Business Context

Northstar Mutual wants a model to flag potentially high-risk auto insurance claims for manual review before payout. The data is ambiguous: several fields are incomplete at claim creation, some features are weak proxies for downstream outcomes, and label quality is imperfect because fraud investigations are not always conclusive.

Dataset

You are given a historical claims dataset collected over 24 months.

• Size: 148K claims, 40 candidate features
• Target: Binary — claim later confirmed as high-risk for manual escalation (1) vs standard processing (0)
• Class balance: 11.6% positive, 88.4% negative
• Missing data: 18% missing in external/vendor fields, 9% missing in behavioral fields, and some categorical levels drift over time

Success Criteria

A good solution should achieve strong ranking quality while remaining explainable enough for operations and compliance teams. Target PR-AUC >= 0.42, ROC-AUC >= 0.80, and precision >= 0.45 at 70% recall on a held-out time-based test set.

Constraints

• Predictions must be generated in under 50 ms per claim in an online API.
• The final model must be explainable at both global and per-claim levels.
• Avoid data leakage from post-claim investigation fields.
• Retraining is allowed monthly; feature computation must be reproducible in batch and online settings.

Deliverables

1. Define a feature selection strategy for ambiguous, partially noisy data.
2. Build at least one baseline and one stronger production candidate model.
3. Design a validation plan that accounts for temporal drift and ambiguous labels.
4. Report metrics, threshold selection logic, and feature importance.
5. Explain how you would detect overfitting, leakage, and unstable features before deployment.