By Dr Scott Zoldi, Chief Analytics Officer, FICO

Although anti-money laundering (AML) teams are doubling down on compliance efforts, money laundering levels continue to rise year on year. According to the United Nations Office on Drugs and Crime, an estimated 2-5% of global domestic gross product is lost to money laundering every year.

While multiple factors play a role in this growing challenge, the difficulty of finding the 95%+ of money laundering missed today is persistently high false positive rates and inefficient processes.

Current AML systems are overwhelmingly scenario-based and having to grow in complexity to keep up with increasingly demanding regulations. As a result, hundreds of rules now drive every Know Your Customer (KYC) and Suspicious Activity Report (SAR) filing.

Shortcomings of rule-based systems

• Linked to high false positive rates

More rules beget more cases that are flagged for investigation, which are leading to higher false positive rates. At the same time, sophisticated criminals are learning how to work around the transaction monitoring rules and avoid known suspicious patterns of behaviour, which is making it harder to pin down true instances of money laundering. This places incredible pressure on financial institutions, which have to deal with a burgeoning workload without a parallel growth in their compliance staff numbers. They simply can’t keep up.

With the implementation of the 3rd EU ML directive, banks have applied a Risk-Based Approach (RBA), which ensures riskier customers get more due diligence. Although an increased focus on higher-risk customers has helped lower false positive rates, false positives are still too high. Furthermore, a risk-based approach allows one to prioritise riskier consumers but does not address the inherent money laundering missed by rule-based scenarios.

• Detection is limited to known behaviours

Detection today only produces alerts on known scenarios. However, in the fast-changing world of financial crime, where criminal methods are constantly evolving, this just isn’t good enough. Criminals will simply learn and avoid the scenarios designed to catch them.

Under these circumstances, adaptability, accuracy and efficiency are of the essence – all of which are ingredients new machine learning techniques can bring to the table.

Machine learning is successfully reducing money laundering

According to McKinsey, machine learning techniques are reducing false positives by 20-30%, in turn reducing investigators’ workloads by 50%. These significant figures have been achieved simply through introducing tighter segmentation.

Examples of such segmentation can include learning that a customer has financial relationships outside of the home country, is a high net worth individual, or is a small business owner. In this way, machine learning challenges the status quo of KYC processes by drawing on real-time behavioural analytics based on financial transaction activity.

Unlike rules, machine learning algorithms can detect new unseen patterns that reveal illicit activity.  Machine learning also excels at taking many different behavioural features and combining them optimally to maximise detection and minimise false positives. These different behaviour features are the nets cast broadly across customer behaviour and the machine learning combines them in ways that make it very difficult for criminals to learn.

Here’s how the main machine learning for AML methods work to effectively reduce false positives.

• Customer behavioural soft-clustering and monitoring

Traditional AML solutions resort to hard segmentations of customers based on the KYC data (often inaccurate or aged) or special sequences of historical events. However, customers are too complex to be assigned to such hard-and-fast segments and need to be monitored continuously for anomalous behaviour through soft segmentation.

The best approach is to aggregate customers’ banking transactions and generate archetypes of customer behaviour. Each customer is a mixture of these archetypes, which are adjusted in realtime based on their financial and non-financial activity.  Looked at in an “archetype space”, good customers are similar in behaviours, giving the appearance of customer cluster grouping.

Machine learning models can detect and rank-order with scores those that fall in areas of low customer density (where very few customers behave like the customer in question) or areas of extremely high similarity in behaviour (for example, mules being told how to circumvent transaction monitoring).

• Rank-ordering AML alerts

Machine learning can be used to bring the most urgent alerts to a human officer’s attention, which helps reduce false positives. For example, the FICO AML Threat Score prioritises investigation queues for SARs, to ensure the most pressing cases are looked at first.

Moreover, machine learning can generate its own alerts for money laundering instances that are not detected by the transaction monitoring system. Combining these types of machine learning will help in reducing false positives of the transaction monitoring system, while enabling the efficient identification of missing money laundering cases plaguing most financial institutions.

Remaining compliant with Explainable AI 

Recent regulations such as the General Data Protection Regulation (GDPR) are mandating explainable decisions when using analytic models. To remain compliant, business users need to provide clear reasoning behind each AML decision; to explain why certain cases were flagged or missed, for example.

It is an important criterion to keep in mind when choosing a machine learning-driven AML technique. For compliance reasons, only consider those that have explainable AI built into them; better yet, choose systems that are explainable from their architecture design.

The greatest value machine learning can add to AML processes is the ability to continuously learn and adapt. As transaction volumes and datasets continue to surge, machine learning’s “superhuman” boost will be crucial in identifying the needles in the haystack — consistently, quickly, and accurately.