AI has been, is, and will continue to be one of the hottest topics in trading, with large-scale language models (LLM) making headlines in the latest news. However, panelists at FIX EMEA in London asserted that the technology is still best suited for support services rather than decision-making.
“You’d be surprised at their eloquence when they’re just lying or hallucinating,” said Eric Bos, global head of trading at Allianz Global Investors. “And we’re talking about proper industry-grade, large-scale language models here, not open source.”
“I would say that clients are using AI more as a funnel to widen the flow of information and gather unstructured data to support decision-making, research notes, and morning briefings. I don’t think anyone I’ve talked to is using AI as something like cheap, independent algorithmic trading,” added John Bryant, senior vice president and field chief technology officer at Options IT.
One reason for this is the lack of certainty regarding the results of LLM, Voss explained, contrasting this with previous generations of AI, which were used to build algos through machine learning techniques.
“Sophisticated algos know exactly what they should do, and they know when to deviate from what they should do. You can’t reverse engineer AI models to the same degree. This is part of the big problem with AI in financial services. Algorithms are not linear decision trees, so the same problem doesn’t always lead to the same outcome.”
He added that the line between Argos and AI can be “very blurry” from an outsider’s perspective. Where do you put regular stock and FX algorithms that are already self-optimizing by taking intraday trading patterns and volume trends? This one learns a bit and adapts to the given market environment. I don’t want to apply this to the field of AI, but from the outside it looks like it’s operating almost autonomously. ”
Traders have been using machine learning, an early version of AI, for years. This technology can recognize patterns and make decisions and predictions based on them. Generative AI, popularized by Chat GPT in 2022, can create “new” content. We are now on the verge of the emergence of agent AI that can make decisions and adapt to new situations without continuous human intervention.
BlackRock published a paper last year on how multi-agent systems can handle stock picking.
“By employing a team of expert agents who can reason with market data and fundamentals, we demonstrate that this collaborative framework can improve investment processes and outcomes,” they said. They also considered how agent systems work to combat cognitive biases in trading.
“AI agents can process information at scale, but they don’t have the contextual judgment or accountability of humans. If we want to take the next step, we need to build governance around possible outcomes,” said Petros Kyriakoudis, head of trading research and analysis at Baillie Gifford.
“The buy-side is still in the early stages of incorporating AI into live trading workflows. Full automation requires a level of governance and trust architecture that most companies are still building.”
Predictability and explainability are important factors.
“The key is traceability, tracking the agents and understanding what is going on,” said algorithmic trader Hanane Dupuis. Education is also key, she stressed.
“We need to let people who are using these kinds of systems know what their limitations are. If we want to validate something, we need to understand what we’re validating.”
Panelists agreed that, at least for now, traders should use AI tools instead of relying entirely on technology.
“Some of the smart auto-routing features incorporate historical data and automatically adjust for market conditions such as volatility and spreads. However, market conditions are different every day. Yes, you need to understand the history, but the conclusions you draw from it are different from day to day. That is where the creativity of trading comes into play and where humans are still needed,” Boss commented.
“Completely automating trading is not even my goal. I think that’s probably a bad idea to begin with,” he added.
