Question:
Can we build a lightweight surrogate model with high recall of Fraud transactions that can be used as a preprocessing filter before a more complex model/human checks for Fraud?

Problem:
Financial fraud is a growing challenge that costs financial institutions worldwide billions of dollars annually.

Project proposal: Build a lightweight surrogate model that screens transactions for potential fraud, and then sends those transactions through more intensive models/human analysts for verification. The goal is to reduce computational load, speed detection, and focus resources.

Data:
JPMorgan Chase Payment Data (synthetically generated)
1.49 million transactions (electronic transfers, bill payments, deposits, withdrawals) spanning approximately 50 years.

Approach:
We engineer both transactional and network features. Network features track fraud communities and identify proximity to fraud.

Models: Logistic Regression, XGBoost,