Odd Lots

How Hudson River Trading Actually Uses AI

October 31, 2025

Key Takeaways Copied to clipboard!

  • Hudson River Trading's current trading strategy is entirely driven by large, data-consuming AI models that have superseded older, handcrafted feature engineering methods, similar in concept to how large language models like ChatGPT are trained. 
  • For intraday trading predictions, the most useful data source is raw, low-level market event data (quotes, trades) purchased from exchanges, rather than alternative data sources like news feeds. 
  • Competitive advantage in modern quantitative trading relies on optimizing the entire technology stack—from data ingestion and model training (compute/electricity constraints) to low-latency inference and rigorous operational risk management—rather than solely on minimizing physical latency. 

Segments

HRT Business Model Explained
Copied to clipboard!
(00:04:52)
  • Key Takeaway: Hudson River Trading operates as a service provider to markets, primarily through market making across various asset classes by quoting tight, optimal prices.
  • Summary: HRT functions as a sophisticated middleman, providing liquidity by being ready to buy or sell stocks, futures, options, and crypto at the best possible prices. They profit by capturing the spread between the bid and ask prices, analogous to picking up pennies in front of a steamroller. This service provides utility to market participants by ensuring a counterparty is always available for trades.
AI vs. Traditional Quant Trading
Copied to clipboard!
(00:06:09)
  • Key Takeaway: The shift to AI represents a seismic change from traditional quant trading, moving from handcrafted features based on human intuition to large models that consume internet-scale data autonomously.
  • Summary: Older algorithmic trading relied on smart people crafting features (e.g., order book imbalance) and using simple models like linear regression. Modern AI trading uses massive models trained on all available data, which has overtaken the traditional approach entirely within HRT’s operations. This new method is fundamentally different because the models learn patterns without human structural bias.
Data Crunching vs. Pattern Spotting
Copied to clipboard!
(00:09:19)
  • Key Takeaway: The utility of AI in trading stems from its ability to process internet-scale market data volumes, which is necessary because financial market events generate data sets too vast for manual feature engineering.
  • Summary: The power of AI here is both in execution speed and pattern recognition, driven by the sheer volume of low-level market events (quotes, trades). The ‘bitter lesson’ in AI suggests feeding massive, generic neural networks with large data sets is more effective than over-engineering features. The resulting models are often not interpretable, which is expected when the system is already superhuman at short-term prediction.
Short-Term Price Prediction Capacity
Copied to clipboard!
(00:11:17)
  • Key Takeaway: Despite efficient market hypothesis skepticism, AI models can achieve slightly better-than-random (e.g., 50.1%) predictive capacity for stock prices on short time scales like minutes or hours.
  • Summary: These models can make informed, albeit slightly biased, predictions about the near future price of a stock, which becomes highly profitable when executed at scale. The predictive power comes from sucking up micro-signals from the flow of buyers and sellers, which are the primary drivers of intraday price movement. Predictions beyond the day scale, relying on fundamentals, are outside the scope of these market-flow models.
Most Useful Data Ingredients
Copied to clipboard!
(00:16:43)
  • Key Takeaway: For intraday prediction (minutes/hours), raw market data feeds detailing quotes and trades are overwhelmingly the most useful input, while alternative data becomes more relevant for multi-day horizons.
  • Summary: The most counterintuitive finding is that raw market data feeds, which are widely available, provide the clearest expression of market intent for short-term forecasting. Alternative data like SEC filings and news feeds are more relevant when predicting prices over days, where fundamental analysis plays a larger role. The market for alternative data is vast, but much of it may lack significant predictive value.
Interpretability and Model Reasoning
Copied to clipboard!
(00:20:16)
  • Key Takeaway: The lack of interpretability in high-performing neural networks stems from their internal processing methods being fundamentally different from human cognition, making them opaque ‘blobs of numbers’.
  • Summary: Neural networks are trained to be free of human structure, leading them to learn in ways that do not map to human reasoning, hence the difficulty in explaining their decisions. While some models show specific interests (like the Golden Gate Bridge example), mapping internal processing back to human thought remains challenging. Solving this interpretability gap is crucial for productivity gains across many AI applications.
LLM Training Process Similarities
Copied to clipboard!
(00:22:42)
  • Key Takeaway: The training processes for HRT’s trading models and frontier Large Language Models (LLMs) have become highly similar due to shared requirements for handling long sequential data strings and demanding high compute efficiency.
  • Summary: Both types of models deal with long histories and require efficient churning of vast amounts of data, making the research applicable across modalities. Both systems must also serve predictions/responses promptly, necessitating optimization for inference speed. HRT’s ’tokens’ are market events, analogous to text tokens processed by LLMs.
Data Moats and Off-Exchange Volume
Copied to clipboard!
(00:24:49)
  • Key Takeaway: The increasing prevalence of off-exchange, dark volume creates an anti-AI trend where hidden flow data provides an advantage to firms ‘in the room,’ challenging the democratization of data access.
