Validate Loan Default Model

Context

LendWise built a binary classification model to predict whether a personal loan applicant will default within 90 days. The model is scheduled for production deployment to support underwriting decisions, but the risk team is concerned that headline accuracy may hide costly approval mistakes.

Current Validation Results

The Problem

The model looks strong on accuracy and AUC-ROC, but it only identifies 42% of actual defaulters at the current threshold. Missing defaulters is expensive because approved bad loans create direct credit losses, while false positives reduce approval volume and customer growth.

Requirements

1. Explain whether this model is ready for production and which metrics matter most.
2. Interpret the tradeoff between accuracy, precision, recall, and F1 given the 12% default rate.
3. Describe how you would validate the model beyond a single holdout set before deployment.
4. Recommend threshold, calibration, and error analysis steps before launch.
5. Propose specific improvements if the business wants to reduce missed defaults without sharply lowering approval rates.

Constraints

• Each missed defaulter costs about $4,800 on average.
• Each falsely declined good applicant costs about $320 in expected contribution margin.
• The underwriting team requires a clear validation plan before approving deployment.