications over 4 years
• Target: default_12m = 1 if account becomes 90+ days delinquent within 12 months
• Class balance: 8.7% positive, 91.3% negative
• Missing data: 18% missing in bureau attributes for thin-file applicants; 6% missing in relationship features for new customers

Success Criteria

A good solution should outperform a regularized logistic regression baseline while remaining explainable enough for risk review. Target ROC-AUC >= 0.84, PR-AUC >= 0.42, and recall >= 0.70 at precision >= 0.35 on the holdout set.

Constraints

• Predictions are used in a regulated credit workflow, so model choice must be defensible
• Batch scoring must finish in under 10 minutes for 100K applications
• Explanations must be understandable to risk analysts and auditors
• Retraining is allowed monthly, not daily

Deliverables

1. Build and compare at least two candidate models and justify the final choice
2. Define preprocessing for mixed data types and missing values
3. Explain feature engineering decisions and leakage checks
4. Choose an evaluation strategy and threshold aligned to business cost
5. Provide model interpretation outputs suitable for a Microsoft risk review