Context

LendWise uses a gradient boosted binary classifier to predict whether a personal loan applicant will default within 90 days. In the last model review, the team found that the new model improves overall ranking performance versus the production scorecard, but default rates in one applicant segment appear materially underestimated.

Current Performance

The Problem

You need to decide what validation steps are required before presenting these findings to the Head of Risk and recommending rollout. The ranking metrics look better, but threshold metrics, calibration, and segment behavior raise concerns about whether the model is truly safer for deployment.

Requirements

1. Identify which findings need additional validation before stakeholder review.
2. Explain how you would validate metric stability, calibration, and threshold choice.
3. Propose an error analysis plan for the underpredicted segment.
4. Recommend whether the current evidence is sufficient for rollout.
5. Describe what you would present to stakeholders if results remain mixed.

Constraints

• False negatives are costly because missed defaults create direct credit losses.
• Risk policy requires explainable approval decisions.
• Model retraining takes 10 business days, and rollout must be decided this quarter.