(Summary featured here. For full report, download below.)

The Project Hertha report by the BIS Innovation Hub and Bank of England investigates how real-time retail payment systems can be leveraged to detect financial crime through advanced analytics. By using a highly realistic synthetic dataset consisting of 1.8 million bank accounts and 308 million transactions, the project simulates the detection of illicit financial activity such as money laundering. It hypothesizes that payment systems, which have a unique network-wide view, can provide valuable risk indicators to financial institutions without requiring access to sensitive customer information. This approach respects data privacy while aiming to enhance the financial system’s integrity.

The study found that payment system analytics, while slightly less effective on their own compared to individual banks or PSPs, significantly boost detection accuracy when combined with institutional models. This collaborative approach improves the identification of known crime patterns by 9% and new patterns by 26%, with a corresponding reduction in false positives. The most effective outcomes were achieved through active collaboration, where payment system insights are selectively integrated based on past performance, reinforcing a continuous feedback loop for model refinement.

Project Hertha emphasizes the importance of supervised learning, demonstrating that models trained on labeled historical data outperform unsupervised methods, which often produce high false-positive rates. It also confirms that models can be calibrated to focus on a small subset of high-risk accounts, making them useful even in resource-constrained environments. Cutting-edge deep learning models such as UniTTab showed additional gains, especially in detecting emerging fraud patterns, reinforcing the potential of AI in payment system surveillance.

The project highlights key practical implications, such as the need for a robust feedback infrastructure between payment systems and banks/PSPs. This includes methods for sharing risk scores and investigation outcomes securely and efficiently. It also advocates for explainable AI methods to assist institutions in understanding and acting on flagged accounts. Privacy-preserving techniques were central, illustrating that effective monitoring can occur with minimal data, aligning with principles of data minimization and lawful processing.

Looking forward, the report outlines three promising areas for further research: using transaction tracing to map broader criminal networks, enabling collaborative investigations across institutions, and applying analytics to large-value, cross-border, or cryptoasset payment systems. These extensions could make detection more robust across varied financial environments. Overall, Project Hertha presents a blueprint for integrating advanced analytics into payment systems to enhance the global fight against financial crime.