Technologies like artificial intelligence and machine learning have the potential to transform surveillance systems from the traditional rule-based methods to more predictive, risk-based processes.
Most firms, though, have not begun introducing these technologies into surveillance programs.
During a session at Financial Industry Regulatory Authority’s RegTech conference held in New York City recently, a poll of the room showed that the vast majority (69%) are not using any machine learning techniques for surveillance and monitoring.
Only 2% of the room said they were substantially using machine learning techniques in production, while 15% said they were using these techniques in small-use cases in production and 10% were just starting to experiment with machine learning techniques.
During this same session, executives from Citadel Securities, E-Ttrade Financial, Barclays and FINRA discussed how they were using machine learning technologies.
Jon Kroeper, executive vice president of FINRA’s quality of markets group, is responsible for surveillance and investigations in the equity and fixed income markets at FINRA. He gave a detailed background on FINRA’s surveillance program as it exists today.
In 2018, FINRA took in nearly 67 billion events per day; at its peak, the regulator took in 135 billion events per day. “Events” include orders, order routes, order cancellations, quotes and trades.
According to Kroeper, FINRA combines the data from the billions of events it receives and makes what it calls a “cross-market data model.” FINRA then runs its surveillance programs — which it calls surveillance patterns — against that data.
Currently, the surveillance program has more than 172 surveillance patterns that cover equities, fixed income, options and cross-product surveillance. In those 172 patterns, there are more than 308 distinct threat scenarios.
“We have a mature and robust environment of rule-based surveillance patterns,” Kroeper said.
A surveillance pattern covers a general type of behavior FINRA is trying to identify. A surveillance pattern may contain one or more threat scenarios. Each threat scenario seeks to identify a discrete type of “potentially violative behavior,” according to Kroeper.
For example, the wash sale surveillance pattern includes threat scenarios that look for wash sales, prearranged trading, painting the tape and money pass, according to Kroeper.
“The output of that is provided to our analysts, analysts evaluate that — whether they are indicative of intentions or behavior — and then we go out to the firms and do an investigation if we can’t explain away those exceptions,” Kroeper explained.
A move to the cloud a few years ago has helped FINRA to start implementing machine learning in some areas.
FINRA moved its market-surveillance data storage and processing to the cloud between 2014 and 2016. According to Kroeper, the cloud has enabled FINRA to store large data sets not possible through existing data appliances and access the computing power necessary to best leverage machine learning and other advanced analytic techniques.
From there, FINRA has started using machine learning models to develop the parameters for a number of its existing rules-based surveillance patterns.
In addition, Kroeper said his team is “running a few machine-learning surveillance patterns in pilot, parallel to some of our existing surveillance patterns, so that we can evaluate their performance, results, and how they can best supplement our program going forward.”
Kroeper believes machine learning and other technologies will play a key role in FINRA’s surveillance program in the coming years. However, he said, it may not entirely replace its existing rule-based pattern portfolio.
“There still may be a place for rules-based programs in certain areas,” Kroeper said.
Firms like Citadel Securities and Barclays have, similar to FINRA, started using these technologies to enhance current systems.
At Citadel Securities, Gregg Berman, director of research for market integrity, monitoring and surveillance, said the firm does use some machine learning and some artificial intelligence techniques. But, he added, “it’s not to replace the rule-based systems. It’s to enhance.”
Berman said these technologies can come in use, for instance, when problems arise where there’s a huge set of data that has some missing points.
“Computers can do that,” Berman said. “A human could do that, [but he or she would get bored]. Computers are great because they don’t get bored.”
Similarly, John Stecher, managing director and chief innovation officer at Barclays, said his firm’s approach has bee to look at AI and machine learning as “human augmentation” rather than replacing humans altogether.
“Leveraging these techniques across the board free up people to actually do what they were trained to do. Which is solve interesting problems for customers, not just sit there and screw around on spreadsheets,” Stecher said.
At other firms, like E-Trade Financial, using machine learning technologies in surveillance programs widely is a few years away.
John Davidson, senior vice president and global head of anti-money laundering at E-Trade Financial, explained that many broker-dealers like E-Trade —which he called “large by broker-dealer standpoints so we may have resources that others don’t,” — are still in very early stages when it comes to implementing machine learning technologies.
From Davidson’s perspective a lot of that is because of the regulatory pressure.
E-Trade has some machine-based learning that’s incorporated into money movement modules, but, according to Davidson, it’s “very minimal.”
Meanwhile, on the trading front, he said E-Trade is “not doing machine learning and I think we are several years away from any practical implementation of that.”
— Related on ThinkAdvisor: