Using Machine Learning to Reduce AML False Positives
07/19/2018 by Eric Hale Financial Crimes
Are your AML investigations flooded with false positives? Do you have an endless backlog of alerts, new risks to identify and monitor, and new products and services that your business partners want to introduce? Are you worried that your team can barely manage the continually evolving regulatory demands? Have you heard the buzz about how artificial intelligence (AI) and machine learning can help, but you have no idea where to start? If your answer to any of these questions is yes, then Zencos can help. Getting started is not as overwhelming as you may think.
The AML community has been dealing with the problem of receiving alerts that turn out to be false positives since the introduction of the Bank Secrecy Act (BSA) in the 1970’s. In recent years, there has been increasing pressure from regulators to cover more risk areas, enhance governance of the scenarios covering each risk area, and to identify and monitor emerging risks that come from growth and expansion of products and services. AML investigation teams are now being pushed to their limits. The historical response to solve this problem has been to hire more investigators, driving compliance costs up, and directly impacting the bottom line.
For some time now, larger banks have already been utilizing machine learning techniques for credit scoring and fraud detection. In the last ten years, these same organizations have recognized the potential benefits of applying machine learning techniques to the AML false positive challenge. The results have been promising, allowing some banks to reduce alert volumes by as much as 60%, depending on the channel and monitoring focus. A common approach to evolve from this effort is the process of “hibernating” alerts that are predicted to have a low probability of being productive while continuing to monitor the alerted activity until it hits a defined threshold that signals for an active review.
Adding machine learning to an AML program represents a significant shift in how alerts are triaged and investigated. Incorporating machine learning for alert prioritization does not happen overnight. BSA leadership needs time to understand the impact on coverage and investigations; internal audit needs time to understand the models used, and analysts need time to interpret and incorporate the model results into the investigative process.
To get started, Zencos recommends an iterative approach that begins with conducting a pilot program that targets a reduction in false positives beginning with one customer segment and one monitoring focus. An excellent place to start is with a smaller customer segment and a monitoring focus area with moderate-to-high false positive rates. Once models are built and validated to address the pilot area, the results must be integrated into the general investigation process. One way to do this is to incorporate the alert score and associated business rules into the transaction monitoring process to automatically ‘hibernate’ alerts below a certain threshold. In situations where your AML system or process are not yet ready to implement this approach, Zencos recommends creating and using an AML Customer Dashboard to augment your investigative process. This approach allows you to supplement and challenge your current AML Monitoring system, and avoid an untimely or unnecessary “rip and replace” scenario.
An AML Customer Dashboard will provide your investigations team with a unique score for each customer, and a recommendation on customer investigation productivity. To facilitate the most effective understanding of this score, the dashboard couples the score with a scorecard that shows the number of points each risk attribute contributed to the overall rating. The scorecard snapshot of critical insights provides guidance that makes the investigative process more efficient. In this way, the customer dashboard can serve as a one-stop investigative tool to visually explore a customer’s transactions, prior alert activity, and other relevant risk events.
Starting with a pilot allows your organization to experientially learn how machine learning can benefit your AML program while working with data specific to your organization. During the pilot, the AML Customer Dashboard can be rolled out to a select group of investigators who will focus on the customer segment and triage alerts targeted by the pilot program. Conducting the pilot will enable your team to identify potential efficiencies and real insights. The results can be used to build the business case in support of expanding the pilot efforts to cover the full transaction monitoring effort.
By utilizing machine learning techniques, organizations can be proactive in identifying potential gaps in existing transaction monitoring systems and quickly address these gaps before regulators scrutinize them. Quite often, these gaps can be treated using reports or by developing new transaction monitoring rules or scenarios.
Check out our recent webinar AML Transaction Monitoring: The True Cost of False Positives featuring Eric Hale, VP of Analytical Solutions at Zencos, and SAS colleagues Carl Suplee and David Stewart.
If you would like to learn more details about the Zencos AML machine learning and Customer Dashboard methodology or to arrange a time to speak with the author, Eric Hale, CAMS please contact us!