Improving Trade Surveillance Through Technology

Financial institutions are expected to demonstrate compliance by reviewing all of their trades adequately and generating useful alerts. Nasdaq recently participated in a panel discussion at XLoD Global, which focused on the technologies that can alleviate the pain points that undermine efficiency and increase costs in trade surveillance.

At a fundamental level, if surveillance teams don’t have good quality data, they could end up dealing with a mountain of false positives. Admittedly, false positives are necessary because that’s how surveillance teams prove to regulators and internal auditors that they’re doing their job. On the other hand, false positives can also take capacity away from operations. There must be a balance.

Part of the problem is that trade capture systems were designed to make trading more efficient and increase profitability, not to enable trade surveillance. While data quality is paramount to being able to set thresholds and tune out the noise, it’s tough to capture all of the data and establish a lineage map to track its journey. 

Ultimately, trade surveillance systems run alert logic programs and generate alerts based on the data that’s input into the system. If the system is missing key pieces of data, then it can only provide alerts based on what it has received in the first place. Some alerts can be resolved relatively quickly by asking the right questions to the sales and trading team. But as one panelist put it, it would be much easier for surveillance analysts to recreate a scenario if they could see what the traders see on their screen when they quote a price. The more data surveillance analysts have, the better the questions they can ask.

Surveillance analysts wear a lot of hats nowadays. They’re expected to be data scientists, project managers, presenters and writers as well as investigators. However, machine learning, artificial intelligence and behavioral analytics can provide them with valuable support in navigating their various roles. 

Machine learning algorithms are extremely useful, but they also have some limitations. One is that surveillance analysts have capabilities, skillsets, knowledge, backgrounds and gut feelings that aren’t captured in a set of rules that can be applied to a machine learning process. To this end, machines could come to the right conclusion, but the logic is accidental.

Another issue is that you can’t tell a machine learning algorithm to go through a massive dataset and detect potentially suspicious activity, such as layering and spoofing. You need to be specific, and that involves: 

• identifying a task that human beings can break down into rules and then solve;
• communicating those rules to the machine, so the machine executes in the way that humans execute that task; and
• comparing the outputs to see if there’s a predictable relationship between the output from machines and humans.

Essentially, the machine needs to understand what it is about a dataset that makes it interesting, and you’ve got to describe it in a way that the machine can understand. Instead of trying to detect layering and spoofing, for example, it’s better to break the activity down into components. To illustrate, machine learning algorithms can be taught to detect when the market is trending, and there’s large order cancellation. They can be taught to differentiate between aggressive and passive trading, as well as to identify depth of the market and cancellation near the depth vs. away from the depth. You can then apply an overarching algorithm that knows when these conditions are met, and a case of layering and/or spoofing is likely. 

According to the panelists, behavioral analytics systems are the icing on the cake. Behavioral scientists interview analysts and get them to describe exactly how they do their job. They baseline the behavior of traders, their peers, trading desks, as well as people trading the same products in order to be able to identify exceptions. Through this research, they provide cognitive assistance to normalize and standardize analysts’ tasks.

A key takeaway from this session was that financial institutions need to be able to explain technology-enabled decisions. Artificial intelligence and machine learning-based systems should be able to produce a set of human-readable rules, which can be presented to compliance officers and regulators. In addition, financial institutions need to prove that they’re constantly testing their program and validate that they’re targeting the right risks.