Project Hertha, a major joint initiative between the Bank for International Settlements (BIS) Innovation Hub and the Bank of England (BOE), has become a milestone in the ongoing quest to improve financial crime detection in real-time retail payment systems. As the scale and complexity of retail payments skyrocket, criminal networks are increasingly exploiting gaps between financial institutions and payment service providers (PSPs). Project Hertha is an ambitious response, leveraging innovative analytics, artificial intelligence, and inter-institutional collaboration to strengthen anti-money laundering (AML) defenses across the payment landscape.

The initiative recognizes that as financial transactions move ever faster and become more fragmented, the traditional approaches to detecting illicit activity face severe limitations. The ability of individual banks to spot criminal behavior is often limited by their view of customer activity, which can be obscured when funds are transferred through multiple accounts, banks, and payment methods. Project Hertha’s core ambition is to test whether new technological tools—especially advanced network analytics—can bridge these gaps, enabling a broader, more precise, and privacy-respecting approach to financial crime detection.

The Objective: Enabling Smarter, Privacy-Preserving AML in Real-Time Payments

At its heart, Project Hertha was established to answer a challenging but crucial question: can financial crime be detected more accurately in real-time retail payment systems by using sophisticated network analytics, without compromising user privacy or operational efficiency? BIS and the Bank of England were determined to show that significant improvements in detection could be achieved with a minimal data footprint, moving away from invasive, data-heavy solutions that risk violating customer trust.

More specifically, Project Hertha set out to:

• Assess whether advanced transaction analytics, powered by artificial intelligence and machine learning, could meaningfully boost the detection rates of money laundering and other financial crimes in real-time retail payment ecosystems.
• Demonstrate how network-wide data, viewed from the perspective of a payment system operator, could supplement (rather than replace) the monitoring systems of individual banks and PSPs.
• Ensure that any gains in detection did not come at the expense of customer privacy, focusing only on the smallest necessary set of data points.
• Identify how collaborative analytics might reduce the burden of false positives, streamline compliance, and minimize friction for legitimate customers.

The project was underpinned by a recognition of several pressing realities in the modern financial ecosystem. Today, criminal groups move illicit funds with ever-increasing sophistication, using chains of transactions that cross not only institutional but often national boundaries. As technology brings new payment methods—including digital wallets and instant payment services—there is a growing risk that traditional, siloed AML controls will become less and less effective. Project Hertha was conceived to confront these risks head-on, positioning the BIS and BOE as leaders in next-generation AML strategy.

Methodologies: Synthetic Data, Machine Learning, and Realistic Simulation

A key innovation of Project Hertha lies in its methodological rigor and reliance on advanced data science. To ensure results were both robust and privacy-respecting, the project team constructed a complex synthetic dataset representing an entire national retail payment ecosystem. No real customer data were used; instead, AI-driven models simulated 1.8 million bank accounts and 308 million transactions over a one-year period, reflecting realistic consumer behaviors, income, spending patterns, and use of various financial services.

Building a Synthetic Payments Ecosystem

The development of the synthetic dataset unfolded in several stages:

• Data Modeling: Generative AI models were used to create detailed transaction histories for a large universe of artificial customers, mirroring the demographics, economic activity, and spending profiles found in UK retail banking statistics.
• Account Assignment: Customers were distributed across eight simulated banks, including both digital-first and traditional players, to capture the diversity of the real payments landscape.
• Validation and Iteration: The generated data were repeatedly validated and improved, ensuring that transaction volumes, value distributions, and customer behaviors closely matched real-world figures, despite being fully synthetic.
• Financial Crime Simulation: Expert input and published typologies informed the creation of 2,000 simulated money laundering schemes, covering ten distinct patterns of criminal behavior, from scatter-gather and fan-in/fan-out schemes to simple cycles and cross-border layering.

This comprehensive, artificial ecosystem enabled the BIS and BOE teams to rigorously test the limits of network analytics for AML purposes, while avoiding the legal and ethical complications of handling real transaction data.

