Artificial intelligence is one of the biggest buzzwords of the year, and it can seem as if the world and his wife are advertising ever more innovative and ambitious solutions. But where is AI having the most material impact – can it really help traders to improve execution, is it yet being used at the coalface, and what impact will the latest regulations have? Laurie McAughtry looks at some of the latest solutions.
A broad church
It might be used as a sweeping ‘catch-all’, but AI in trading encompasses a broad church: with applications ranging from pattern recognition and analysis to market data processing and scaling, workflow automation to liquidity sourcing, TCA improvement to chat interfaces, and much, much more. Some have compared AI and machine learning to the original electronification of trading, with those not investing now likely to get left behind.
A multitude of new players have sprung up to create a vibrant new market, from fintechs such as ALFO DeepTech developing multi-asset AI-driven risk management and hedging tools to Switzerland-based Starfetch specialising in the research and development of AI-based trading algorithms. DataMotion estimates that the fintech sector will spend over US$26bn on AI by 2026, while UBS analysts in May 2023 predicted that the AI hardware market could hit US$90bn by 2025, expanding at a compound annual growth rate (CAGR) of 20%. It’s a trend that has the potential to change the face of financial markets – but will it be for the better, and how widely are these tools actually being used yet?
“As much interest as AI is attracting right now, we’re in the early days as far as understanding the true opportunities for these technologies to revolutionise the financial industry,” warned Anju Kambadur, head of AI engineering, Bloomberg.
“AI is reaching peak hype – everyone’s doing it, everyone’s talking about it and in some cases people are just trying to find a use case to stuff it into,” agreed Andy Mahoney, managing director of FlexTrade in EMEA.
New applications
Bloomberg has been using AI, machine learning, and natural language processing for more than a decade (since 2009) – not only to create efficiencies within its solution stack and optimise functionality on the Bloomberg terminal, but also to develop entirely new signals for financial decision making.
“We’re constantly speaking with customers to understand the workflows – on and off the Terminal – that they are looking to automate, considering the role AI can play in that, and evaluating how we might be able to help,” said Kambadur.
The big AI splash this year might be ChatGPT and all its myriad applications, but these chat interfaces are already making their way into the finance space, and could be the easiest access point for AI opportunity. In March 2023 Bloomberg introduced its own BloombergGPT, a 50-billion parameter large language model built from scratch for the financial markets, which it believes represents “the first step in the development and application of this new technology” for the financial industry.
The model will help to improve existing financial natural language processing (NLP) tasks: such as sentiment analysis, named entity recognition, news classification, and question answering, as well as developing new ways to curate the vast quantities of data available on the Bloomberg Terminal. But it’s early days. Although the firm is exploring possible applications for the tech with its product teams and clients, there is as yet no commercial product nor user interface, Kambadur told Best Execution. Instead, it is likely to be used as an “ingredient model” to power AI-enabled applications on the Terminal, such as making searches more intuitive and unstructured data (i.e. text, tables, charts) more liquid.
“We believe that institutional investors are more likely to use AI tools for streamlining idea generation and pre-screening tasks so that the staff who conduct these processes can work more efficiently,” added Kambadur. “For example, an analyst might use AI tools to identify a selection of instruments that meet certain criteria and risk tolerances, and then conduct human review, validation, and further analysis before making recommendations. One trend in particular that we’re watching is new strategies that are a synthesis between traditional discretionary and systematic groups.”
Another possible use case is to enable clients to use conversational English to generate code when composing queries using Bloomberg Query Language (BQL), to pinpoint data for import into data science and portfolio management tools. Some products already use NLP to convert unstructured communications into structured content so they can quickly be analysed for investment decisions – such as finding the best price to trade a given instrument or reducing the time between RFQ and quote. Another example is financial sentiment analysis, a new way to use textual data in quantitative models.
