By Will Towning
New research published by the Bank for International Settlements (BIS) shows that graph or network-based machine learning analysis significantly improves the effectiveness and efficiency of anti-fraud and money laundering controls.
The BIS Innovation Hub Nordics’ Project Aurora explored – among other important topics – the effectiveness and efficiency of different approaches to AML monitoring across three scenarios. The scenarios included analysis of synthetic transaction data at the individual financial institution level, national level and cross-border level. The effectiveness and efficiency measures refer to the ability to detect money laundering activity while keeping the number of false positives low.
Under each scenario, graph-based machine learning models detect twice as many money launderers than the traditional rule-based approach, the study finds. Graph-based machine learning models also outperform all other approaches researchers tested, such as isolation forest, logistic regression and artificial neural networks.
On the efficiency side, the study finds graph-based machine learning models also reduce the number of false positives compared to a rule-based approach at the individual financial institution level by between 40-85%.
One important finding shows all approaches are significantly improved when conducting the analysis at the national and cross-border levels. A graph-based machine learning approach works best at these levels because the model can learn from a larger observed network structure.
Why is network-based machine learning so effective?
Money laundering is often a network-based event. Money launders use a number of methods to conceal the source and destination of funds. This includes moving funds through different accounts, payment methods, payment systems and jurisdictions. The activity creates complex and sophisticated transaction trails that span across industries and borders. In Figure 1, the BIS provides a nice example of how money launders use different payment ecosystems.
Figure 1: Simplified view of how money launders use different payment ecosystems
When individual financial institutions are required to identify suspicious activity without access to more holistic system or cross-border level datasets, these sophisticated money launderers can more easily go undetected. The BIS also helps illustrate this with Figure 2.
Figure 2: Difference in the visibility of suspicious networks under a siloed approach compared to a system-level approach
The challenges underscore the importance of understanding transaction networks for AML efforts. Without network-based analysis, experts are naturally limited in their ability to detect network-based money laundering events.
FNA’s advanced network analytics anomaly detection
FNA’s advanced network analytics capabilities help leading central banks and financial market infrastructures to examine the connections between individuals and businesses. This enables experts to uncover hidden patterns in the flow of funds and more sophisticated money laundering techniques.
Our anomaly detection models combine network analytics with machine learning to not only capture information within the transaction data – volumes, timing and frequency – but also the network features. This means that when we are trying to understand if a new payment is suspicious or not, we are also accounting for the network context in which the payment was made, such as the centrality of the sender or receiver, the network distance between them or the communities they belong to.
For further information or to discuss anomaly detection with FNA, get in touch via www.fna.fi
Footnotes:
(1) BIS (2023). Project Aurora: The power of data, technology and collaboration to combat money laundering across institutions and borders. Available at: https://www.bis.org/publ/othp66.pdf