Machine Learning: A Game-Changer in the Fight Against Money Laundering

Financial crime, including money laundering, is an illicit growing industry, and criminals are getting increasingly sophisticated. Financial institutions (FIs) are deeply involved in combating this. In North America alone, they spent a little less than $50 billion on compliance in 2021.¹ But banks typically use rule- and scenario-based tools, derived from industry red flags and expert judgment, for transaction monitoring—and this always seems to put them a step behind the bad guys.

Now, there is an opportunity to get out in front. Recent enhancements in machine learning are helping banks to improve their anti-money laundering (AML) programs significantly. Regulators are supporting these efforts² and working with banks to test new approaches.

In theory, banks can apply machine learning against all kinds of money laundering. But we believe that transaction monitoring—specifically, combining machine learning with other advanced algorithms—is where they can reap the most immediate and significant benefits. Machine-learning models can apply detailed, behavior-indicative data to build sophisticated algorithms; moreover, they are much more adaptable than rules and scenarios because they can quickly adjust to new trends and continually improve over time. By replacing rule- and scenario-based tools with machine learning models, one leading FI improved suspicious activity identification by up to 40% and efficiency by up to 30%.

Transitioning to Machine Learning

In making this transition, banks first need to understand when and how machine learning can be used—and when it cannot. Machine learning works well when there is a high degree of freedom in choosing data attributes as well as sufficient availability of quality data (for example, in scenarios where there is a rapid movement of funds and a large number of attributes can be considered). Machine learning is also appropriate when it becomes difficult to identify the dynamics and relationships between risk factors.

However, it is not useful when there is not enough existing data to build forward-looking intelligence. Bad data inevitably leads to poor model performance. In these cases, a traditional approach (rule- and scenario-based tools, for instance) could be more effective.

Some institutions are exploring how to improve their data, looking at modeling against individual transactions or cases or client relationships terminated for AML reasons and data from historical subpoenas and other law enforcement (LE) sources. In effect, as machine learning develops, it may allow banks to move away from a compliance mindset and toward risk management³ —a shift that could not only enable them to follow the law but also improve their overall competitiveness.⁴

FIs will have to choose their own unique path to bring machine learning to transaction monitoring. That said, three best practices will apply widely.⁵

Get everyone on board. When machine learning projects fail, it is often because of a lack of buy-in from various stakeholders, including the data, technology, line-of-business, model risk management and compliance teams. It is critical to engage everyone from the outset to create a common vision and to make architectural design choices that work for all processes. This helps to ensure that the business can continue as usual and that ongoing regulatory actions are considered. Gathering multiple perspectives improves transparency and can help to uncover and reduce risks. It may be wise to meet with regulators well before machine learning development even begins—and then throughout the development process—to avoid surprises.

Develop a transition plan. Technology transformations are complicated: Employees often resist, and new technologies can introduce unforeseen risks. To deal with these issues, consider running existing rule- and scenario-based processes in parallel with machine learning-based scenarios to build confidence. Banks can choose projects that can use platforms employees are already comfortable with and integrate new components one at a time, starting with those that offer significant potential rewards with manageable risks.

Empower model risk management teams. To incorporate machine learning solutions into the transaction monitoring framework, model risk management teams need to expand their capabilities to work closely with the data science team in the model development and validation process.⁶ Ideally, model risk management teams should have the expertise to educate data scientists about potential risks; define precise performance and monitoring requirements; and address the specific risks associated with machine learning models during validation. Addressing these risks will require policy decisions on what to include in a model inventory,⁷ as well as determining risk appetite, risk tiering, roles and responsibilities, and model life-cycle controls. But many banks will not have to reinvent the wheel; existing frameworks can be adapted to this purpose.

Conclusion

Financial crime is a global industry, and the emergence of new technologies and digital currencies means it is ever more complicated. Fortunately, the degree of global collaboration⁸ among regulators, LE and FIs is unprecedented. And the advent of machine learning models could be a game-changer. This could be an inflection point in the fight against financial crime, enabling FIs to take more effective action and to spend less time on low-reward efforts.

In short, machine learning gives banks a chance to catch up with the bad guys. They can improve transaction monitoring dramatically by reducing false-negative and false-positive rates—and by sending higher-quality alerts to downstream AML investigators. This will likely require investing significant time and resources. But given the stakes, it is well worth the effort.

PK Doppalapudi, associate, McKinsey & Company, Charlotte, NC, pk_doppalapudi@mckinsey.com,