Wall Street banks have spent the past year putting AI digital assistants to work across research, coding and other knowledge tasks, with reports describing lenders ramping up the tools in a productivity race.
Adoption has been broad, but one area has proved harder than the headlines suggest. Live trading depends on market microstructure data that moves on a millisecond cadence and is largely numerical, which is precisely where general-purpose models are weakest.
In this written Q&A, Yianni Gamvros, CEO and co-founder of Quantum Signals, a Paris-based firm building finance-native AI for intraday systematic trading, discusses where today’s AI still falls short on the desk, why so many AI trading projects fail in production, and what would have to change before the technology becomes genuinely reliable.
Wall Street banks are rapidly rolling out AI digital assistants, yet you argue that AI still falls short in trading specifically. Where is the gap between broad adoption across the sector and genuine usefulness on the trading desk?
The AI assistants being rolled out are used mainly for reviewing knowledge bases, researching news and coding. Large language models are very good at all of those tasks.
During intraday trading and execution, a large language model has no visibility into the price and volume data that changes on a millisecond cadence and has to be bought from the exchanges.
The trading desk needs to work out where the market is going over the next few minutes or the next few hours, whether there is enough liquidity to cover a trade, and whether expected volatility will be higher later in the session. Those answers sit in the market microstructure data, which the model cannot access and is poor at interpreting, because it is largely numerical time-series data rather than text.
You say most AI trading projects fail because building production-ready systems is far harder than expected. What tends to break between a promising model and a system that can run live in the markets?
Getting to a promising model is very hard to begin with. The team that builds it typically performs several offline pre-processing steps and normalisations, all done once and offline. The model is then trained, and the results may be promising.
The problem is that every step the researchers took now has to be implemented robustly so it can be repeated in real time, every time the model makes a prediction. ‘Robustly’ and ‘in real time’ are the real challenges. There are cases where the two teams have to compromise and find something that is good enough to feed the model and can also be calculated quickly and accurately in real time.
Why is general-purpose AI not enough for financial markets, and what does a finance-native model do differently?
General-purpose AI is very good at reading the news, product announcements and earnings calls. It can understand the broader macro context and whether there are supply-chain risks. If you are building a macro strategy, those capabilities all help. If you are a quant building intraday, mid- frequency strategies, or a trader who needs to execute a large trade, they are largely irrelevant.
You place proprietary data and infrastructure above simply using bigger models. For a firm without either, is a competitive AI trading capability realistically out of reach?
Yes. Infrastructure, data and sophistication, meaning AI and data engineers, are all key requirements. A firm that does not have them needs to look at buying third-party solutions.
How much of the edge comes from the AI itself, and how much from the financial expertise around it, and how do you combine the two in practice?
As with general AI, it can get you 70 to 80 per cent of the way but not all the way. You can ask a model to write a review, an email or an article and it will get you down that path quickly, but you have to re-prompt it several times, guide it and correct it to reach something genuinely useful. We have done some of that, and we are providing prediction streams that have been curated to show what the technology can do.
Your focus is intraday systematic trading. What can AI do well there today, and what remains out of reach for now?
It can detect key market dynamics such as price movement, liquidity and volatility. It can find patterns around these and predict them with some confidence. Converting those into trading strategies takes more work: how much to trade, and when to time the entries and exits. Sophisticated traders will also be able to ask the AI to test and research more complex patterns across contracts, sectors and market dynamics.
AI can also uncover patterns in the data without connecting them to their underlying causes. Sophisticated quants know the causes and can look for different patterns driven by real-world events. It is unclear when, or whether, AI will close that gap.
What is the industry most often getting wrong in how it applies AI to trading?
Many people think of AI only as text review, text generation or a coding assistant, and nothing more. On the quant side, many have a preconceived notion that AI does not help, because the traditional machine-learning techniques used over the past few decades had limited capabilities. People tend to point to the zero-sum nature of markets, the small signal-to-noise ratio and the way market characteristics constantly change.
Newer AI techniques have shown that all of these can be addressed. We have seen AI models beat Go champions, detect patterns in medicine better than humans, and adapt constantly to changing conditions, as self-driving models do.
Over the next couple of years, what has to change, in models, data or infrastructure, before AI becomes genuinely reliable in trading?
Foundational finance models, in the way there are now foundational text, image, video, robotics and driving models, are possible today, and public information suggests they are already used at firms such as Citadel, Jane Street and HRT, and other tier-one funds. They do not solve every problem, but they already do a great deal. Everyone else, without access to those models, is falling behind.
The post AI is Everywhere in Finance, but the Trading Desk is Where it Still Falls Short, Says Quantum Signals’ Yianni Gamvros appeared first on The Fintech Times.
