Business Context

PayWave processes roughly 8 million card transactions per day. Fraud losses are rising, and the risk team needs a binary classifier that identifies fraudulent transactions while keeping false positives low enough to avoid blocking legitimate customers.

Dataset

You are given a historical transaction-level dataset for supervised fraud detection.

• Size: 2.4M transactions, 36 features
• Target: is_fraud (1 = fraudulent transaction, 0 = legitimate)
• Class balance: 0.9% fraud, 99.1% non-fraud
• Missing data: 18% missing in merchant risk features for new merchants, 6% missing in device attributes

Success Criteria

A good solution should achieve strong minority-class detection without relying on accuracy. Target PR-AUC above 0.35, recall above 75% at precision above 20%, and provide a thresholding strategy the fraud operations team can tune based on review capacity.

Constraints

• Batch scoring every 5 minutes; average inference latency should stay under 50 ms per transaction
• Model outputs must be explainable enough for fraud analysts to review flagged transactions
• False positives have direct customer impact, so threshold selection matters as much as model choice

Deliverables

1. Build a classification pipeline for highly imbalanced fraud data.
2. Explain and implement at least two imbalance-handling techniques.
3. Choose evaluation metrics appropriate for rare-event classification.
4. Recommend a decision threshold based on business tradeoffs.
5. Describe how you would monitor precision and recall after deployment.