Context

LendWise uses a binary classification model to predict whether a personal loan applicant will default within 90 days. A recently deployed gradient boosting model was expected to reduce credit losses, but post-launch performance is below target and the risk team is questioning whether the model should remain in production.

Current Performance

The Problem

The model is approving more applicants than expected while missing a larger share of true defaulters. Credit losses rose from $1.9M to $2.7M per month over the last quarter, and leadership wants a structured plan to diagnose whether the issue is thresholding, drift, feature quality, or model generalization.

Requirements

1. Interpret the metrics and explain what they imply about model behavior.
2. Identify the most likely root causes of underperformance.
3. Propose a practical error analysis plan using the available outputs.
4. Recommend specific model or evaluation improvements.
5. Explain what tradeoffs you would make if the business wants to reduce losses quickly.

Constraints

• Retraining requires 10 business days.
• Loan decisions must remain explainable for compliance.
• False negatives are much more expensive than false positives.
• The business cannot reduce approval rate below 58% without hurting growth targets.