FLASH FRIDAY is a weekly content series looking at the past, present and future of capital markets trading and technology. FLASH FRIDAY is sponsored by Instinet, a Nomura company.
Automated decision making has always been a mixed bag. This year has signified a turning point in using AI and machine learning tools for the key execution processes. Not long ago, we perceived any kind of automation – electronic, algorithmic, self learning, etc. – as a helper, not as a front office instrument.
AI and ML: From Signal to Strategy
In a recent episode as part of the risk management summit in London this summer, the event organizers offered a very unique experiment. They simulated some kind of economic stress scenario – a sort of black swan, so to speak – that would affect broad markets by catching the traders by surprise. The idea was to study the depth and speed of a human versus a machine operator’s reaction facing quick losses and overall instability of their investment portfolios. Ironically, that event was hosted by major European banks and some big tech companies. As part of that man-made experiment, the bot surprisingly managed to safeguard over 90% of the portfolio assets and even attained some modest gain. It acted quickly and boldly – unlike human managers, who seemed lost in assessing cross-asset impacts and other interdependencies – before they even realized it was too late to start bailing out the sagging portfolio.
However, industry players get more and more grasped by what they perceive as the immense scale of AI’s applicability and usability. I tend to take a more cautious approach.
According to one recent study by Bloomberg, European portfolio managers are reluctant to deeply integrate these technologies into decision-making. Only about 13% of them admitted the acceptability of outsourcing existentially important processes in their companies. I think the outcome of that poll makes a lot of sense. Also, beyond any doubt, although these technologies have made a huge forward leap recently, their structural and functional algorithms remain largely in limbo. The very fact that we don’t know how AI executes self learning, what exact historical inputs it uses, and how deeply any particular application is able to study such historical patterns – yes, I’m talking about the intricacies of backtesting – leads me to believe that we are still essentially at the experimental phase.
Compliance in the Age of Complexity
In terms of the regulatory and compliance impact of these new technologies, I see them mostly as an additional daily burden on a trading desk, forcing brokerage companies to appoint an independent actor whose only task would be to monitor and implement new directives and practices as they emerge and become new regulatory norms. That said, the regulatory and compliance practices are not uniform across the Atlantic. For example, in Europe, its recently introduced AI Act classified various systems of AI used in trading without further separation of them into particular categories – such as aiding routine processes, helping analyze big data, comparing patterns and number arrays from those directly involved in reporting and deduction making – labeling them ‘high risk factors’. That was intuitively understood, but, honestly, not entirely helpful.
Beyond the Buzzwords
Another emerging class, where AI and machine learning tools find more and more use, are crypto platforms. They obtained a huge momentum recently due to the rapid development of digital assets. Decentralized exchanges, along with various tokenized asset issues and smart contracts, are facing their own unique challenges, given complexities and the lack of uniformity of European and U.S. regulations. In the U.S., the Securities and Exchange Commission and CFTC remain stringent about the acceptance of various DeFi protocols. Furthermore, they require rigorous know your customer and anti-money laundering procedures, as well as enforced audits of smart contracts, stablecoins, their collaterals, and the necessary public disclosures.
My biggest concerns about the applicability of these tools stem from a few questions that everyone is free to independently investigate. If we assign an AI to a particular task, sourcing of our queries may include broken or non-existent links and sources. In a situation where we can manually check all of the offered materials, that may not be a really big issue. What we usually do is discard the trash. However, if such a process becomes self-operated and self-controlled, then such zero output can create a real mess in the entire algorithm.
Another familiar example is the AI’s blindness towards the date stamps. It may address the inquiry irrespective of their historical time, which means once left unattended, the AI may give an inaccurate and outdated trading recommendation based on some conditions and factors, that now appear to be no longer valid.
