Business Context

PayFlow processes roughly 8 million card transactions per day across web and mobile checkout. The risk team wants a fraud detection model that scores each transaction before authorization so high-risk payments can be declined or routed to manual review.

Dataset

You are given a historical transaction dataset labeled after chargeback resolution.

• Size: 3.2M transactions over 9 months, 41 features
• Target: Binary — fraudulent transaction (1) vs legitimate transaction (0)
• Class balance: 0.7% fraud, 99.3% legitimate
• Missing data: 18% missing in device fields, 6% missing in merchant risk features, sparse missingness in customer history for new users

Success Criteria

A strong solution should achieve high recall on fraud while keeping false positives low enough that approval rate is not materially harmed. Target recall >= 75% at precision >= 20%, PR-AUC >= 0.30, and p95 online inference latency < 50 ms.

Constraints

• Real-time scoring at checkout
• Severe class imbalance and delayed labels
• Model outputs must support analyst review with clear feature importance or reason codes
• Retraining budget is daily or weekly, not continuous

Deliverables

1. Build a binary classification pipeline for transaction-level fraud detection.
2. Explain feature engineering choices and how you prevent leakage.
3. Describe how you handle imbalance, missing data, and categorical variables.
4. Choose decision thresholds for decline vs manual review vs approve.
5. Report offline metrics and discuss production monitoring for drift and calibration.