The biggest challenge in AML is presented by the number of alerts generated by transaction monitoring and fuzzy logic screening, in particular, by the high rate of false positives ratios, which often exceed 90%.
At the same time, regulators increasingly insist both on ensuring full coverage (many installed systems only cover a part of a financial institution’s activities) with all resulting alerts investigated.
Together with the high number of alerts and the high rate of false compliance departments are drowning under investigation tasks and are in dire need of ever more staff.
However, throwing additional staff becomes not only prohibitive on the cost side, but creates other problems as well:
High staff turnover: fastidious, repetitive and seemingly uninteresting work, under time and performance pressure quickly leads to loss of motivation and short tenure;
Error risks: inevitably, true positives or interesting alerts which should be promoted to case investigation are missed and mistakenly filed as false positive;
Flooding FIU’s with SARs: in response to stricter regulatory requirements financial institutions adopt a take-no-risk approach; however, this only shifts the burden to the FIUs; some have already flagged the problem and now push for better quality SARs;
Reduced budget allocation for investment in AML systems: bloated operational costs (OPEX) consume the major part of the budget, depleting resources for capital expenditure.
The problem is further aggravated by the lack of qualified investigators in the job market. It is hard to find good candidates for such positions, which leads to staff poaching. Even though wages rise many positions cannot be filled. In the Netherlands alone, it is rumoured that more than 1000 positions are permanently open.
The obvious challenge for the industry is to enhance the quality of the detection. This is not an easy problem to solve, as fraudsters and money launderers keep changing their modii operandi, constantly probing detection thresholds to be able to fly under the radar.
We have looked at all options and concluded that it is best to focus on the alert triage. This approach complements, not replaces the fine-tuning of detection. The results are highly promising. With the help of machine learning, alerts are scored to reflect their propensity of being “interesting”, i.e. the likelihood of identifying real money-laundering behaviour. Overall, we have come to the conclusion that 40% to 50% of the alerts could be classified as low-risk and should then be treated accordingly. This applies not only to transaction monitoring but to fuzzy match alerts as well.
Appropriate alert risk scoring considerably reduces the effort needed for investigations, which translates into huge savings on operating costs and the possibility of reallocating resources to more in-depth investigations that could elevate the quality of reporting to an entirely different level.
The caveat is that building such models and moving from laboratory experiments to industrialising the process requires not only advanced data science knowledge but simultaneously broad and deep AML domain expertise. Without the proper background in both of these fields, the desired results cannot be achieved or transferred into real operations.
This blog is intended to offer insight into the nitty-gritty of a working solution by Complidata. Filip is a co-founder of Complidata, a consultancy focused on the optimization and automation of AML processes through AI, and Compliserve, the only systems integrator focusing 100% on AML.