Asking the right questions
Bloomberg aren’t the only players making waves when it comes to converting unstructured data. In May 2023 FlexTrade rolled out an artificial intelligence application within its OEMS which it claimed would “revolutionise” how trading teams interact with their data and trading technology solutions.
‘FlexA’ (working title) is a new AI natural language interface layer across FlexTRADER EMS which aims to simplify the interaction between the trader and their technology platform, allowing them to more easily and intuitively process large quantities of information. The application can be controlled via both voice commands or a text-based chat interface, meaning that users can ask either verbal or written questions to achieve what FlexTrade call a “deep situational awareness” of orders within their blotter, using a combination of specific criteria or characteristics (for example: “What is the value of my Swiss orders over 10% ADV?”). They can also request further information – such as ad-hoc reports on historical trading to help with pre-trade decision-making, using questions like “show me a chart of my US traded venues for US orders in February”. They can also request actions to be automated to speed up workflow, for example: “Create a ticket for my US orders less than 500k using the VWAP AlgoWheel”.
“Initiatives such as ‘FlexA’ present a glimpse into the future of trading technology and what is possible,” said Mahoney. The solution uses the API from ChatGPT to “reduce the barriers between human and machine” whilst retaining human control.
“We have constrained what the AI can do to make sure that it’s safe,” emphasised Mahoney. It allows you to reduce your workflow down to a single interaction. I think the most valuable impact is being able to grab any amount of technical trading data without the need for complex syntax knowledge. The interface layer no longer needs to be a table or a spreadsheet where you have to go back to your data provider asking questions. You can just freeform your trading query and receive an actionable answer.”
The firm has also utilised machine learning to achieve real-time market impact analytics within its EMS – an area that has traditionally been dominated by financial models and cost estimates. “Looking at the ways in which trading analytics can be developed, one area suggested by our clients was a pre-trade analysis resource that can run continuously in the background, that can vary based on market conditions, rather than having a standard number of inputs and outputs,” said Mahoney. “If you continuously run a standard econometric pre-trade model over and over again, it will pretty much just give you the same result. If you can evaluate instantaneous market impact instead, this gives traders faster access to actionable insights and better decision-making.”
Data-driven
Gabino Roche, CEO and co-founder of New York-based AI fintech Saphyre, has spent the past few years working out how to use AI to help traders with their original computer algorithms – to review market data, and determine the best possible trades. “Although this technology exists, the challenge is feeding the AI clean and correct contextual market data in order to train itself so that it can determine such decisions consistently and reliably,” he explained.
Saphyre’s current solution automates documentation to help with the trade lifecycle. It digitises, structures, standardises and maintains memory of shared data and documents (as well as permissioned counterparties) in pre-trade; which then automates and expedites downstream trade lifecycle processes.
One example of this could be in the case of a block trade with, say, 20 funds allocated. Right now, there is no real-time status transparency at the allocation level. In such cases, the broker pushes through the allocation, only to discover issues during reconciliation that means they have to amend the trade or allocation. This only elongates the time to settle – a huge disadvantage, especially given the current move towards T+1. “Having real-time statuses on allocated accounts, enriched by AI driven data from pre-trade, can get you not just to T+1, but even T-0,” said Roche.
There are huge opportunities for efficiency improvement here, and not just at the coalface. “The traders have most of the latest and greatest tech, but you’d be surprised when it comes to the middle/back office,” revealed Stephen Roche, Saphyre president and co-founder. “They still rely on manual processes and old systems – even faxes, believe it or not. Many have automated solutions but many of these are just one-trick ponies. What’s missing out of their automated solution is ‘intelligence’. Something that can recall information and correlate it to the lifecycle of a fund with context.
“We will [soon] be announcing an exclusive project where we will transform the post-trade in the middle/back office. Not to just help prepare their firms for the T+1 initiative but to even provide them with T-Zero scenarios by using this pre-trade data.”
