Business Context

PayWave processes roughly 8 million card transactions per day. Fraud is rare but costly, and the risk team needs a model that identifies suspicious transactions for manual review without overwhelming analysts with false positives.

Dataset

You are given a historical transaction dataset for binary classification.

• Size: 1.2M transactions, 40 features
• Target: is_fraud — 1 if the transaction was confirmed fraudulent within 30 days, else 0
• Class balance: 0.9% positive, 99.1% negative
• Missing data: ~12% missing in merchant risk fields for new merchants; ~4% missing in device attributes

Success Criteria

A good solution should improve minority-class detection while keeping review volume manageable. Target at least 70% recall on fraud cases with precision >= 15% on the flagged set, and clearly justify threshold selection. Accuracy alone is not acceptable.

Constraints

• Batch scoring every 5 minutes; per-transaction inference should remain under 50 ms
• Model output must support analyst review with explainable top drivers
• False negatives are expensive, but false positives increase manual review cost
• Training pipeline should be reproducible and avoid leakage from future information

Deliverables

1. Build a baseline and an improved classifier for this imbalanced dataset
2. Explain how you handle class imbalance in training and evaluation
3. Design preprocessing for missing values, categorical variables, and skewed numeric features
4. Choose decision threshold(s) based on business tradeoffs, not default 0.5
5. Report evaluation on a held-out test set using appropriate metrics
6. Summarize what you would deploy and how you would monitor it in production