AI's Achilles' Heel: Why Trading Desks Lag Behind in Fintech, According to Quantum Signals Expert

Wall Street banks have increasingly integrated AI digital assistants into various operations over the past year, leveraging these tools for research, coding, and other knowledge-intensive tasks to boost productivity. However, this broad adoption masks a significant challenge in one critical area: live trading.

According to Yianni Gamvros, CEO and co-founder of Quantum Signals, a Paris-based firm specializing in finance-native AI for intraday systematic trading, general-purpose AI models struggle where market data moves on a millisecond cadence and is largely numerical. This "market microstructure data" is precisely where current AI falls weakest, hindering its genuine usefulness on the trading desk despite widespread enthusiasm.

The disparity arises because the AI assistants widely deployed in the sector are primarily designed for tasks like reviewing knowledge bases, researching news, and assisting with coding, areas where Large Language Models (LLMs) excel. In contrast, intraday trading and execution demand real-time visibility into rapidly changing price and volume data, which must be acquired from exchanges. Trading desks need to ascertain immediate market direction, assess liquidity for trades, and predict intra-session volatility. These crucial insights are embedded in market microstructure data, which general models cannot access and are poorly equipped to interpret due to its numerical, time-series nature rather than textual format.

Furthermore, many AI trading projects encounter failure during production deployment, largely because building robust, production-ready systems is far more arduous than initially anticipated. Even achieving a promising model is difficult, often involving extensive offline pre-processing and normalizations. The critical hurdle is then implementing these steps robustly and in real-time for every prediction the model makes. This requirement for both speed and accuracy in live environments often forces compromises between research and implementation teams, highlighting the immense practical challenges.

The limitations of general-purpose AI in financial markets necessitate a shift towards finance-native models. While general AI is adept at understanding broader macro contexts, interpreting news, product announcements, and earnings calls—which are valuable for macro strategies—it is largely irrelevant for quants building intraday, mid-frequency strategies, or traders executing large trades. Finance-native models, by contrast, are specifically engineered to handle the unique characteristics of financial data, particularly the high-frequency numerical time-series data crucial for tactical trading decisions.

Beyond specialized models, proprietary data and robust infrastructure are paramount for establishing a competitive AI trading capability. Firms lacking these foundational elements, along with skilled AI and data engineers, will realistically find it challenging to compete and may need to resort to third-party solutions. The edge in AI trading is not solely derived from the AI itself; financial expertise plays an equally vital role. AI can accelerate progress significantly, potentially getting users 70-80 percent of the way, but human intervention—through multiple prompts, guidance, and corrections—remains essential to achieve genuinely useful and actionable outcomes.

In intraday systematic trading, AI already demonstrates significant capabilities. It can effectively detect key market dynamics such as price movement, liquidity, and volatility, identifying and predicting patterns with a degree of confidence. However, converting these predictions into full-fledged trading strategies—determining trade size, and precise entry and exit timings—still requires further development. While AI can uncover patterns in data, it often does so without inherently understanding their underlying causes. Sophisticated quants leverage their domain knowledge to connect these patterns to real-world events, a gap that AI may or may not close in the foreseeable future.

The industry often misapplies AI in trading due to several misconceptions. Many confine AI's role to text review, text generation, or coding assistance. On the quantitative side, a preconceived notion persists that AI is unhelpful, stemming from the limited capabilities of traditional machine-learning techniques used in past decades. Critics often point to the zero-sum nature of markets, the small signal-to-noise ratio, and constantly changing market characteristics as insurmountable barriers. However, newer AI techniques have shown promise in addressing these challenges, mirroring advancements seen in areas like Go mastery, medical pattern detection, and adaptive self-driving systems.

Looking ahead, fundamental changes in models, data, or infrastructure are crucial for AI to become genuinely reliable in trading. The emergence of foundational finance models, analogous to those in text, image, video, robotics, and driving domains, is already a reality. Public information suggests these are actively employed by top-tier funds such as Citadel, Jane Street, and HRT, performing a great deal of work even if they don't solve every problem. Consequently, firms without access to these advanced models risk falling increasingly behind in the competitive landscape of financial markets.