Close

Are you in compliance?

Don't miss out! Sign up today for our weekly newsletters and stay abreast of important GRC-related information and news.

×

Status message

This is subscriber-only content, you are viewing with temporary unrestricted access. For full access, being your free, no obligation 5-day trial.

Primary tabs

On the front lines with machine learning and AML battles

Joe Mont | January 30, 2018

Late last year, Oversea-Chinese Banking Corporation Limited, best known as OCBC Bank, became the first bank in Singapore to implement artificial intelligence and machine leaning in its efforts to root out suspicious activities, specifically money laundering and other financial crimes.

To do so, it turned to ThetaRay, a global cyber-security and big data analytics company headquartered in Israel.

In turning to the promise of cutting-edge technologies, the bank confronted where existing AML efforts fell short.

On any one typical workday, an OCBC Bank anti-money laundering compliance analyst would log into the bank’s transaction monitoring system and find up to hundreds of potentially suspicious transactions they had to review. These transactions are flagged for having fulfilled one of several “rules”, such as a sudden large transfer into or out of an account, that the system had been programmed to recognize, the bank says.

This rule-based approach, however, meant that, with the multiple alerts, AML analysts had to manually trawl a slew of false positives and dead ends as they went on the hunt for viable evidence of a financial crime. It was a time-consuming process that could take days, even weeks, for complex transactions.

“Multiply this process across the many days in a month and year for the bank’s transaction monitoring team, and the demands of AML monitoring become apparent,” the bank explained in a press release. “This also does not account for the potentially suspicious transactions that the system did not detect because they were not captured under the pre-defined list of rules.”

Having reached a breaking point, to tackle the increasing scale and complexity of AML monitoring, OCBC Bank became the first in its nation to turn to artificial intelligence and machine learning. The goal, and the promise of the technology, was increasing the bank’s operational efficiency and accuracy in the detection of suspicious transactions.

OCBC Bank’s transaction monitoring team and its FinTech unit, The Open Vault at OCBC,   conducted a proof of concept with ThetaRay, which concluded earlier this year.

The bank is now in a pre-implementation phase, with advanced testing of additional test data. “This allows the bank to further verify the efficacy, security and robustness of the solution while gaining a more comprehensive understanding of its workings and capabilities,” said in a statement.

Upon the successful conclusion of this additional testing, the bank targets to fully implement the technology, which will run in parallel with its existing transaction monitoring software.

“Financial crimes are evolving in complexity and sophistication,” says Loretta Yuen, OCBC Bank’s head of group legal and regulatory compliance. “Banks play a central role in foiling illegal activity such as money laundering and constantly have to be one step ahead of financial criminals. This is why we strongly believe in embracing technology and tools that will increase our proficiency in transaction monitoring.”

“Banks play a central role in foiling illegal activity such as money laundering and constantly have to be one step ahead of financial criminals. This is why we strongly believe in embracing technology and tools that will increase our proficiency in transaction monitoring.”

Loretta Yuen, Head of Group Legal and Regulatory Compliance, OCBC

“We are now able to leverage a FinTech solution to sharpen the detection of suspicious transactions,” she added. “There is tremendous potential for the application of artificial intelligence and machine learning to review these flagged suspicious transactions… We hope to be able to fully use technology to make the entire AML process efficient, more accurate and secure.”

Jim Heinzman, executive vice president of financial services solutions at ThetaRay, has more than 25 years of experience in the world of financial institution compliance and the technology underlying those efforts.

His firm’s technology is already proving its value at the bank. It used the software to analyze one year's worth of transaction data. Doing so reduced false positives and alerts not requiring further attention, by 35 percent. The accuracy rate of identifying suspicious transactions increased four-fold sorting flagged activities using a risk-based approach.

ThetaRay’s solution uses algorithms that are not reliant on pre-programmed rules to flag suspicious transactions. Instead, the analysis, and search for anomalies, is holistic, broader, and spans a wider slice of all transaction data.

This sort of contextual data analysis is also able to “learn” from and adjust to changes in transaction patterns. Over time, this on-the-job learning increases effectiveness and accuracy. The number of false positives, a bane of manual-process steeped compliance officers, is reduce the ability to apply and prioritize risk-based categorizations and responses.

