Business Context

PayLink processes roughly 8 million card transactions per day. Fraud is rare but expensive, so the risk team needs a model that catches fraudulent transactions without overwhelming manual reviewers with false positives.

Dataset

You are given a historical transaction dataset for binary classification: predict whether a transaction is fraudulent (is_fraud=1) or legitimate (is_fraud=0). The data is sampled from 6 months of production traffic.

• Size: 1.2M transactions, 30 features
• Target: Binary — fraudulent transaction (1) vs legitimate transaction (0)
• Class balance: 0.9% positive, 99.1% negative
• Missing data: ~12% missing in distance_from_home_km, 7% in ip_risk_score, and sparse missingness in merchant metadata

Success Criteria

A good solution should achieve recall >= 0.85 on fraud while keeping precision >= 0.25 at the operating threshold. The model should also improve ranking quality enough to support manual review queues.

Constraints

• Batch scoring every 5 minutes; per-row inference should remain low-latency
• False negatives are costly, but false positives create reviewer load and customer friction
• The fraud team needs feature-level explanations for flagged transactions

Deliverables

1. Build and justify a modeling approach for this highly imbalanced classification problem.
2. Show how you would preprocess missing values and mixed feature types.
3. Compare at least one baseline against a stronger model.
4. Choose evaluation metrics appropriate for class imbalance and explain threshold selection.
5. Describe how you would deploy, monitor, and retrain the model in production.