Business Context

PayFlow processes roughly 8 million card transactions per day. Fraud is rare but costly, and the risk team needs a model that flags suspicious transactions without overwhelming manual reviewers with false positives.

Dataset

You are given a historical transaction dataset for binary classification.

• Size: 1.2M transactions, 40 features
• Target: is_fraud (1 = fraudulent, 0 = legitimate)
• Class balance: 0.7% fraud, 99.3% non-fraud
• Missing data: ~12% missing in device fields, ~4% missing in merchant metadata, sparse missingness elsewhere

Success Criteria

A good solution should improve fraud capture materially over a majority-class baseline. Target at least 75% recall on fraud while keeping precision above 15% for the review queue, and achieve PR AUC above 0.30 on the holdout set.

Constraints

• Batch scoring every 15 minutes; average inference latency should stay under 50 ms per transaction.
• The fraud operations team needs feature-level explanations for flagged transactions.
• False positives are expensive because analysts can review only ~5,000 transactions per day.

Deliverables

1. Build a classification pipeline that handles the severe class imbalance correctly.
2. Explain how you would compare class weighting, resampling, and threshold tuning.
3. Choose evaluation metrics appropriate for rare-event detection and justify them.
4. Produce a validation strategy that avoids leakage from customer or time-based patterns.
5. Recommend a production-ready model, decision threshold, and monitoring plan.