HRT Trains AI Models on Trading Data
The quant firm has developed transformer-based models using decades of market microstructure data.
Hey, it’s Matt. Welcome back to AI Street. This week:
HRT on Building “Foundation Models for Automated Trading”
Top Papers on AI in Finance Q4 2025: SSRN
JPMorgan’s Dimon: banks must invest in AI or get left behind. + More News
Hudson River Trading is building foundation-style models trained on decades of global market data, applying techniques similar to those used in frontier language models for automated trading.
The firm is training these models on more than two decades of data spanning equities, futures, and cryptocurrencies, totaling over 100 terabytes. That translates into “something like trillions of tokens, in the same realm as what you train frontier language models on,” said Marc Khoury, a researcher on HRT’s AI team, speaking at an academic conference.
At a high level, HRT’s goal is to model markets as sequences of interactions. Electronic markets generate detailed streams of activity, including full limit order books, executed trades, and order-level events such as placements, cancellations, and fills. According to Khoury, much of the predictive signal lies in how these sequences evolve over time, especially during fast-moving conditions.
(We’ve previously discussed how training transformers on specific domains — the weather, payments, grocery store sales — have shown promising results.)
Khoury showed charts indicating that more data boosted the model’s predictive performance. “As I increase the model size, the model continues to improve,” Khoury said, adding that the pattern mirrors what researchers see when scaling large language models.
To train a transformer model on markets, trade data has to be converted into tokens. In language models, tokenization typically breaks text into subword pieces. See the below example from OpenAI’s tokenizer:
In market data, tokenization means mapping events into discrete units a model can learn from. One approach groups events into fixed time intervals, such as one-minute windows. Another bundles a fixed number of events together. HRT described both as active design choices, each with tradeoffs for model performance and cost.
“Language modeling people would call this tokenization,” Khoury said. In this context, each token represents a unit of market activity rather than a word. The goal is to preserve market structure while keeping training and inference computationally feasible.
For more background on tokenization of market data, check out my chat with Juho Kanniainen, a professor at Tampere University’s Data Science Research Centre, on treating limit order books as “language” here.
Transformer models offer one way to address this problem. Because they are designed to process long sequences of diverse events, transformers can adapt to changing conditions without relying entirely on fixed, hand-engineered features.
Training models at this scale requires substantial infrastructure. Khoury did not disclose how many GPUs the firm operates, but he said HRT runs its own state-of-the-art data center and that at one point, the firm’s hardware purchases were large enough that he joked it “bottlenecked” GPU deliveries on the U.S. East Coast.
The research effort comes as HRT posts record results. Bloomberg reported that the firm generated an estimated $12.3 billion in net trading revenue in 2025. The company has also been expanding beyond high-speed trading into longer-horizon strategies, while continuing to invest heavily in artificial intelligence and data-driven trading models. The firm accounts for roughly 10 percent of U.S. equity trading volume.
An HRT spokesperson declined to comment on the firm’s AI training efforts.
Takeaway
After tokenizing 100 terabytes of market events into a trillion-token-scale training set, HRT showed internal results where predictive accuracy kept improving as they scaled both data and model size, mirroring the scaling behavior seen in LLMs.
Outlook
There’s a lot to unpack with the implications of HRT building a foundation model for finance. Khoury’s presentation shows they’re getting promising results, which to me leads into all sorts of other questions:
How many other funds are pursuing similar market-sequence models?
Are computing resources becoming a meaningful edge for market makers?
If transformers keep outperforming traditional AI models, how big of an impact is that on GPU demand and data-center buildouts?
Further Reporting
If you’re building in this space, please reach out by replying to this email. I’m still mapping the landscape and have a lot more questions.
Related Reading
RESEARCH
Top Papers on AI in Finance
SSRN just published its Q4 2025 leaderboard for AI-in-finance research: the 10 most-downloaded papers posted that quarter, plus which organizations did the downloading. The list is a quick snapshot of what’s getting traction right now, with “agentic” workflows, LLMs in systematic investing, and GenAI model risk showing up alongside core quant portfolio work.
The most downloaded paper was “Agentic Artificial Intelligence in Finance: A Comprehensive Survey” by lead author Irene Aldridge. I interviewed her last month about how AI agents are the next evolution in trading.
Big themes: agentic AI (survey + real estate), ChatGPT for systematic investing, GenAI model risk management, ML-meets-Markowitz portfolio construction
Also notable: learning firm characteristics for asset management, investor automation and effort allocation, LLMs for central bank communication
COMPUTE
Tracking “Compute” Constraints
We hear more about AI spending and the building out of data centers than we do about the overall demand for compute, or the hardware that trains and runs AI. Or at least, I’ve found it hard to track down this sort of information.
So I “built” a compute capacity constraint index using Claude Code. The name is a bit misleading. You do not spend much time coding with it. It does the coding for you.
My goal is to track whether the Amazons, Microsofts, Googles, and other infrastructure have more customer demand than they do compute capacity. My thinking is that this gives some sense of AI demand and whether or not all this supply being built will be met with more demand.
This is very much a work in progress, so if you have ideas on how to structure this, please reply to this email.
