Context

LendWise uses a binary classification model to predict whether a personal loan applicant will default within 12 months. The team is unsure how to interpret a seemingly strong ROC-AUC score because business results at the current decision threshold are only moderate.

Current Performance

The Problem

The risk team sees ROC-AUC = 0.86 and calls the model “strong,” but loan officers point out that recall is low and many defaults are still being approved. You need to explain what ROC-AUC actually means here, what it does not guarantee, and whether the model is good enough for deployment at the current threshold.

Requirements

1. Interpret the ROC-AUC score in plain language for this model.
2. Explain how ROC-AUC should be read alongside precision, recall, and accuracy.
3. Assess whether threshold = 0.50 is appropriate for this use case.
4. Identify what additional evaluation would help before changing the model or threshold.
5. Recommend practical next steps to improve decision quality.

Constraints

• Each false negative costs about $7,500 in expected credit losses.
• Each false positive costs about $900 in lost profit and customer friction.
• The underwriting team can manually review at most 2,000 additional applications per month.