I recently had the pleasure of being part of the FCA Market Abuse Surveillance Tech Sprint (MASTS). The eflow team, together with eight other firms, presented our findings to a wide range of Surveillance and AI specialists at the FCA’s headquarters in Stratford, London.
The aim of the Tech Sprint was to have firms and research institutions approach three problem statements:
- Improved accuracy of market abuse detection: How can AI in post-trade market abuse surveillance help reduce false positives, produce a higher proportion of true positives, resulting in more accurate alerts and signals?
- Detection of instances of complex market abuse: How can AI identify signals or instances of more complex types of market abuse, involving the analysis of multiple data sets that traditional rules-based surveillance tools currently struggle to identify?
- Transforming market abuse surveillance by incorporating anomaly detection: How can AI help identify previously undetected anomalies indicative of market abuse, manipulative strategies and disruptive trading practices that may give participants an unfair advantage?
The eflow team focused on problem statement 1, looking at how AI and Machine Learning (ML) can be used to drive an improvement in both analyst efficiency, by highlighting high-value alerts, and also examining how clients can have parameters that are dynamically adjusted – being fit for purpose for how they trade both now and in the future.
By utilising an alert feedback loop to allow an ML process to learn what could be considered high or low value alerts, we can understand commonalities in the results. Whether that be time traded before the release of Material Non-Public Information (MNPI) for an insider dealing alert, or the time for cancellations of a non-bona fide alert for spoofing, patterns that highlight what a client deems high risk can help drive both alert risk scoring and the tuning of thresholds.
This approach was further reinforced through the training of ML models on the huge real-world dataset provided by the FCA, where traditional, parameter based, surveillance behaviours help further refine and identify those events that would be considered high risk (wherein both ML and traditional parameters agree). It also allowed us to identify the value-added proposition; alerts identified by the ML process that were not identified by the traditional parameter solution, which links back to the dynamic parameterisation proposition.
Exploring new innovations in the world of regulatory compliance
Of course, we were not the only firm with some great ideas about how to approach trade surveillance. We were all amazed to see students from UCL and King’s College London present their approach to problem statement 2 studying market dynamics and market behaviour through the integration of an econophysics model. This demonstrated how approaching the market as a physical system with mass and velocity can lead to novel approaches in identifying unusual trading activity before news events.
We were also treated to some wonderful presenters, Benjamin Levy from B-Next had us searching for invisible (or very well hidden!) snakes in the grass whilst drawing analogies with German radar in World War II. Ross Pearson and Kenn Rodrigues from DHI-AI utilised their current work with the Australian Securities and Investment Commission (ASIC) to drive a greater understanding of how Bayesian networks can drive anomaly detection, all wrapped up in a dynamic presentation style that closed the presentations with vim and a call to the bar!
In addition to the presenters from the vendor and research side, we were also fortunate to have some great speakers from the FCA share their wisdom and insight into the markets with us. Jamie Bell, Head of Market Oversight, made the point that as AI and ML can be used to drive trading decisions, so should we understand that there’s a growing need for ‘more sophisticated market abuse detection tools that can monitor evolving market dynamics and detect complex forms of market abuse’. That is the crux of what we were all trying to do; to ensure that the integrity of the marketplace can be maintained in the face of increased risk.
What does this mean for the application of AI and ML in financial regulation?
So what are the next steps? Well, for eflow, the next step is to continue to build out a feedback loop and work with clients to get real-world alert risk ranking feedback. This will ensure that we can create a best-in-class mechanism to allow for the review of risk-scored alerts.
While it is not the role of a trade surveillance provider to tell a firm how to approach their risk, but if we can help them to get to their risk quicker then we will ensure that they can approach the compliance processes with more certainty and conviction.
However, for all of us at the Tech Sprint, the next step was a fantastic social get together on the roof terrace of the FCA building where we were all treated to pizza, ice cream (I might have avoided the surveillance radar to get an extra 99 flake) and a chance to feedback what we had learnt from each other’s presentations.
For me, it was the sense that everyone knows what they need to do. AI and ML are undoubtedly the future, but one that needs to be approached in lock-step with testing, tuning and current, parameter based, solutions. Whilst no one vendor had the solution for future surveillance ready yet, I do not doubt that we all, particularly here at eflow, will be at the forefront of where the technology will be.
Follow the links for more information on eflow’s trade surveillance and eComms surveillance solutions to combat the threat of market abuse.
