Machine Learning can bolster our defences against financial crime

As winter approaches, so does the season of coughs, colds and ‘flu. Many of us will take the opportunity of having a ‘flu vaccine to ward off this seasonal lurgy. Unfortunately, ‘flu vaccines are not 100% effective because the ‘flu virus mutates rapidly, making it difficult for medical experts and scientists to predict precisely which strain will dominate from year to year. However, this year, breakthroughs in both the US and Australia using artificial intelligence and machine learning promise solutions to the annual battle against ‘flu.

In a similar manner to biological viruses such as the ‘flu, the techniques used by criminals to launder money and finance terrorism mutate and evolve, driven by increasingly sophisticated uses of technology and the ongoing digitisation of finance. Recently, criminals have been found to use on-line games as a mechanism for laundering money. There is a rise in the use of ‘money mules’ which are notoriously hard to detect and new methods such as ‘cuckoo smurfing continue to arise.

The first wave of technological solutions to the problem of money laundering and transaction monitoring in particular employed rules-based logic to identify suspicious transactions. Creating these rules effectively depended on understanding the known patterns of money laundering – the types of payment patterns that may indicate smurfing, for example, or transactions of a size that is unusual for a particular customer. 

Of course, rules-based systems are a huge improvement on manual transaction monitoring and have been very successful in improving the detection of illicit financial flows but their effectiveness is limited by three key factors:

  • their reliance on known patterns of behaviour
  • the need to set quantitative thresholds for the rules to operate effectively
  • The requirement for ongoing recalibration and rule adaption when new money laundering typologies have emerged

However, as criminals become ever more sophisticated in their use of technology and devise multiple and various ways of hiding their proceeds of crime, transaction monitoring products based on rules alone are no longer proving to be an adequate defence against money-laundering. Figures from Europol suggest that whilst levels of reporting are increasing year on year, only around 10% of SARs lead to further investigation and only about 1% of criminal proceeds in the EU are ultimately confiscated.

Over the last few years, the RegTech market has seen a huge growth in products designed to fight financial crime that are making use of other, more innovative types of technology to improve levels of detection accuracy and efficiency in transaction monitoring.

Transaction Monitoring products in the RegTech Directory

Machine learning (ML) in particular has the capability to be more adaptive than rules-based systems and able to spot anomalous patterns in data that indicate something suspicious is going on, in real-time. Indeed, a recent joint report from the FCA and the Bank of England has highlighted that anti-money laundering is a key use case for ML in financial services, and one where these firms see real benefits.

ML is a type of artificial intelligence which uses advanced statistical models to parse huge data sets to identify patterns and make predictions. In the case of transaction monitoring, ML systems will use models to flag patterns of behaviour in transaction data that appear suspicious and do this in real-time. Broadly, there are two types of ML models – supervised and unsupervised. Supervised ML models are ‘trained’ on large sets of historical data so they can recognise known patterns of behaviour that are likely to be suspicious, based on what has gone before. Unsupervised ML models are not trained – instead, they can identify patterns of behaviour without reference to existing typologies and are thus used to detect anomalies in transaction data which are likely to be suspicious.

Generally in financial services, we are extremely good at detecting and understanding risks such as market and credit risk where there are huge sets of historical data that can be analysed statistically to help us predict risk in the future. This is analogous to supervised machine learning – historical data is used to train ML models to detect known outcomes. When it comes to financial crime, we can detect a proportion of suspicious transactions based on known money laundering typologies, but unfortunately, as we have seen above, new and unknown patterns continue to emerge.

For these Rumsfeldian unknown unknowns, we need to think more about financial crime in terms uncertainty and less about risk. This is a distinction made by Knight as far back as 1921 and hinges on the idea that we are dealing with risk when, even if we do not know the outcome, we can measure the probabilities of different outcomes occurring. Uncertainty, however, means we do not have all (or any) of the information in order to set these probabilities in the first place. And this sounds very much like unsupervised machine learning – which can detect new and emerging financial crime typologies that have not before been encountered.

On this basis, we would expect that transaction monitoring systems based on unsupervised ML would be the most desirable solutions for financial institutions. However, there is an important trade off that has to be considered. Regulatory expectations (and good practice) around ML models require that the application of these models is transparent, and that the outcomes and decisions reached by ML systems can be explained. Unfortunately, levels of explainability decrease significantly when unsupervised ML models are used and firms must weigh up this balance between accuracy and explainability when implementing ™ systems using ML.

As far as trends in the RegTech market for transaction monitoring tools go we have observed the following:

  1. Optimisation of existing rule-based systems through the use of data analytics to improve their tuning and calibration
  2. Layering of products making use of ML techniques on top of incumbent rules-based systems to improve detection accuracy and reduce false positives
  3. Replacement of incumbent rules-based systems with products based on a combination of rules and ML models – both supervised and unsupervised
  4. Vendors with products that use ML models are developing repositories of crime typologies and seeking ways of sharing these across the industry

Hopefully, these advances in the use of ML and sharing of data will continue to be adopted throughout the industry and improve the level of detection, providing a more robust and adaptive solution to the virus that is financial crime.

For more information on our market assessment of the transaction monitoring tools in our RegTech Directory, click here.