Testing Transaction Analytics

The crux of Project Hertha’s experimental design was to compare the effectiveness of financial crime detection under three main scenarios:

1. Banks/PSPs in Isolation: Each institution analyzed its own customer transactions and payment data, reflecting the standard approach to AML.
2. Payment System Operator Alone: The operator examined network-level transaction flows using only a limited set of data points (such as pseudonymous sender/receiver identifiers, transaction time, amount, and purpose), with no access to underlying customer identity.
3. Collaborative Analytics: Insights from network analytics (run by the payment system operator) were fed back to banks and PSPs, who combined these risk indicators with their own monitoring systems to target suspicious activity more accurately.

A wide variety of machine learning models were evaluated, including both traditional supervised algorithms (which rely on labeled examples of past illicit activity) and unsupervised models (which look for statistical anomalies without prior examples). The project also assessed new deep learning architectures specifically designed to handle large, complex transaction datasets with high class imbalance—a realistic challenge in AML, where illicit activity is rare compared to legitimate transactions.

Measuring Success

Key metrics included:

• Recall: The proportion of illicit accounts correctly flagged as suspicious.
• Precision: The proportion of flagged accounts that were truly illicit (i.e., not false positives).
• Average Precision: A combined metric summarizing model performance at various thresholds, with higher scores indicating better overall discrimination between legitimate and criminal activity.

Results: Network Analytics Offer Real Gains in Detection and Efficiency

Project Hertha produced a series of concrete findings that reshape how the financial industry might approach AML in the future.

Improved Detection Rates through Collaboration

One of the headline results was that collaboration between payment system operators and banks/PSPs leads to clear improvements in the identification of illicit accounts. Specifically:

• Combined Approaches Outperform Isolation: When banks and PSPs leveraged network-level risk scores generated by the payment system, detection rates improved by 12% on average compared to when each party acted alone.
• Detection of Novel Criminal Typologies: For previously unseen money laundering schemes—those not captured in historical data—network analytics delivered a 26% improvement in detection rates. This is crucial given how rapidly criminal tactics evolve.
• Focus on Complex Schemes: The benefits of network analytics were most pronounced in schemes that spanned multiple institutions and relied on coordinated activity across several accounts—precisely the types of threats that are hardest to catch using traditional, institution-centric monitoring.

Machine Learning and Deep Learning Insights

The experiments revealed important distinctions between different types of analytical models:

• Supervised Learning Is Key: Supervised machine learning, which requires labeled data on past cases, proved far more effective than unsupervised anomaly detection. While the latter produced many false positives, supervised models could zero in on genuine risk with much higher accuracy.
• Deep Learning for New Patterns: Deep learning models, particularly those based on transformer architectures for structured data, excelled at detecting new and evolving patterns of financial crime. This adaptability is essential in a threat landscape where innovation is constant.
• Data Efficiency: All these advances were achieved using minimal, non-personal transaction data. The focus on pattern analysis over personal identifiers shows that robust AML can be compatible with privacy protections.

Balancing Detection and False Positives

A persistent challenge in AML is the trade-off between catching more illicit accounts and generating false alarms that disrupt legitimate customers. Project Hertha demonstrated that by carefully calibrating the thresholds for risk scoring, payment system operators could tune their models to prioritize either higher precision (fewer false positives) or higher recall (catching more illicit activity), according to their operational needs.

The Critical Role of Feedback Loops

The research underscored the importance of continuous feedback between payment system analytics and the banks/PSPs that investigate and confirm suspicious activity. High-quality, labeled training data—often derived from confirmed investigations—are essential for keeping detection models effective over time. The feedback loop concept tested by Project Hertha establishes a template for future collaborative AML frameworks.

