Problem
Context
LendWise uses a binary classification model to predict whether a personal loan applicant will default within 90 days. The model was trained six months ago and recently expanded from prime borrowers into near-prime segments, but risk leaders are concerned that headline accuracy still looks strong while portfolio losses are rising.
Current Performance
| Metric | Validation at Launch | Current Holdout | Change |
|---|---|---|---|
| Accuracy | 0.91 | 0.89 | -0.02 |
| Precision | 0.72 | 0.61 | -0.11 |
| Recall | 0.68 | 0.49 | -0.19 |
| F1 Score | 0.70 | 0.54 | -0.16 |
| AUC-ROC | 0.84 | 0.78 | -0.06 |
| Log Loss | 0.29 | 0.37 | +0.08 |
| Default rate in sample | 0.12 | 0.16 | +0.04 |
| Monthly charge-off losses | $1.8M | $2.6M | +$0.8M |
The Problem
The model still appears acceptable on accuracy, but recall has fallen sharply and losses have increased after the borrower mix changed. You need to assess whether the model is robust enough for current production use and identify what evaluation steps would give confidence before retraining or changing thresholds.
Requirements
- Interpret the metric changes and explain why accuracy may be misleading here.
- Diagnose the most likely robustness risks affecting current performance.
- Recommend a validation framework to test robustness across time, segments, and thresholds.
- Propose specific actions to improve model reliability and business outcomes.
- Explain the tradeoff between approving more loans and catching more likely defaulters.
Constraints
- False negatives (missed defaulters) cost about $4,200 per loan.
- False positives (incorrectly declining good applicants) cost about $320 in lost contribution margin.
- Compliance requires explainable decisions and stable approval rates across major borrower segments.
You are practicing as a guest. Sign up free to get your answer graded with AI feedback. Your draft stays right here.