HRT Just Trained An LLM On Market Microstructure Data. Here’s What That Means — And What It Still Doesn’t Solve.

At the recent ICML 2025 conference, Mark Khoury from Hudson River Trading (HRT) described the firm’s development of foundation-style transformer models trained on decades of market microstructure data. The result is a trillion-token-scale dataset spanning equities, futures, and crypto. Matt Robinson from AI Street, provided some great coverage of the presentation.

It’s a huge step forward in modeling the markets as language. And it’s part of a broader trend: domain-specific LLMs, trained not on Reddit or Wikipedia, but on granular, structured data sequences like limit order books, trade fills, and cancellations.

The excitement is justified. These models can spot patterns across millions of micro-movements that no human could process. And they’re showing improved predictive performance as both model size and data volume scale.

But here’s the catch:

Even the best predictor still needs a reliable executor.

Where Predictive Models End, Business Logic Begins

Imagine you’re a quant or product manager at a hedge fund or trading platform. Your model predicts that a specific pricing condition will trigger a trade. Great.

Now what?

How do you calculate position sizing under internal compliance rules?
How do you validate margin requirements under jurisdiction-specific regulation?
How do you adjust risk thresholds based on client-specific logic?

These aren’t probabilistic guesses. They’re deterministic calculations, and they need to be:

Accurate
Auditable
Versioned
Reproducible

Every one of these requirements can trigger a custom code development project. That often results in a brittle solution with high maintenance costs. A platform like TrueMath offers a much more efficient alternative.

Predictive Models Are Getting Better. Execution Still Needs Structure.

What HRT and others are building is undeniably impressive. But those LLMs are built for reasoning, not executing.

They don’t:

Store and version logic over time
Keep an immutable audit trail
Ensure the same inputs always produce the same result
Respect business rules baked into your org or regulatory environment

TrueMath does.

We’re a deterministic, graph-based math engine built to serve as the execution layer beneath LLMs so that once a model makes a prediction, your system can calculate exactly what to do next, under provable assumptions and documented rules.

Think of It This Way:

HRT models forecast the weather
TrueMath calculates how much to charge for flood insurance
LLMs say, “This scenario might happen”
TrueMath says, “Here’s how to respond with the right math”

LLMs and TrueMath: Better Together

Just as vector databases didn’t become obsolete when embeddings improved, the rise of market-modeling LLMs actually increases the need for deterministic execution infrastructure.

That’s because the more you rely on AI to interpret complex data, the more you need a trusted system to turn those interpretations into repeatable, compliant decisions.

If you’re building LLM-based finance tools, this is your moment to ask: