The FICC Market Standards Board (FMSB), has published its latest Spotlight Review, which looks at the impact of machine learning on market surveillance functions as part of the broader challenges facing institutions as they establish oversight frameworks. It argues, “Machine learning could lead to fundamental changes in market surveillance given its ability to process complex, large and poorly structured datasets. FICC markets involve huge amounts of disparate structured and unstructured data including quotes, trades, email and voice communications data. In particular, machine learning techniques can proactively combine trade and communications data in a more systematic way through sophisticated natural language processing.
“The need for this will only increase with the growth of alternative data,” the paper continues, adding that market surveillance in FICC has undergone, and continues to undergo, significant change as a result of regulation, the evolution of market structure, and technological developments. It considers these structural and technological changes, in particular the emergence of machine learning trading strategies, and sets out some of the challenges associated with these developments for surveillance teams in FICC markets.
The review then examines the role of technology as a potential solution to these challenges, creating as it does opportunities to improve market surveillance through the application of machine learning.
Market surveillance has been an area of regulatory focus over recent years. In particular, the UK’s Financial Conduct Authority (FCA) has focused on the need for firms to continue to improve surveillance in FICC markets and to enhance both the quality and number of suspicious trade submissions relating to FICC markets activity. Furthermore, in the current remote working context, maintaining robust surveillance and suspicious transaction and order reporting has been flagged as a regulatory priority.
Further enhancing FICC surveillance in increasingly fast moving, complex and data-driven markets, however, is not a simple task, FMSB says, reflecting upon what it sees as two key structural challenges surveillance teams face in FICC markets: namely data quality and availability, and the increasing sophistication of trading strategies and technologies deployed to support them.
There is little doubt that the amount of data available to surveillance functions has increased significantly in recent years, driven by regulatory reporting requirements and the proliferation of electronic trading, however, FMSB says, extrapolating signals from these data sets remains challenging given variances in the accuracy, robustness, timeliness and consistency of such data, in particular across different FICC asset classes. It also points to the challenges associated with the growth in algorithmic trading, systematic investment strategies and the nascent adoption of machine learning in trading, suggesting they are materially increasing the speed and complexity of FICC markets. “This combination of increased data and trading complexity, and the possibility of new market abuse risks emerging as a result of these developments, may drive the adoption of new, or the improvement of existing, surveillance techniques,” the paper states.
The paper argues that technological innovations are creating opportunities to change market surveillance. Whereas as historically the “rudimentary nature” of traditional automated alert systems produced a remarkably high number of alerts, only a small number translated into suspicious transaction and order reporting (STORs) being captured, reported and investigated in FICC markets. “Machine learning techniques, with their ability to process large, complex data sets efficiently from both structured and unstructured data sources, offer the opportunity to make surveillance significantly more effective,” FMSB states.
“Given the increasingly data driven nature of FICC markets and the potential for technological developments to significantly change the nature of market surveillance, a greater understanding of data science and technology is becoming central to the future of market surveillance professionals,” it continues. “However, it is likely that the full potential of the application of machine learning techniques to market surveillance will only be realised in the long term.”
The paper focuses on four key areas; the first being the factors driving the pace of change – namely the weight of data and need for technology solutions to help process it. It also looks at the surveillance challenges related to remote working.
The second key area is what the paper describes as the “acute impact of data of surveillance effectiveness, specifically the amount and completeness of the data,its accuracy and robustness and relevance, as well as its timeliness and consistency. This section also looks at the appropriate use of unique and correlated data, highlighting how the increased complexity and interdependence makes it important for surveillance professionals to understand the sources of price- forming trades and to avoid conducting surveillance of one financial instrument or venue in isolation.
Thirdly, the paper looks at the surveillance of complex algorithms and machine learning, something that was discussed in a recent In the FICC of It podcast with Martin Pluves, CEO of FMSB. The paper follows up FMSB’s first Spotlight Review in April which looked at algo model risk by looking at the market abuse risks and challenges faced when conducting surveillance of machines learning strategies. It highlights evidence of intent, complexity, “gaming” and collusion as the key challenges in this area and stresses the need for oversight teams to be able to quickly interrogate the decision making of the computer.
It adds that a machine learning mechanism, “may not be able to understand the limits of permissibility” and that there may be a need for an “ethical governor” that tests the optimisation process against ethical benchmarks. The paper also highlights the risk of the machine seeking to exploit fault lines in the surveillance governance framework and establishing some sort of “surveillance arbitrage” strategy.
FMSB does acknowledge that computer-generated decisions can in principle be scrutinised, which might not always be the case for human decisions, adding that with the right governance, both the data used to train the algorithm, and the algorithm itself, can be investigated. “There are many market participants that believe the existing governance and surveillance framework for algorithms should be extended to machine learning techniques,” it says. “…As the adoption of machine learning grows, more consensus needs to be built around the best practice in governance structures for managing these challenges.”
The fourth key area investigated looks at the benefits and risks of using machine learning and kicks off with the ominous warning that, “despite the marketing hype, it is not uncommon to see new vendor products fail to deliver on [their] promises when put to the test in real world-market conditions”.
It observes that machine learning is not bound by traditional frameworks when it comes to surveillance techniques and can therefore better learn from experience and leverage historical data by finding relevant market conditions of structures.
The report concludes by stressing the need for agility in effective surveillance, noting that given the view of regulators, there is a need for financial institutions to improve the suspicious trade submissions relating to FICC markets activity “…The ability of machine learning programmes to process large complex data sets efficiently, [means]it is highly likely that machine learning will play a role in the future of market surveillance of market abuse risks,” the paper states. “Working side by side with humans, over time, machine learning programmes may be better able to understand the semantics of data and the evolution of behavioural patterns and to adapt their machine learning algorithms. Consequently, a greater understanding of data science and technology is becoming central to the future of market surveillance professionals so they can effectively specify and test machine learning functionalities.
“As the adoption of machine learning techniques in financial services matures in terms of usage, best practice on safe and effective deployment will emerge, and there may be an important role for practitioner-led industry standards to improve the consistency and effectiveness of market surveillance in fast- developing markets,” the paper adds.
“The role of market surveillance remains a high priority for regulators and financial firms face new challenges during the outbreak of the COVID-19 global pandemic with high volatility and large numbers of sales and trading staff working remotely,” says Pluves. “This ‘perfect storm’ forms the backdrop to this important Spotlight Review which looks at how innovation, including machine learning, can present new problems for surveillance professionals, but conversely may also play an important part in delivering creative solutions for managing market abuse risks in the front office.”
Lukasz Szpruch, programme director for finance and economics at The Alan Turing Institute, which assisted in the paper, adds, “In the Finance and Economics programme at The Alan Turing Institute we are addressing the key challenges of adopting machine learning techniques in the financial services industry by relying on transparent, reliable, and reproducible research. We welcomed this opportunity to work with FMSB on its review of market surveillance. By examining the significant opportunities and risks that machine learning methods offer relative to more traditional rules-based algorithms this paper facilitates an important discussion about the most relevant factors that could impact industry best practices across financial market participants.”