Context

LendWise uses a binary classification model to predict whether a personal loan applicant will default within 12 months. The team reports that the model has 92% accuracy, but default losses are still higher than expected and risk analysts believe the metric being emphasized may be misleading.

Current Performance

The Problem

The business wants to understand which evaluation metrics matter most for this model and why accuracy alone is not enough. You need to interpret the current results, explain the trade-offs between common classification metrics, and recommend what LendWise should optimize for before the next release.

Requirements

1. Explain what each primary metric (accuracy, precision, recall, F1, AUC-ROC) tells you in this case.
2. Identify why the model can have high accuracy but still perform poorly for the business goal.
3. Recommend which 2-3 metrics should be prioritized for loan default prediction and justify your choice.
4. Describe how threshold changes would affect precision and recall.
5. Suggest concrete next steps to improve evaluation and model performance.

Constraints

• False negatives are expensive: each missed default costs about $8,000 on average.
• False positives reduce approvals and revenue: each wrongly rejected good applicant costs about $1,200 in expected profit.
• Risk leadership wants a simple metric dashboard for weekly review.