Market operators are awash with big data, and computing technology has matured to the extent where we can gain significant insight from it. Nowhere is this trend more evident than in market surveillance. Today’s algorithmic rules-based approach used to detect market abuse is now being complemented by machine learning and artificial intelligence, providing the opportunity to eliminate bias in the analysis and discover new patterns in the data.
SMARTS is already deploying artificial intelligence to achieve 360 degree holistic market surveillance. Unstructured data from electronic communications such as emails, instant messages and social media are being tied to structured surveillance data such as orders, cancels and amendments. This provides a complete view of individual traders’ communications including who they corresponded with internally and externally across every channel. The underlying technique is called cognitive computing – where artificial intelligence is used to process large volumes of unstructured data including natural language and speech to detect intent, collusive communications, and relationships between different parties and entities.
Other artificial intelligence techniques create additional paths to tackle opportunities to increase the efficiency in market surveillance such as trader profiling and the predictive evaluation or scoring of abnormal events. Through trader profiling, machines group particular types of traders together based on to their styles of trading – e.g. block trading or high frequency trading with large numbers of small orders. Then they look for clusters of traders to establish what is considered normal. Cluster specific trading patterns or outliers can then point towards a suspicious change in behavior or potentially illicit trading activities. Profiling relies on a technique called unsupervised learning, where machines can find previously unknown groupings within the data.
“It is still early days,” says Tony Sio, Head of Nasdaq SMARTS Market Surveillance. “However we have already managed to get good results, and we are now incorporating it into our product and into the surveillance team’s daily workflow.”
SMARTS will also be adding the alert scoring capability to its offering for regulators and exchanges in the upcoming SMARTS 7, allowing alerts to be classified based on the level of risk. Alert scoring leverages supervised learning, where the machine relies upon historical cases to predict which new abnormal events are most likely to require investigation. In this case the data is the key, and SMARTS has one of the largest training sets available to teach the machine.
Today’s surveillance systems rely on contextualized, finely calibrated, detection algorithms, a programmatic approach where a set of rules is used to look for certain behaviors, calibrations can be run or new rules are generated to capture novel behaviors. This mirrors what human analysts do when they apply their own knowledge and experience to interpret an alert. Artificial intelligence is increasingly being leveraged by firms to provide additional context during investigations of specific behaviors triggered by rules-based alerts. The combination of both rules-based alerts as well as artificial intelligence reduces false positives, enabling firms to become more efficient in their monitoring activities, and increases true positives, allowing them to proactively detect activities that may have not been identified with trade surveillance data alone.
“In the current deployment we will augment rules-based alerting with artificial intelligence to give users the best of both worlds,” says Sio. “The rules-based approach can capture all apparent attempted and actual manipulations, and then the machine can score the alerts so analysts can prioritize their investigations.”
In practice, rules-based alerts have thresholds. If a trade has moved the price of an instrument by a certain percentage, then that will be enough to generate a rules-based alert. But analysts also use additional levels of filtering. For example, they may know that a security was issued recently, which accounts for the change in trading behavior.
Machine learning can improve the rules by capturing analysts’ knowledge and implicit connections within the data to discover new patterns that are not on the radar screen. The algorithms can also detect patterns of behavior that are not necessarily part of the alert logic – for example, if a firm normally only trades in companies in a specific sector – and apply them automatically.
“We're taking black and white rules-based alerts, and we're adding grayness to them,” Sio adds.
SMARTS is looking at the way alerts have been categorized or managed historically. Market behavior and analysts' input are being used to build artificial intelligence models of what makes an alert good or bad. When a new alert is generated, it will be possible to predict in real time whether it is useful or not, and whether it is a high-priority alert or not. Artificial intelligence will be used for alert scoring and electronic communications in the upcoming version of SMARTS. When these tools were run against Nasdaq’s own surveillance data, they were shown to have predictive capabilities.
The next challenge is for surveillance professionals to consider how the technology may change the way they do their job in practice and how to make best use of these new capabilities.
Food for Thought
- Here are some questions that surveillance teams should think about when they implement artificial intelligence based tools:
- Alerts can be measured in several ways. What determines a good alert, and are there multiple axes that the machine can learn upon?
- What is the best practice to incorporate discoveries made by the machine learning tool into the rule-based alerts? If the machine learning tool discovers a new pattern, should it be left for the machine learning tool, or should that discovery be incorporated? It may make sense to have a process to operationalize the incorporation of those discoveries into surveillance processes.
- Should artificial intelligence be used as a tool for real-time prediction or oversight? As a real-time prediction tool, it can improve the efficiency of the analysts’ workflow by directing them to the most interesting cases. By not showing the analyst the prediction, it can be used as an oversight tool to find cases that had been incorrectly handled.
- Can we train a machine on a subset of analysts, and then use it to help train new analysts? The machine learning tool could check the less-experienced analysts’ work, and if necessary, escalate an alert to a more experienced analyst.
Want to hear more about Nasdaq’s Machine Learning in Surveillance, read more and register for our breakfast discussion at World Exchange Congress.
Contact us for more information: firstname.lastname@example.org
The views and opinions expressed herein are the views and opinions of the author and do not necessarily reflect those of Nasdaq, Inc.