Oxford scientists develop GPU-accelerated limit order book sim to teach AI how to trade

A multidisciplinary research team from the University of Oxford recently developed a GPU-accelerated limit order book (LOB) simulator called JAX-LOB, the first of its kind. 

JAX is a tool for training high-performance machine learning systems developed by Google. In the context of a LOB simulator, it allows artificial intelligence (AI) models to train directly on financial data.

The Oxford research team created a novel method by which JAX could be used to run a LOB simulator using only GPUs. Traditionally, LOB sims are run using computer processing units (CPUs). By running them directly on a GPU chain, where modern AI training occurs, AI models are able to skip several communication steps. According to the Oxford team’s pre-print research paper, this gives a speed

Training an AI system to understand LOB dynamics is a difficult and data-intensive task that, due to the nature and complexity of the financial market, relies on simulations. And the more accurate and powerful the simulations, the more efficient and useful the models trained on them tend to be.

According to the Oxford team’s paper, finding ways to optimize this process is of the utmost importance:

“Due to their central role in the financial system, the ability to accurately and efficiently model LOB dynamics is extremely valuable. For example, it might allow a financial company to offer better services or may enable the government to predict the impact of financial regulation on the stability of the financial system.”

As the first of its kind, JAX-LOB is still in its infancy. The researchers stress the need for further study in their paper, but some experts are already predicting that it could have a positive impact in the fields of AI and fintech.

Jack Clark, co-founder of Anthropic, recently wrote:

“Software like JAX-LOB is interesting as it seems like the exact sort of thing that a future powerful AI may use to conduct its own financial experiments.”