  • Summary: While market data feeds are democratized, significant trading volume occurs in opaque venues where data is not promptly reported or accessible for machine consumption. This hidden flow data creates an advantage for those physically present in those trading environments, which is contrary to the data-rich environment AI thrives on. This lack of transparency is a long-term issue for purely data-driven models.
Hardware Stack and Latency Evolution
Copied to clipboard!
(00:26:56)
  • Key Takeaway: The competitive edge has shifted from minimizing physical wire length (latency arbitrage) to maximizing decision intelligence (smartness) for a given response speed, utilizing custom hardware like FPGAs alongside GPUs.
  • Summary: The race to shorten physical wires is largely complete, meaning the focus is now on the intelligence of the decision made within the available response time (latency vs. throughput curve). HRT uses custom hardware (FPGAs/chips) and off-the-shelf GPUs to optimize this trade-off, as they cannot batch requests like web-based LLMs due to the non-stoppable speed of market events. Training occurs in cloud/private data centers, but inference requires co-location near exchanges.
Constraints: Electricity Over GPUs
Copied to clipboard!
(00:32:30)
  • Key Takeaway: For large-scale AI training infrastructure, the primary long-term strategic constraint is securing reliable and affordable electricity, not the availability of GPUs, which has improved significantly.
  • Summary: While GPU availability was a crunch point previously (late 2023), supply has ramped up, making electricity the binding factor for expanding data centers. HRT finds that negotiating power supply, sometimes requiring the immediate installation of gas turbines, is the bottleneck for growth. This power demand raises significant concerns about the feasibility of massive future data center expansion plans.
Competitive Edge and Scale
Copied to clipboard!
(00:35:55)
  • Key Takeaway: Competitive advantage is derived from optimizing the entire proprietary stack—talent, data collection/storage, model training, and deployment—making it extremely difficult for new firms to enter the space.
  • Summary: Talent remains competitive, requiring individuals skilled in both research and engineering, as ideas must be implemented immediately. The true edge lies in the seamless integration of petabyte-scale data handling, expensive training runs, and reliable serving infrastructure. Scale provides a significant barrier to entry, as the initial engineering lift required to build this comprehensive stack is immense.
Trading Horizon Expansion
Copied to clipboard!
(00:38:20)
  • Key Takeaway: HRT has expanded its business beyond pure high-frequency trading into medium-frequency trading (holding positions for days), where models inform longer-term acquisition strategies.
  • Summary: The firm now operates as both a high-frequency and medium-frequency trading entity, moving away from being purely a liquidity provision service provider at longer horizons. Longer-term trading involves acquiring positions based on a view of what a stock should be in several days, potentially using shorter-term models to optimize the timing of that acquisition. This shift means they sometimes act as liquidity takers, paying transaction costs when acquiring a position based on a directional tilt.
Trading as Positive Sum Game
Copied to clipboard!
(00:40:31)
  • Key Takeaway: Financial trading is fundamentally a positive-sum game driven by differing utility horizons among participants, unlike zero-sum games like chess or Go.
  • Summary: Trading works because participants have different time horizons and risk preferences; for example, a long-term investor is happy to trade with a high-frequency firm for better immediate liquidity, even if the price moves slightly against them later. If markets were purely competitive deathmatches between the smartest players, trading volume would cease, as everyone would wait for the perfect moment.
Balancing Open Research vs. IP Protection
Copied to clipboard!
(00:42:28)
  • Key Takeaway: The historical tension between engineers wanting open research and trading firms needing secrecy has lessened as frontier AI labs have become more secretive, making proprietary IP protection the norm across the industry.
  • Summary: Previously, researchers preferred Big Tech for the ability to publish, but now the most cutting-edge AI work is inherently secretive, aligning with trading firms’ need to protect intellectual property. This shift means that the value of process knowledge gained through years of internal work is now recognized across the tech landscape, justifying strong IP protection measures like non-competes.
Avoiding Rogue Algo Disasters
Copied to clipboard!
(00:45:06)
  • Key Takeaway: To prevent catastrophic failures like the Knight Capital incident, HRT employs deep operational paranoia, relying on multiple layers of human-audited risk checks between the AI model’s plan and live order execution.
  • Summary: The AI model provides a plan, but heavily audited, risk-checked layers execute the final actions, ensuring no neural network directly sends orders to exchanges. This includes rigorous pre-release checks, daily sanity checks on model outputs, and extreme vigilance regarding regulatory compliance to maintain trust with global regulators. The culture prioritizes long-term survival in the game over ‘move fast and break things.’
Post-News Event Market Reaction
Copied to clipboard!
(00:48:32)
  • Key Takeaway: In the microsecond after major news releases like jobs reports, markets move based on automated systems reacting to initial data flashes or text before any human can process the information.
  • Summary: Automated systems, using low-latency headline feeds or AI models trained on text, react instantly to new data, often before the full context of the report is understood by humans. The inability to reliably backtest LLMs (which have memorized past events) against future unpredictable news events highlights a key challenge in using them for high-speed reaction trading compared to traditional finance backtesting.