No silver bullet
Traders themselves are very aware of the importance of this new technology. “The field of AI is rapidly evolving and holds immense potential to reshape our jobs in ways we may not fully comprehend yet,” said Elke Wenzler, head of trading at MEAG and our cover star for this edition (see p.14). “We maintain a strong collaborative relationship with our technology and innovation team, actively exploring opportunities for leveraging AI. One particular project that excites me is our endeavour to reconstruct our FX trading process using AI, covering every step from initial trade decisions to execution.”
But they are also aware of its limitations.
“While AI is often discussed as a magical solution, it is not always the answer to every problem,” said Wenzler. “Engaging with AI provides invaluable learning experiences, showcases future possibilities, and highlights the necessary skills for future success. As data continues to proliferate, it may become increasingly challenging for individuals to manage all available inputs effectively. AI can help us access and interpret information in novel ways. However, it is crucial to blend our expertise with AI to ensure the accuracy and relevance of the information we receive. Merely accepting AI output is insufficient; we must delve beneath the surface to scrutinise and understand the underlying mechanisms, ensuring its contribution to profitability. Ultimately, it is about utilising our tools in the most efficient manner possible.”
The human touch
The applications are endless – and exciting – and there are many more firms out there pushing the boundaries of what can be achieved. But there are still reservations around how far AI could, or should, be taken.
“Much of the concern around AI comes down to accountability,” explains Mahoney. “We need filters and restrictions in place – and who decides what those filters will be, and how you know you’ve got the right ones? Even in the case of a low touch order, a machine is making a rule based on its own assessment. Who owns the risk of that logic? Right now, AI is a way of bypassing long workflow operations, but it doesn’t have the ability to make decisions. You can set alerts, you can set Algo Wheels, you can set rules and recommendations, but the next challenge is going to be how regulators perceive this. Where an AI is determining which orders go to a broker, how does that get assessed? We need to make sure the right factors are being taken into account for best execution purposes.”
“AI still needs human intervention in the near-term,” agreed Stephen Roche. “It’s in its infancy state, but we’ll get closer and closer to the singularity point. I think within 10 years’ time it will be possible to see fully mature products and solutions pumped out for industry at an exponential rate. What that rate will be, well, that’s anyone’s guess. On the trading side, it can be very scary. The algorithms could be considered so much more reliable that they beat any human trader. Then again, an outside economic force that it hasn’t calculated or seen before could provide a human advantage in those instances (hence human oversight).
“Even worse, it could collude with other AI systems to manufacture market movements for certain outcomes that could be very hard to track. The need to have oversight and seeking continuous vigilance has never been more important than the times we are entering now.”
Get ready for regulation
That regulation is already underway. In April 2021, the European Commission proposed the first EU regulatory framework for AI, defining AI systems that can be used in different applications and requiring them to be analysed, classified and regulated according to different risk levels. On 14 June 2023 the EU’s AI Act vote passed with an overwhelming majority, making it the world’s first set of regulatory rules for the space – described by Roberta Metsola, the president of the European Parliament, as “legislation that will no doubt be setting the global standard for years to come”.
Structured in a similar way to the EU’s current Digital Services Act, the AI rules take a risk-based approach with firms submitting their own risk assessments. It introduces a ban on ‘emotion recognition’ AI (and a potential ban on facial detection and analysis), along with a ban on real-time biometrics, predictive policing, and social scoring, all of which fall under the category of “unacceptable risk”.
More importantly for the fintech and financial services space, the draft laws discuss the possibility of regulating generative AI, particularly around the use of copyrighted material within large language models. All gen AI models, such as ChatGPT, would have to comply with transparency requirements: including disclosing that the content was generated by AI, designing the model to prevent it from generating illegal content, and publishing summaries of copyrighted data used for training.
It’s a big step forward, and one that could have potentially significant repercussions for the application of AI within Europe. However, there is still some way to go. Now that the draft has been approved, the legislation moves to negotiations with the European Council, with the goal of reaching an agreement by the end of the year – after which, the real process of implementation begins.
©Markets Media Europe 2023
