Explain XGBoost for Loan Default

Business Context

FinEdge, a digital lending platform processing about 250K consumer loan applications per month, wants to improve its default-risk model. The credit team currently uses logistic regression, but they want a stronger baseline and a clear explanation of when XGBoost is the right choice for tabular classification problems.

Dataset

You are given a historical loan application dataset and asked to build a binary classifier that predicts whether an approved loan will default within 12 months.

• Size: 420K loan records, 36 features
• Target: default_12m — 1 if the borrower defaults within 12 months, else 0
• Class balance: 11.4% positive, 88.6% negative
• Missing data: 9% missing in employment and verification fields, 3% missing in some bureau attributes

Success Criteria

A good solution should outperform logistic regression and random forest on recall and PR-AUC, while remaining practical for batch scoring. A strong answer should also explain what XGBoost is, why it works well on structured data, and its tradeoffs versus simpler models.

Constraints

• Daily batch inference on 250K applications must finish in under 10 minutes
• Risk analysts need feature importance and reason codes for adverse action review
• Retraining should be feasible weekly on standard cloud CPU instances

Deliverables

1. Explain what XGBoost is and how it differs from a single decision tree and random forest
2. Train an XGBoost classifier for loan default prediction
3. Compare it against a logistic regression baseline
4. Evaluate with metrics appropriate for imbalanced classification
5. Summarize key tradeoffs around accuracy, interpretability, and operational complexity