With a million-wide customer base, banks have introduced automated transaction monitoring systems. Traditionally, they are based on a limited number of predefined money laundering detection scenarios, known as rules or conditions. If an account holder violates any of these rules, an alert is generated.
This is very imperfect. Firstly, the rule-based systems are usually based on patterns observed in historical cases. Thus, many banks can only detect money laundering cases that were similar to these observed in the past.
Secondly, most of the existing transaction monitoring systems are producing an unwieldy number of alerts that require to be reviewed manually by AML analysts. Apparently, around 95 percent of alarms raised by current systems are false alarms, as reported by Reuters. However, our experience shows that in some banks even more than 99 percent of alerts are irrelevant. What does it mean in practice? It means that AML analysts spend most of their efforts on alerts that should never have been created.
What can we do about it?
We believe that artificial intelligence (AI) is the future of AML systems. We have proved that machine learning algorithms can be used to boost the effectiveness of the overall AML process. The figure below illustrates the AI-enhanced AML process, where you can see alerts ranked by importance:
For each alerted case or client, we estimate the risk of money laundering with the use of the most powerful classification algorithms. They include (but are not limited to) random forests, gradient-boosted methods (XGBoost, LightGBM, CatGBM) and deep neural networks.
Based on ranked alerts, we are able to prioritise. Below a certain risk threshold, we can cut off or hibernate alerts, as their risks are negligible. It basically means less work to be done – 60 percent less in one of the cases we worked on with a major bank.
In search of the ‘why’
As more and more machine learning algorithms were being proposed, AI practitioners observed that there are significant obstacles to use some of them in the real-world applications. More advanced algorithms with larger accuracy are usually more complex and loose on transparency. When using ensemble methods or neural networks, known to have large predictive power, we cannot clarify the reasoning behind each particular prediction. The results are more difficult to interpret – it becomes challenging to explain why an AI system claimed one case to be very suspicious and the other – not to be suspicious at all.
Banking institutions are obliged to detect and report on customers engaged in money laundering and terrorist financing. But not only that. Regulators demand transparent and detailed documentation. It has to include, among others, a clear reason every time the compliance department closes a case and marks it as not suspicious (it does so in around 95% of cases).
Understand, not just analyse
That is why we turned to a different approach, taking advantage of novel techniques known collectively as explainable artificial intelligence (XAI). How do these techniques work and how are they different from traditional AI-based systems?
Traditionally, based on given data, a machine learning model addressing a specific task is built. It is evaluated and optimised. In many cases, one would end up with a black box model that solves the problem very well but even the model creators can’t really inspect how the model is accomplishing what it is accomplishing. On the contrary, our approach is to consider the so-called glass box models. In such kind of models we not only use powerful AI algorithms, but also methods of XAI. Using such approach, we understand why the system derived a particular prediction, why it erred, and we know how to improve it.
No, money laundering will not disappear overnight. But yes, with time, AI can come to much rescue.
One of the easiest XAI methods is called permutation feature importance. Permutation of a sequence is a way in which its elements are rearranged. In short, the method uses different permutations of feature values to estimate its importance.
Say you want to predict how many bikes will be rented on day X. Permutation feature importances can tell you that the temperature is what matters most here, while what matters least is whether the day is a business day or a holiday.
Get more information about financial services software.