Business Context

LendWise, a digital consumer lending platform processing ~250K loan applications per month, wants to improve its loan default model before expanding credit approval volume. You need to tune a deep learning model that predicts whether an applicant will default within 90 days of origination.

Dataset

The training data contains applicant, bureau, and behavioral features collected at application time.

• Size: 320K applications, 48 input features
• Target: Binary label indicating default within 90 days
• Class balance: 11.6% default, 88.4% non-default
• Missing data: 12% missing in bureau features, 4% missing in self-reported income/employment fields

Success Criteria

A strong solution should improve ranking quality over a logistic regression baseline and produce a stable model suitable for monthly retraining. Target business performance is AUC-ROC >= 0.84, PR-AUC >= 0.42, and recall >= 0.70 at precision >= 0.35.

Constraints

• Batch inference must score 250K applications in under 10 minutes.
• The model will be retrained monthly on the latest 18 months of data.
• The risk team needs reproducible tuning and clear justification for selected hyperparameters.
• Avoid data leakage from post-origination features.

Deliverables

1. Define a hyperparameter tuning strategy for a deep learning classifier.
2. Build a preprocessing and training pipeline for mixed tabular data.
3. Compare at least one baseline model against the tuned neural network.
4. Select an operating threshold using business-relevant metrics.
5. Explain tradeoffs among model quality, training cost, and production reliability.