Context

LendWise uses a binary classification model to predict whether a small-business loan applicant will default within 90 days. The current gradient boosting model was retrained last quarter and is now used to auto-approve, auto-decline, or send applications to manual underwriting.

Product and risk teams report conflicting feedback: approval volume increased, but early default losses also rose. Manual underwriters say many borderline applications being routed to review are actually low risk, while some defaults are slipping through auto-approval.

Current Performance

The Problem

Despite higher accuracy and AUC, the current model is performing worse on the business outcome that matters: catching likely defaulters before approval. You need to evaluate the model, explain the metric tradeoffs, and recommend how to iterate based on stakeholder feedback.

Requirements

1. Interpret why accuracy and AUC improved while recall and F1 worsened.
2. Diagnose likely causes using the metrics and business feedback.
3. Recommend how you would analyze errors by segment and decision threshold.
4. Propose concrete model or policy changes for the next iteration.
5. Explain what success metrics you would use in the next validation cycle.

Constraints

• Manual underwriting capacity is capped at 25% of applications.
• False negatives are costly because defaults create direct credit losses.
• False positives reduce funded-loan volume and hurt revenue growth.