Banks and other financial institutions face thousands of decision problems every day. Some are simple, and decisions can be made quickly and efficiently by a single person or handled through automation. Others are on a different scale and require much more effort to be solved properly. These problems often have a significant bottom-line impact and, in the case of Terrorist Financing (TF) or Money Laundering (ML), also regulatory impact.
Consider, for example, a bank like Banking Circle which has built the first and only real-time clearing and settlement network for 24 currencies to deliver faster, lower-cost payments. The AML team’s role is crucial to this mission as they ensure all solutions are fully secure and compliant all the way from onboarding to live TF/ML monitoring of payments.
This requires a more sophisticated and powerful payment screening decision system than the traditional rules-based AML setup where decisions are made on one or two variables such as amount, date, location of the payee, etc.
The problem with a rules-based approach is that, from a machine learning perspective, the rules look a lot like a shallow, badly tuned decision tree (see figure 1). They are most often tuned by hand, and decision points are adjusted using intuition and an understanding of the way things have always been done. This approach generally leads to false positive rates of 97-99%. Rules-based systems are usually improved by adding rules and tweaking where decision points are set. For e.g., what amount and location combination correlates with a higher likelihood of a SAR filing?
Figure 1: A traditional AML rule (left) vs a simplified representation of a machine learning-based model (right)
Automating the improvement process and using many more variables gathered from a wide range of sources is the essence of machine learning applied to this space. It provides a decision-making method capable of identifying and utilising correlations that are impossible for humans to find and use by hand. We often refer to this set of more complex rules as a model.
While hand-crafted rules are not completely outdated – they are still useful to filter out very obviously suspect payments – they should be seen as the first step in a modern AML pipeline. A modern system contains multiple machine learning steps and ends with analysts investigating payments. In a world of increasingly sophisticated financial criminals, it becomes not only a value add to utilise the proper machine learning tools, but simply necessary.
Machine learning is trained on historical data and expects the payments evaluated to be similar to the payments of the past. This is also true for rules. However, machine learning models can be improved by diversifying the data they learn from, and diversification models can be used as part of an AML pipeline for this purpose. For example, knowing how different payment is compared to previously analysed payments, or how much a customer’s behaviour (e.g., frequency of payments) has changed over time can improve model performance.
An important consideration when introducing such systems is to evaluate the AML risk. A rules-based system is binary and will not consider or assign any risk for most payments as only a fraction of the total number of payments is flagged and assessed. This hides information about a large part of the flow. Machine learning-based models assign a risk level to all transactions. Not doing this leaves banks unaware of their flow risk. Crucially, a machine learning model finds suspicious flow that a normal rules-based setup cannot. It can reduce false positive rates by 10x or more, automatically adapt to new scenarios and increase detection of true positives, allowing analysts to focus on relevant payments without being flooded with payments they don’t need to see.
Embracing machine learning opens up new opportunities for collaboration within banking and finance that simply do not exist when using traditional rules. Banking Circle has recently developed a federated learning system. Federated learning is when multiple participants contribute to training a model, see Figure 2, without sharing data.
Figure 2: Banking Circle’s federated learning system
Each participant trains using their own local data and shares the findings with a server which uses them to build an improved global model while retaining the knowledge gained from each participant. Only the machine learning updates from each participant are sent to the server, never data. This enables a number of interesting possibilities. For Banking Circle, the internal use case is sharing knowledge to train a model to analyse US payments without exposing European data.
Here, we can train a much more effective US model while remaining fully compliant with data protection laws in Europe. Having one participant in Europe training on EU data, and one in the US training on US data, both contribute to a single federated model that retains the knowledge of both participants. Looking beyond this, when considering the need for machine learning in modern AML and the privacy preservation that federated learning provides, the opportunity exists for banks to collaborate and train a model more performant than any of them could train alone.
Together with the ability to identify complex flow patterns, utilise more data to discover previously hidden pathologies, and provide an improved understanding of flow risk, it is clear that not using machine learning in AML risks enabling money laundering.
This editorial was initially published in the Financial Crime and Fraud Report 2023 which dives into the captivating world of fraud management, digital onboarding, and financial crime in the financial services industry. You can download your free copy here.
Ruben, PhD, has a background in engineering and computing with a focus on optimisation and control theory.
Robert has a PhD in computer science and a background in working with data and automation in the banking sector.
Christian has a PhD in engineering and worked across network analytics, optimisation and AI/machine learning.
Banking Circle is a fully licensed next-generation Payments Bank, designed to meet the global banking and payments needs of Payments businesses, banks, and online marketplaces. Banking Circle solutions power the payments propositions of 300 financial institutions across the globe, including 16 banks.
Every day we send out a free e-mail with the most important headlines of the last 24 hours.
Subscribe now