How Hudson River Trading Actually Uses AI
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this episode, Joe Weisenthal and Tracy Alloway speak with Iain Dunning, head of AI research at Hudson River Trading, about how artificial intelligence is transforming quantitative trading. Dunning explains how modern AI models differ from traditional algorithmic trading approaches, revealing that HRT's trading is now entirely driven by machine learning systems that process massive amounts of market data. (08:18) The conversation covers the practical realities of deploying AI in high-frequency trading environments, from hardware constraints to risk management protocols.
- Core discussion focuses on the evolution from handcrafted trading features to AI-driven models that can predict short-term price movements with greater accuracy than traditional methods
Speakers
Iain Dunning
Iain Dunning is the head of AI research at Hudson River Trading, a major US market maker and quantitative trading firm. He previously worked at DeepMind, Google's AI research lab, bringing deep expertise in artificial intelligence to the financial markets. At HRT, Dunning leads efforts to apply cutting-edge AI techniques to trading strategies, moving the firm beyond traditional algorithmic approaches to fully AI-driven market making and prediction systems.
Joe Weisenthal
Joe Weisenthal is co-host of Bloomberg's Odd Lots podcast and a Bloomberg Markets Live blogger. He focuses on financial markets, economics, and the intersection of technology and finance.
Tracy Alloway
Tracy Alloway is co-host of Bloomberg's Odd Lots podcast and a Bloomberg Markets Live reporter. She covers market structure, trading, and financial innovation with particular expertise in derivatives and risk management.
Key Takeaways
AI Models Can Achieve Short-Term Market Prediction
Contrary to efficient market hypothesis skeptics, AI models can successfully predict short-term price movements with accuracy rates of approximately 50.1% - only slightly better than random, but sufficient to generate significant profits at scale. (13:15) Dunning explains that while the predictions aren't highly accurate in absolute terms, the marginal improvement over randomness becomes extremely valuable when applied consistently across thousands of trades. This challenges the conventional wisdom that markets are too noisy or non-stationary for machine learning approaches to work effectively.
Market Data Trumps Alternative Data for Intraday Trading
For short-term trading horizons (minutes to hours), traditional market data feeds from exchanges provide far more predictive value than alternative data sources like social media sentiment or satellite imagery. (18:36) Dunning emphasizes that market data represents the "most true expression of everyone's intents" as participants quote, buy, and sell in real-time. While alternative data may have value for longer-term strategies, the immediate flow patterns captured in order book data contain the richest signals for intraday prediction models.
AI Trading Requires Massive Engineering Infrastructure
Successful AI trading isn't just about having better models - it requires building an entire technology stack capable of collecting, storing, and processing petabyte-scale datasets in real-time. (41:18) This includes custom hardware, data centers, and teams that can reliably stream vast amounts of market data to training systems and then deploy models with microsecond-level latency requirements. The engineering complexity creates a significant barrier to entry that favors established firms with deep technical resources.
Electricity Has Become the Primary Constraint
Power availability has emerged as the main bottleneck for scaling AI trading operations, even for mid-sized firms like HRT that consume "tens of megawatts" rather than gigawatts. (37:22) Dunning describes electricity negotiations as the key factor when planning new GPU-based training data centers, with some firms resorting to installing gas turbines outside buildings to get power quickly. This constraint affects not just hyperscale cloud providers but any organization trying to deploy AI at meaningful scale.
Risk Management Requires Multiple Layers of Human Oversight
Despite full automation, AI trading systems require extensive human-audited risk controls to prevent catastrophic failures like the Knight Capital incident. (50:40) Neural networks don't directly send orders to exchanges - instead, they provide plans that pass through traditional risk management layers with multiple sanity checks throughout the trading day. This approach reflects the zero-tolerance culture from regulators and the recognition that operational errors can permanently damage a firm's ability to access markets globally.
Statistics & Facts
- HRT's AI models achieve approximately 50.1% accuracy in predicting short-term price movements - only slightly better than random but sufficient for profitability at scale. (13:35)
- Hudson River Trading consumes "tens of megawatts" of electricity for their AI operations, which Dunning notes is "more than most towns and cities." (40:04)
- NVIDIA reported producing 4 million Blackwell class GPUs compared to 1 million Hopper class GPUs, indicating rapid supply scaling in the GPU market. (38:38)
