Business Context

PayWave processes millions of card transactions per day and wants a fraud detection model to score transactions before approval. Fraud is rare, but missed fraud is expensive, while too many false positives create customer friction and declined legitimate payments.

Dataset

You are given a historical transaction dataset for supervised binary classification.

• Size: 2.4M transactions, 42 features
• Target: is_fraud where 1 = fraudulent transaction, 0 = legitimate transaction
• Class balance: 0.9% positive, 99.1% negative
• Missing data: ~12% missing in device fields, ~4% missing in geo features, some unseen categories in production

Success Criteria

A strong solution should improve fraud capture materially over a majority-class baseline and be evaluated with metrics appropriate for severe class imbalance. Good enough means achieving recall >= 70% at precision >= 20% on the fraud class, with PR-AUC > 0.18 on a held-out test set.

Constraints

• Inference must complete in under 50 ms per transaction.
• The fraud operations team needs feature-level explanations for flagged transactions.
• Training can run offline daily; scoring happens online.
• Avoid data leakage from post-transaction or chargeback-resolution fields.

Deliverables

1. Propose a modeling approach for highly imbalanced binary classification.
2. Define preprocessing and feature engineering for mixed data types and missing values.
3. Choose evaluation metrics and threshold selection strategy appropriate for fraud.
4. Explain how you would validate the model without leakage.
5. Provide production-ready Python code for training, evaluation, and threshold tuning.