There’s a fair share of confusion in the marketplace buzzwords like artificial intelligence and machine learning, Heinzman says.

“It's really about a supervised machine learning approach,” he explains, regarding traditional rule-based “if/then” screening technology used in AML efforts. “What that means is that I have a pattern I know, I've seen it before, and I want to train the system to recognize those things very accurately. That’s a supervised model. What we do is unsupervised. I just want to find the anomalies within the data and where there is a behavior that I've never seen before, or a new pattern. We find unknown unknowns. Once we find it, it is not unknown anymore. I can label that thing and the system will continue to find more of them.”

The potential for ThetaRay’s technology to detect previously unknown patterns of money laundering will, if all goes to plan, deepen OCBC Bank’s understanding of financial crime and “help prevent the bank from running afoul of AML regulations as financial crime becomes more sophisticated,” the bank says. Also, as AML analysts are freed up from performing the most basic aspects of transaction monitoring, they will be able to take on more sophisticated and value-added analyses,” it says.

“The use of technology will also potentially allow OCBC Bank to tap a wider pool of professionals who are qualified to perform its transaction monitoring function, data analysts and scientists for example, thereby easing the crunch in the hiring of AML analysts,” the bank added.

“It's a way to really interrogate the data and to look at all of it,” Heinzman says. “Before, I was the guy writing rules. I was saying, ‘Ok, so if the transaction is this size, and there's a frequency like this, and there's a payee like that, and it goes through this channel, that's probably suspicious. Then the bank has to write the rule for that, but the rules are only as smart as I am.

“There's a lot of smart people writing rules, but what's happening is that the criminals, the money launderers, are becoming more sophisticated,” he added. “They are using the banks’ rules against them. Most of those rules that the banks are using in systems are published on the dark web. So, if I want to do money laundering or commit fraud, and I know what your rules are, I can I can easily devise a scheme to go around them. Bad actors, fraudsters, and cyber-criminals are using AI to attack banks. If I'm using a rules-based system to defend myself, and someone's attacking me with AI, I'm showing up to a gunfight with a knife.”

The use of semi-supervised and unsupervised data purges, carry no such weaknesses because they are not strictly rule-based.

“When I see the output of our system, there are things I never would have thought of and would not have thought to write a rule to find,” Heinzman says.

Heinzman recalled a recent deployment at a bank that was generating roughly 10,000 alerts a month, 99 percent of which were false positives. “They were paying a lot of people to look at things that weren't even issues,” he says. “They had a few events that happened that that they didn't catch. The regulators found them. They had a problem because they had a really expensive system that wasn't catching the things that they needed to catch.” The pattern-based analysis also uncovered problems well ahead of when regulators arrived on scene.

“We uncovered a couple of cases of terrorist financing and human trafficking,” he says. “That made us feel good because we actually could point to real cases where we identified criminals they otherwise couldn't catch.”

“It was clear from how they how they went after the bank that they were able to they had an understanding of the established controls and were able to exploit them,” he added. “We found for example where they were using multiple geographies multiple channels and multiple products. It wasn't just one customer. It was a number of customers that were acting in tandem over time. They actually moved a very large amount of money through the bank a few hundred euros at a time.”

Knowing, for example, that the bank had higher monitoring thresholds on new accounts and on a certain amount, the thieves played it safe for the first several months after opening an account in order to avoid detection. Then, they only did one transaction a month, and always kept it under a million dollars.

“They knew the threshold and they were able to move a tremendous amount of money through this bank,” Heinzman says. “They moved hundreds of millions of dollars undetected.”

Correspondent banking relationships, he says, are also rife with the potential for abuse. Other uncovered cases used foreign currency settlements to launder money.

The challenge for firms is how to leverage new technology as a helping hand for the compliance department. “You try hard you spend a lot of money on hiring a lot of people and you just can’t do it alone,” Heinzman says, adding that both integration with legacy systems and regulatory buy-in, which is becoming more common, are important.

Machine learning and AI offer tools that “give you the ability to find things that you couldn't think of looking for on your own, and even the best experts wouldn’t be able to write a rule,” he says.

Order a Reprint