Harnessing Artificial Intelligence in Market Surveillance: an Aquis case study

The development of Artificial Intelligence/ Machine Learning technology has been a generator of conversations – or heated debates – globally, with recognised potential that has increased its deployment exponentially in recent years and months.

Aquis, in partnership with the University of Derby and supported by Innovate UK, has been working for the past two years on a proprietary market surveillance tool which utilises artificial intelligence and machine learning to provide comprehensive monitoring and anomaly-detection.

Presenting to The AI Summit in London in June, project leaders David Attew (Chief Regulatory Officer) and Dr Baqar Rizvi (Data Engineer) provide an update on parameters, considerations and progress to date.

Huge potential benefits
Over the years, markets have evolved with diversification in trading practices, globalisation and sheer competition with more modern businesses being added every day. Aquis Exchange plays a significant role in the economy and is an integral part of the modern business world. Indeed, it is being monitored by several regulatory bodies, market analysts and researchers in a bid to detect and identify manipulation. However, market surveillance involves a huge amount of data – on Aquis, trades generating daily message traffic in the millions making it computationally expensive both in terms of time and resources.

Additionally, to cope with the complexity of the dataset, technical automation is required for parameters to be optimised as frequently as required (often multiple times a day) and across the many thousands of individual stocks listed on any given exchange.

And it’s not just the market that evolves; so, do the strategies deployed by those engaging in market manipulation. When deployed appropriately, technology is able to detect new patterns quicker than a human team and ensure that new anomalies are flagged and surveyed.

But not without challenges
There are significant challenges involved with deploying AI to meet the roles and needs described above.

The speed of high frequency trading is an obvious one – bid/ask orders are currently executed within nanoseconds, and developments in quantum technology indicate that this may become even faster – picoseconds perhaps – in future. Any technology deployed to monitor this activity requires a high level of robustness and processing power.

The data on an exchange is, for the most part, a huge, heterogeneous, and unlabelled dataset. Even on Aquis, which is one of the newest and most technologically-advanced exchanges, we hold an enormous dataset for any technology to process and means that there is an extensive ‘training’ process required in harnessing that data for learning and pattern identification purposes.

In contrast, there is a lack of publicly-available data around past market manipulation (which would be an invaluable learning tool for future detection), given many of the examples that made it to prosecution stage are not in the public domain.

The Aquis solution
The Aquis AI project is working to create a system which combines traditional logic-based alerts with research-driven anomaly detection techniques.

The data analysis began with Aquis’ own data – raw orders and trades – which was then overlaid with anonymised examples of manipulative schemes that were available in the public domain. This helped to train a market-manipulation detection model based on natural immune system known as Dendritic cell algorithm (DCA). The work suggesting a significant improvement in reducing false positives and increasing the detection rates is now accepted by a peer reviewing committee and published here.

A role for human oversight
Regulation in financial services is principles-based – it’s quite technology neutral, with the onus on regulated firms employing new technologies to analyse and consider the risks prior to implementation.
So, what did Aquis look at when we embarked on this AI project two years ago? We considered things like customer implications, risks, verification (how do we know it works?), resourcing and whether we had the appropriate level of expertise within the firm. As a business headquartered in the UK but also operating in the EU, we had a raft of regulatory considerations, including those prescribed by the FCA, AMF, those under GDPR and the EU AI Act, and best practice as outlined by the UK Government and ESMA.

And then we came to ethics. We had to consider whether what we were doing – and the benefits that we were looking to achieve – and whether there were any less obvious or undesired risks around accountability, transparency, bias, fairness, confidentiality and social benefit.

Ultimately, we realised that AI had to be employed as a tool that sat in the hands of a human agent. The human has to train the system appropriately and has to be the filter for the outputs: to be able to put the information they receive in context and to be held accountable for the judgement calls that are made around use of that information.

The goal
There are clear benefits associated with utilising AI to provide broader and more effective market surveillance of the trading ecosystem. However, the ethical and contextual understanding required means that this can never be a pure technology process or solution.

At Aquis, we want to equip our human surveillance team with the data processing tools made available to us by technological developments, and ultimately run a trading venue analyst oversight system with cutting-edge anomaly detection based on advanced modelling.

AI is not a replacement for market surveillance, but clearly it can be harnessed by human teams to vastly improve data monitoring, analysis and anomaly detection.