Context

LendWise uses a binary classification model to predict whether a personal loan applicant will default within 90 days. The model was deployed 6 months ago to support automated approvals and manual underwriting, but risk leaders now report worsening portfolio quality despite similar application volume.

Current Performance

The Problem

The model still identifies many risky applicants correctly, but it is missing a larger share of actual defaulters than it did at launch. Product and risk teams want to know whether the decline is caused by threshold choice, data drift, changing applicant mix, or model staleness.

Requirements

1. Interpret what the metric changes suggest about the model's current behavior.
2. Identify the most likely root causes of the decline and rank them by likelihood.
3. Explain how you would use confusion matrix analysis and segment-level evaluation to diagnose the issue.
4. Recommend specific actions to improve performance in the next 2-4 weeks.
5. Discuss what tradeoffs LendWise should make if improving recall lowers precision.

Constraints

• Manual underwriting capacity is limited to 4,000 applications per week.
• False negatives are costly because missed risky applicants become charge-offs.
• False positives reduce approval rate and hurt revenue growth.