Business Context

PayFlow, a mid-size digital payments company, currently blocks suspicious card transactions using hard-coded rules such as amount > 1000 and country != billing_country. These rules are easy to explain but miss new fraud patterns and require constant manual updates. You need to design an ML-based fraud classifier and explain how it differs from a rule-based system in behavior, maintenance, and performance.

Dataset

You are given historical transaction-level data labeled after chargeback resolution.

• Size: 420K transactions over 9 months, 25 input features
• Target: Binary label indicating whether the transaction was confirmed fraudulent
• Class balance: 2.9% fraud, 97.1% legitimate
• Missing data: 12% missing in device fingerprint fields, 4% missing in historical behavior for new users

Success Criteria

A good solution should outperform the current rule engine on fraud recall while keeping analyst review volume manageable. Target at least 75% recall with precision above 20% and provide a clear comparison between static rules and a trained model.

Constraints

• Inference latency must stay under 50 ms per transaction
• Risk team needs feature importance and reason codes for flagged transactions
• The model must be retrained at least monthly because fraud patterns drift

Deliverables

1. Build a baseline hard-coded rule system and report its metrics.
2. Train an ML classification model on the labeled dataset.
3. Compare rule-based vs ML approaches in terms of adaptability, maintenance, and error patterns.
4. Propose preprocessing and feature engineering for mixed data and missing values.
5. Recommend a deployment threshold based on business tradeoffs between fraud loss and false positives.