Machines vs. The Mob: Fighting Money Laundering & Terrorist Financing with Machine Learning
Consider for a moment the problems of criminals: The money they earn, spend, transfer or move – as a result of crimes committed or as part of the planning process for crimes as-yet-to-be-committed – has the high potential to raise the suspicion of law enforcement. If criminals wish to live to fight another day, they must conceal the origins of their ill-gotten gains – typically by “laundering” their money.
Money laundering and terrorist financing
Concealing the origins of their illegally obtained or about-to-be-used illegally money can be quite the undertaking: Criminals and would-be terrorists often pass money through a complex sequence of banking transfers or commercial transactions, with the ultimate aim of returning it to themselves in an obscure or indirect way – now “clean,” having been erased of all traces of criminal activity on its trip through the financial system.
For that reason, the financial services industry has long been tasked with implementing anti-money laundering (AML) detection systems. Organizations have a real incentive to detect and quash money-laundering activities: As fines from regulatory organizations increase, so too has the call for holding compliance officers, senior executives, and board members personally liable for failure to have an adequate AML program and transaction monitoring system (TMS) in place.
Since the 2008 financial crisis, fines levied against financial institutions have been huge and are continuing to increase – we’ve seen $26 billion in fines levied since 2008 alone. Despite all this, money laundering is a $3 trillion per year industry, making up 5% of the global GDP!! So while we’re spending a lot in fines, in enforcement, we’re still not getting to those transactions that are illegal.
Anti-money laundering transaction monitoring systems
Transaction monitoring systems are an incredible boon to the anti-money laundering and compliance profession: The sheer number of financial transactions that occur daily across the globe are too vast for even a legion of very focused humans to monitor. It would be near-impossible for even the sharpest of eyes to detect even a fraction of suspicious transactions without the help of automated detection systems.
However, the systems employed today come with a troubling downside: Transaction monitoring systems throw off a huge number of “false positive” results. Studies have shown that 90-95% of the triggers and alerts these systems generate are false positives and the financial institutions are spending a huge amount of money on humans to then go into these transactions and investigate them.
What does that mean? Any alert generated by a TMS must be investigated, assessed, or otherwise reviewed by a human being who can employ context, skill, and subject matter expertise to either resolve a transaction or escalate it.
And human beings are quite expensive. The more false positives automated TMS creates, the more expensive the compliance process becomes.
The plague of false positives in money laundering investigations
There are a number of reasons for this “plague of false positives.” One way is related to how we try to detect suspicious activity:
The traditional approaches are rules-based, employing if/then statements. As a simple example, one system may flag every transaction sitting just below a reporting level – let’s say $10K. So any transaction at or above $10K may need to be reported to financial regulators, so money launderers would try to structure cash that’s going into the financial system just below $10K so it doesn’t trigger that reporting.
However, money launderers aren’t careless in this way anymore – they’re savvier in the way they introduce cash into the system. Regardless, hundreds of transactions at, say $9K or $9.5K will trigger AML systems that require time-consuming, expensive human review – while real money laundering transactions may go undetected.
Another problem is the massive amounts of data involved in the global financial system.
Huge amounts of data sit within financial institutions – scattered throughout disparate data systems – some sitting in customer documents, in archived financial systems, in emails, even some data can be found in news reports and on social media. Many governments have also built databases to identify ultimate beneficial owners, so data found within those systems must be scoured and matched up with financial transaction data. There’s so much data that a human can’t make sense of it all – let alone see the links needed to identify the criminals hiding within the sea of financial data.
Automating anti-money laundering
Enter two types of machine learning: Unsupervised machine learning, and traditional supervised machine learning. Together, coupled with human expertise and some of the older approaches, these techniques have the potential to end the plague of false positives to help the experts uncover – and stop – the real money laundering and terrorist financing transactions.
The difference between unsupervised machine learning (UML) and traditional machine learning is that in the case of UML, it doesn’t require massive amounts of human training before the system becomes “stabilized,” nor does it require the highly organized, labeled data that you need in traditional machine learning.
Unsupervised machine learning
UML can detect links that traditional, trained machine learning and humans just aren’t able to detect. Oftentimes, in money laundering, the source of the wealth – and the ultimate beneficial owner – are hundreds of layers deep – so you are looking to connect very small pieces of information like IP addresses and locations and so on – and UML is able to detect relationship patterns and links very quickly.
Particularly in the wake of Panama Papers and Paradise Papers, the need to understand ultimate beneficial owners has greatly increased – regulators are certainly going after this more. Government databases in particular are difficult to read, but UML can read and make sense of the data.
UML can also detect patterns which go beyond if/then – so it’s looking at patterns that have been built by known money laundering transactions and these patterns are much more intricate than the rules approach would tell us.
Reducing false positives, increasing safety
Many financial institutions must process the high volumes of false positives – choosing not to do so would leave them vulnerable to astronomical fines and worse – potentially letting funding for the next terrorist attack slip through the cracks.
However, through skillful application of a combination of technologies and techniques – unsupervised machine learning, supervised machine learning, traditional rules-based approaches, and expert review – the high volume of false positives can be lowered, thus freeing up time and resources to investigate potentially dangerous transactions.
The latest UML approaches have been designed to specifically target transactions that mirror the way money laundering and terrorist financing is currently happening. Using these newer approaches allows financial organizations to more easily and efficiently meet their regulatory and moral obligations – successfully applying these models to quickly uncover the transactions that need to be investigated, and put a stop to them immediately.