One of the major European banks has recently admitted that it has failed to spot money laundering in its books for years. As a result, the bank has agreed to pay $900 million in a settlement with prosecutors. While it was a man (the CFO) who stepped down for this negligence, would a machine (Machine Learning) be a better guardian to prevent such situations in the future?
I’ve been working with Machine Learning for a couple of years now. Moreover, I’ve had hands-on experience with the anti-money laundering processes in one of the biggest banks in the world. I’m a huge fan and enthusiast of opportunities enabled by Machine Learning in banking or any other industries.
So, would a machine be a better guardian? Knowing my experience, it would be safe to assume the answer is “yes”. But to be honest, the reality is more complicated.
To get a sense of the challenge, let’s start with some numbers. The exact scale of money laundering, by its nature, is hard to estimate - in 1998 the International Monetary Fund evaluated that the aggregate size of money laundering in the world could be somewhere between two and five per cent of the world’s gross domestic product (GDP).
By the way, the world’s GDP is measured in trillions of dollars. Preventing at least part of these fraudulent transactions is a big deal - both from the macroeconomic and the local financial systems’ point of view.
Finding suspicious transactions in the modern banking systems is not an easy task.
Banks are obliged by law to follow strict procedures concerning anti-money laundering compliance. But this is not as easy as it may seem when we realize how many operations banks have to deal with - as per World Payment Report 2017, the estimated number of non-cash transactions in 2017 exceeded 0.5 trillion. Finding suspicious transactions may be easily compared to finding a needle in a haystack. But what if we could apply some modern Machine Learning methods to finding that needle? In the end, they are trained to look for patterns which the human mind cannot embrace and compute.
Interestingly enough, Machine Learning solutions have one major thing in common with preventing money laundering - they both rely on data, and if you don’t provide consistent and quality data, but input trash, you can not expect to get anything but trash in return. This is why fighting big crime starts with small procedures - such as KYC (Know Your Customer).
In terms of data, there is another aspect that can often stop traditional analytical solutions from getting great results - they have problems with handling unstructured data. This is where many of the modern Machine Learning methods show their power. They are able to identify complicated patterns in loosely structured data, and today they are being incorporated in more and more domains - from language translation to (currently under development) autonomous cars.
And the demand for Machine Learning in banking is growing. For instance, Danske Bank decided to apply some new analytic techniques, including artificial intelligence, to better identify sophisticated types of fraud. The problem was nagging - according to a case study, Danske Bank was dealing with a fraud detection rate in the low forties, and up to 1,200 false positives per day.
Another field where Machine Learning techniques come in handy is credit scoring and building credit risk analyses.
However, there are two things that we have to bear in mind when thinking about incorporating advanced Machine Learning solutions to avoid failure.
Danske Bank applied AI to better identify sophisticated types of fraud. However, it also failed to spot money laundering in one of its branches.
Top challenges ahead of Machine Learning - compliance and machine learning bias
The first challenge is the lack of transparency. What do I mean by that? Many modern solutions, such as deep neural nets, rely on automatically extracted features combined together in a complex way. In turn, it may be not trivial to justify their predictions.
Why does this fact matter to banking executives? Remember, this is a world ruled by rules, well-defined processes, and regulations. For now, it seems that regulators put more trust in simpler, yet easier to understand rules-based systems.
It is comprehensible. It is not easy to regulate the use of Machine Learning in banking procedures when you are not sure how the system is calculating its predictions.
It is also worth adding that counteracting money laundering requires proof of its occurrence - gathering evidence and carrying out reasoning, and therefore the final decisions still are made by humans.
Much of the current research focuses on finding a way to get more insights into how ML algorithms make their final decisions, and we are hearing about more and more ways to understand their reasoning. Recently, an MIT research team presented a model which can perform human-like reasoning and justify its decisions, so the perspectives are bright.
But should we only restrict ourselves to a defined set of rules when money launderers know the stakes are high and adapt their schemes quickly to avoid detection? It will be interesting to follow how things unfold.
The second aspect that limits the usage of these techniques is their true ability to generalise, and the bias that the learning dataset brings to their decisions. This flaw has been clearly shown in another MIT study.
In short, researchers proved that the imbalance of the training data in terms of gender and race had a dramatic effect on final performance - systems were particularly bad at identifying women of colour, being 34% less accurate at recognizing darker-skinned females compared to lighter-skinned males. This problem is definitely one of the challenges modern researchers are trying to solve.
In fact, one of the current competitions at Kaggle (a platform called a ‘home for Data Science’) challenges researchers to face this kind of problem - the training data comes from different geographical regions than the one in which the final solution will be evaluated.
Coming back to the starting question - can Machine Learning help banks spot money laundering more effectively? Definitely yes, but Machine Learning alone is not a silver bullet and a remedy for everything. Banks have to ensure their fundamental procedures are not being neglected. And it is wise to remember that this technology can be a double-edged sword - using it without the proper knowledge and consideration can bring more harm than good.
Machine Learning techniques can’t replace humans, but well-designed systems used as support can significantly improve the efficiency of detecting money laundering.
If you're interested in Machine Learning, subscribe to MLguru: a Netguru's bi-weekly selection of inspiring content about Machine Learning and the latest news from the data science world.