Money laundering has always been an ever changing beast. As financial criminals find more ingenious ways to obscure the money trail, traditional AML systems at institutions have found themselves under greater strain than ever before.
This past decade has witnessed a greater number of financial crime exposés than every before. A huge reason for this is the growing amount of data in digital form and the reducing relevance of paper money as a means of transaction.
You would think that increased digitalisation of financial systems would have simplified AML but this apparently has not been the case
More Data, More Problems
Traditional AML models have always face jurisdictional challenges even between highly regulated systems. The laws governing financial crime in New York are significantly different from the ones in Singapore or Hong Kong. Quite simply, a crime in one jurisdiction is often a gray area in another.
Another major challenge is, quite simply, the size of most financial institutions. While size is a good thing when it comes to convincing a customer to deposit their money, it naturally brings about a lack of flexibility when it comes to processes and decision making.
A lot of innovation gets lost in the corporate chain and financial institutions are no exception. While there has been progress made in this area, via physically separating key units into their own, isolated environments, the fact remains that a behemoth is often no match to a small, nimble operator when it comes to money laundering.
This puts institutions firmly on the back foot when it comes to dealing with the ever changing nature of crime. Take for instance how the concept of a money mule has evolved. What was once a physical operation, that is strapping wads of cash onto your body and walking through airport security, has now turned digital with European students selling their bank accounts for a fee in order to hide financial transactions.
Regulators face the same challenges as institutions when it comes to adopting new processes to combat such crime.
Machine Learning as a Solution
Machine learning offers excellent solutions to these problems by not just learning the existing patterns of crime but by also predicting the possible ways future crime might occur.
Existing AML processes in institutions have hand coded typologies (money laundering patterns) and the sheer amount of data these institutions receive, a function of their size, make it almost impossible to investigate each transaction thoroughly.
What most institutions do is prioritize the transactions by amount and investigate accordingly. This is an approach that is aimed at reducing the impact of any fines that might be levied on them and is a bit like dousing fires in just the important parts of a house while acknowledging that the rest of the house will burn.
So how does AI mitigate these risks and deliver better processes? Well the solution starts with hypothesis generation.
Hypothesis generation is a scientific way of saying "where do we start?" Given the volume of data received this is a huge challenge. Intelligent systems have the ability to analyze mountains of data and develop clear and concise starting points.
Generating a hypothesis is about exploring the relationship patterns within the data and classifying the ones which are weak versus strong ones. Remember all of this is probabilistic, in other words, at this point we don't know for sure whether such relationships even exist.
Therefore it is essential to correct course constantly and keep incorporating feedback as we progress along our analysis. A human doing all of this will take forever and will naturally result in a false positive.
Indeed, it is estimated that 95% of AML investigations conducted by institutions end with a false positive. This clearly speaks to an inability to detect relationship patterns between transactions.
Poor segmentation of data is one of the main reasons for this sort of scenario taking place. Segmentation refers to the grouping or classification of customer and transaction level data. By grouping like data sets together, analysis becomes a whole lot easier.
However, if your grouping is off to a bad start, it compromises every single step downstream. The reason for poor segmentation is often
Data volume in excess of handling capability
Lack of sophisticated data learning techniques
To be fair, this field is a challenge even for existing AI systems and it is here we will see the largest leaps and disruptions taking place. Intelligent segmentation remains the largest selling point for most companies in this space.
Providing transparency and the ability to refine segmentation choices are a few of the key challenges facing firms. AI already suffers from the impression of being a black box with most institutions and demonstrating transparency will go a long way in reducing those fears.
The key to intelligent segmentation is for the algorithm to evolve and incorporate newly arriving data into existing segments and predict emergent behavior. While this sounds simple enough, most AI models out there are actually static. Integrating new data remains a major challenge.
Current institutional procedure is to group activities that seem similar into the usual L1, L2 or L3 categories and go from there. However, as we've seen, if prior segmentation is incorrect, there's not much point in grouping data further downstream.
This is why most L3 alerts end up having the highest monetary amount attached to them and is an admission of an imperfect system. Grouping alerts begins with analyzing the transaction level data and classifying them in order of priority.
Groups of transactions with similar alert levels are then created and their patterns analyzed. This is currently a major challenge for most institutions since the sheer volume of data guarantees an inefficient grouping and requires simplifying the rules surrounding the alert levels.
Hence we end up with the transaction amount being the determining factor of the alert level. While this is simplifying things a lot, the end result ends up following this pattern of thought.
Opportunities and Challenges
Intelligent learning systems with the capability of defining and handling topologies are the solution to all these problems. There remains a lot of room for improvement in this space.
Companies such as Ayasdi provide machine learning solutions to institutions and initial results look promising. By implementing their systems, an institution was able to reduce their investigative volume by over 2o%.
Reducing this volume has a significant impact by reducing the strain and addressing the issue of false positives we saw previously. Ayasdi touts its intelligent topological engines and its ability to handle segmentation as its key selling point.
It remains to be seen though, how such companies address the evolutionary aspect of crime. What that means is, how well do their engines react to change. While its easy to predict what might happen, the proof is in the pudding and we simply won't know until the event actually comes to pass.
The future looks bright, contrary to perceptions, when it comes to AML handling systems in institutions and regulatory bodies. With AI constantly learning and evolving, banks and regulators will face ever decreasing case loads and reduce strain.
The key point to remember is we're still at the beginning of the AI revolution. In absolute terms we're only at the tip of the iceberg when it comes to determining what machine learning can achieve for us.
AML and combating financial crime is no different and this space remains an exciting area to watch for developments.