As the volume of data available to humanity has increased over the last decade, it has exposed a weakness in the way we as humans think. Humans are terrible at making decisions that rely on taking many variables into consideration.
Scientists have suspected this for some time. As early as 1956, George Miller published a paper in Psychological Review which showed that humans process at most five to nine variables when making decisions. Next time you decide on something, take a minute to write down the variable inputs you considered as a part of that decision-making process. Are you on the lower or higher end of that five to nine range?
Miller received so much attention for this paper that he re-published it later with some additions which showed that, with the aid of “several stimulus dimensions, recoding, and various mnemonic devices,” this range could be increased, but not by much.
Intuitively, we understand that more information should lead to better decision making, and so we rationally want as much data as possible. At the same time, we have all experienced information overload and subsequent analysis paralysis that can lead to stalled or sub-optimal decision making.
The limits of rules-based systems
This is true also in financial crime detection, as the volumes of available data have increased with the digitalization of payments, digital identity and “know your customer” data. There are now terabytes of data available. Surely, this should be enough to build a behavioral profile of customers that allows us to tell who is using their bank account to buy a Starbucks coffee on their way to work and who is a crime kingpin sitting at the center of an international money laundering ring. Yet, consistently, the systems responsible for making determinations produce huge numbers of false positives and fail to adapt to new financial crime typologies as they emerge.
The answer lies in the way these systems look at the data related to the coffee drinker and the money laundering kingpin, and the way these systems make decisions based on that data. For more than a decade, the transaction monitoring systems that sit at the intersection between a banking customer and their bank of choice, the system responsible for detecting whether this is another Starbucks order or part of something more sinister, have been relying on a “rules-based” system.
A rules-based system should be familiar to all of us. It is the same system we humans use to make decisions every day. If X is true and Y is false, then Z must be a money launderer. The problem is that this rules-based system is exposed to the same constraints and logical limitations that humans encounter.
The benefits of AI and machine learning based systems in AML
A much better way of looking at the same data is with a machine learning and artificial intelligence-based system. Why is that?
Because machine learning allows us to engineer many features and to use them to make objectively more informed decisions. It does this by adding a greater number of variables which can be used in making a decision and adding variables of a higher predictive probability. An engineered feature is created by taking the same data that a rules-based system has access to and creating new statistically significant combinations with that same data. These can fall into many different categories, such as aggregated features, network features, change in behavior features and entity profile features.
The feature data is then fed into machine learning models to produce an output that could be a binary “yes” or “no” answer, as is the case with classification algorithms, or the output could be a sliding scale of probabilities, as is the case with regression algorithms. Either way, a decision is made, and while a rules-based system might take five to ten variables into consideration, a machine learning-based system will take hundreds or even thousands of these variables into consideration by adding these engineered features into the pool of available information.
Important questions before shifting to AI/ML
As humans, we are not incorrect in our intuition that more and better information leads to superior decision making, but only if it can be used effectively in the decision-making process. This requires new ways of thinking about the problem, ways that machine learning and AI are better suited to do.
So, if you are finding that you are inundated by false positives and failing to keep up with new types of financial crime typologies, or if you are a government, bank, financial institution or one of their customers, ask yourself these questions:
- How many features are being used as a part of the transaction monitoring system’s decision-making process?
- What are the kinds of engineered features that are the most predictive for our statistical population?
- Is our risk-based approach constrained within the same confines as our own human limits when it comes to decision making, or are we using a machine learning- and AI-led approach to risk?
With these questions in mind, this will help guide you through the best approach for your organization.