Policy, Privacy, and Operational Considerations

Project Hertha did not propose changes to the legal responsibilities of payment system operators or banks. Instead, it highlighted how advanced analytics can supplement existing roles, provided there is careful attention to privacy, data protection, and regulatory requirements. Any future deployment of such tools will require:

• Consent from institutions and customers as data controllers
• Strict minimization and protection of any personal data processed
• Transparent model design and, where possible, explainable AI outputs to support investigations and regulatory reporting

The Road Ahead: Next Steps and Industry Implications

The insights from Project Hertha point to a new phase in AML technology and regulatory strategy, but also highlight areas requiring further development.

Extending to Other Payment Types and Networks

While Project Hertha’s focus was on national retail payment systems, its methodologies could be adapted to other areas where financial crime risks are high:

• Large-Value Payment Systems: These often involve fewer, but much higher-value, transactions and pose distinct monitoring challenges.
• Cross-Border Payments: Detecting illicit flows across jurisdictions remains a major AML blind spot, and collaborative analytics could close some of those gaps.
• Cryptoasset Networks: Applying network analytics to digital asset ledgers (such as Bitcoin or Ethereum) could enable earlier detection of laundering and fraud involving privacy coins, mixers, and decentralized exchanges.

Enabling Transaction Tracing and Collaborative Investigations

Future initiatives might move from predictive analytics (spotting high-risk activity before it is confirmed) to transaction tracing—following illicit funds through networks to uncover criminal connections. Banks and PSPs could also benefit from secure, privacy-enhancing technologies that enable collaborative investigations without exposing customer data to unnecessary risk.

Integrating Explainable AI and Model Transparency

To support effective investigations and regulatory compliance, AI systems should increasingly provide clear explanations for their risk assessments—such as the key variables and transaction features that triggered a flag. Explainable AI will become even more critical as deep learning models grow more complex and “black box” in nature.

Cost and Operational Efficiency

Project Hertha estimated that banks globally spend over $200 billion annually on compliance with financial crime regulations. Improved network analytics could allow institutions to reduce the operational burden associated with high false positive rates (which in some cases reach 95%), focusing resources on genuinely high-risk cases and delivering better outcomes at lower cost.

Fostering Regulatory and Policy Support

International bodies such as the Financial Action Task Force (FATF) are already advocating for risk-based, outcome-focused supervision of AML/CFT controls. For network analytics to achieve their full potential, policy frameworks will need to support inter-institutional data sharing, clarify roles and responsibilities, and establish standards for privacy and model governance.

Conclusion: Project Hertha as a Model for the Future of AML

Project Hertha stands as a landmark in financial crime prevention, setting new standards for how technology, collaboration, and regulatory oversight can combine to address the evolving threat landscape. Through the joint leadership of the BIS Innovation Hub and the Bank of England, the project proves that smarter, privacy-respecting network analytics can substantially improve AML effectiveness in real-time retail payment systems.

By designing a realistic, scalable test environment, employing state-of-the-art AI models, and focusing on the operational realities of modern payments, Project Hertha has delivered actionable insights for regulators, financial institutions, and technology providers worldwide. Its legacy will shape the next generation of compliance technology, industry best practices, and regulatory frameworks.

As financial ecosystems continue to evolve and as criminals develop ever more sophisticated tactics, the lessons of Project Hertha provide a robust blueprint for innovation, collaboration, and the relentless pursuit of integrity in the global financial system.

---

Related Links

• BIS Project Hertha: Identifying financial crime patterns in real-time retail payment systems
• Bank of England: Payment and Settlement Services
• FATF Guidance on Risk-Based Supervision
• Europol: The Other Side of the Coin – Analysis of Financial and Economic Crime
• LexisNexis: The True Cost of Financial Crime Compliance

Other FinCrime Central Articles About State-Wide Initiatives

• Innovative Fintel Alliance Boosts AUSTRAC’s Fight Against Money Laundering
• UK Invests £121 Million in Quantum Technology to Combat Financial Crime and Fraud
• Malaysia Proposes Blockchain Identity System to Fight Growing Fraud
• Using Federated Learning for AML in Hong Kong Banks

Source: The BIS report

Some of FinCrime Central’s articles may have been enriched or edited with the help of AI tools. It may contain unintentional errors.