Hudson River Trading is Building Foundation Style Models
Artificial Intelligence & Machine Learning News
According to AI in finance journalist Matt Robinson, Hudson River Trading is building foundation style models trained on decades of global market data. Apparently using methods similar to those behind frontier language models.
At an academic conference, Marc Khoury, a researcher on HRT’s AI team, revealed that the firm is training these systems on more than twenty years of data spanning equities, futures, and crypto markets. The scale is notable. The dataset exceeds one hundred terabytes, which Khoury described as amounting to trillions of tokens. That puts the effort in the same general range as the data volumes used to train leading large language models.
The conceptual shift is equally important.
Rather than treating markets as static snapshots or isolated features, HRT is modeling markets as sequences of interactions. Modern electronic markets generate dense and continuous event streams. These include full limit order books, order placements, cancellations, and executions. According to Khoury, much of the predictive signal emerges from how these sequences evolve over time, particularly during fast moving or stressed market conditions. Timing, ordering, and microstructure dynamics become central rather than secondary inputs.
One of the most striking findings presented was that scaling continues to matter. As both model size and training data increased, predictive performance kept improving. The behavior mirrors the scaling laws observed in large language models, where larger systems trained on more data often show steady gains rather than early saturation. For quantitative finance, this challenges long held assumptions that market noise quickly overwhelms returns to scale in modeling.
HRT’s work fits into a broader research trend gaining traction across applied machine learning. Transformer architectures trained on domain specific sequential data are showing strong results in areas such as weather forecasting, payment systems, and retail transaction modeling. Markets, with their high frequency, rule based structure, and massive data exhaust, appear increasingly well suited to the same approach.
If these results continue to hold, they point toward a future where financial markets are understood less through handcrafted signals and more through large scale sequence models that learn market structure directly from data. The implications for trading, risk management, and market understanding could be substantial.
