Business Context

LendWise, a mid-sized digital lender processing about 120K personal loan applications per month, wants a reliable binary classifier to predict 90-day default risk at application time. The team is less interested in naming optimization libraries abstractly and more interested in whether you can choose, tune, and justify practical optimization tools in a real training workflow.

Dataset

You are given a historical loan-origination dataset for supervised classification.

• Size: 240K applications, 36 features
• Target: default_90d — whether the borrower defaults within 90 days of origination
• Class balance: 11.5% positive, 88.5% negative
• Missing data: 9% missing in employment and income-related fields, 3% missing in bureau variables

Success Criteria

A good solution should outperform a regularized logistic regression baseline and achieve strong ranking quality for underwriting decisions. Target ROC-AUC >= 0.84, PR-AUC >= 0.46, and recall >= 0.70 at precision >= 0.35 on the held-out test set.

Constraints

• Inference must complete in <20 ms per application in an online API.
• Risk and compliance teams require feature-level explainability.
• Retraining should fit within a 2-hour weekly batch window.
• The approach must avoid leakage from post-origination variables.

Deliverables

1. Build and compare at least two models with different optimization/training approaches.
2. Explain which optimization tools or libraries you used and why they fit this dataset.
3. Implement preprocessing, hyperparameter tuning, and threshold selection.
4. Report validation and test metrics, including business-relevant threshold metrics.
5. Summarize production tradeoffs: latency, interpretability, and retraining cost.