We have been asked to explain the differences between rules-engines and machine learning for anti-money laundering/banking secrecy act applications on numerous occasions. This question typically arises from the compliance team, wanting to better understand the difference between their current process and that used by a modern AI-based AML system.
A machine-learning based system of insight differs from traditional rules-based approaches to AML data analysis. They are apples and oranges approaches to data analysis and while both have their individual strengths and can be used in combination, the consensus from industry leaders and regulators is that machine-learning algorithms are the future for BSA/AML processing because of their speed, accuracy and the significant drop in the alert processing workload and subsequent false positives.
One outcome from deploying modern AML using machine learning is a potential drop in alert processing workload by 50%. In the past 5-years, banks have increased regulatory staffing tenfold to cater for increased regulatory requirements. Overall we believe that a 50% reduction in the cost of AML analysis is a realistic objective using machine learning, due a substantial reduction in compliance workload.
Rules-Engines Pros and Cons
A major problem with the current approach to AML/CFT process is poor outcomes from the massive $80+Bn invested annually in anti-money laundering technology, support systems and compliance personnel. Not only is the process inefficient, but money launderers have learned ways to circumvent rules engines to avoid discovery. This is evidenced in the statistics for illicit funds intercepted, of only around $20Bn or 1% of $2+Trillion laundered through the financial system annually.
Rules engines are not new, they have been around for decades. Simplistically, a rules engine is a bunch of if-then statements. A simple example of rules function is “If X, then do Y, else if A, then do B.” The problems with rules engines compound when there are up to 600 rules and dozens of money laundering scenarios that have evolved over the years. Rules engines are deterministic, "based on the inputs, these transactions qualify as an alert and must be investigated".
Other problems could include;
- One rule out of place can cause high false positive counts.
- The person who wrote the rules, left years ago.
- No-one knows all of the rules and therefore adding new rules is highly problematic.
- Rules-engine performance declines as more rules are added and as data sets become very large.
- Another draw back for rules engines is that someone has to update the rules every few months.
- Finally, redundant rules and obsolete rules tend to live on forever. An example of redundant rule is to remove your shoes at TSA checkpoints. This is a redundant rule that creates make work for passengers and TSA employees.
A positive aspect of rules engines are that the rules are visible and can be examined by regulators and edited by bank staff, without requiring a computer scientist.
Machine Learning Pros and Cons
Machine learning recognizes patterns and trends in historic data, and provides a probabilistic output, based on matching transactions with transaction patterns of known bad-actors. To identify patterns, the machine learning algorithms must be "trained". For effective machine learning, algorithms within the machine learning system have to be trained on historic data about outcomes that have been confirmed as money-laundering behavior and illicit transactions.
The advantages of machine learning are well established and include;
- Faster processing speed.
- Reduced anomalous alerts and reduced false positives.
- Insights can be developed across the bank's entire data set.
- Fewer true positives slip through the cracks.
- Machine learning systems can evaluate tens of thousands of virtual attributes to develop new insights, unattainable to rules engines.
Disadvantages are related to the lack of visibility into the system. It does not have rules that can be inspected by compliance personnel or regulators. Instead, the machine learning results are compared with data sets produced by legacy rules-based systems to validate that the models are accurate. AML/BSA regulators are supportive of this approach to vet machine learning accuracy.
In summary, AML systems that use machine learning are the future of anti-money laundering systems and early adopters are proving for themselves that the advantages discussed above outweigh the risk of adopting a new technology.