Introduction

Jane Street is one of the top quantitative trading firms in the world, relying heavily on advanced technologies for its high-frequency trading systems. Their use of computing clusters—high-performance networks of powerful computers—often fitted with top-of-the-line GPUs from the likes of Nvidia, such as the Nvidia H-100 and A-100 GPUs, enables them to train and run inference on Machine Learning models on massive financial datasets quickly. This is vital for predicting market trends and managing risks in real-time.

Why do clusters matter in finance

Financial markets around the world generate huge amounts of data, such as stock prices and trading volumes, which their systems need to access quickly. Clusters like Jane Street’s, fitted with thousands of high-end GPUs used for parallel processing and over 323 petabytes of storage, provide the speed and scalability needed. This allows them to adapt to sudden market shifts and trends quickly and make sub-second trading decisions.

Imitating Jane Street’s approach

For those interested in programming for clusters and parallel processing, languages like CUDA for Nvidia GPUs are helpful. You can start with cloud-based cluster platforms on AWS or Google Cloud, using datasets from sources like Quandl. Alternatively, you could build a cluster of Raspberry Pi’s for a local approach, where just 10 of these small computers offer 40 ARM cores and up to 160GB of RAM—perfect for cluster computing.

Jane Street

Jane Street, a prominent quantitative trading firm based in New York, has established itself as a leader in leveraging technologies for trading in financial markets. With offices in London, Hong Kong, Amsterdam, Chicago, and Singapore, the firm trades across more than 200 venues in 45 countries, handling trillions of dollars in securities annually. Their 2024 revenue reports $14.2 billion, capturing over 10% of the US equity market. Their success is underpinned by their heavy use of machine learning and advanced computational techniques, particularly through the use of large computational clusters.

Their approach is technology-driven, with the majority of their employees writing code (Python, C++, OCaml) as part of their regular work and building almost all of their software in-house including some of their advanced OCaml compilers. A main focus is placed upon real-time visibility and reliability in their systems.

Role of Clusters in Quantitative Finance

What are clusters

Computing clusters are groups of interconnected computers, often fitted with high-performance GPUs, that work together to perform large-scale computations. In finance, where markets generate 10+ terabytes of data every day in the stock market alone, these clusters are essential for training ML models that require immense computational power. Jane Street’s infrastructure is particularly notable, with their machine learning page revealing a cluster of thousands of H-100 and H-200 GPUs with 323 petabytes of storage. The scale of their cluster is critical for handling the demands of AI models, such as real-time analysis for algorithmic trading, risk modelling, and portfolio optimisation. For example, their ability to train deep networks on vast historical datasets enables them to predict market trends and execute trades via their high-frequency trading systems.

How clusters address challenges in Financial AI

Financial markets present several issues for AI, which clusters can help mitigate.

  • Data is often noisy, with much if it being irrelevant, as Jane Street notes on their Machine Learning page describing financial markets as producing a “Torrent of data, mostly noise”. Clusters enable efficient filtering and processing of this data, leveraging parallel processing to handle the large datasets quickly.

  • Additionally, the market conditions undergo frequent structural changes or, as put in the Jane Street article “regime shifts”, due to factors like pandemics, elections and new regulations and legislation. Jane Street’s Machine Learning page highlights this, stating that market data is “regime like” requiring models to adapt rapidly this depends on the computational power of the cluster as smaller networks such as a 500 node LSTM model is able to adapt quickly but deep networks for example with 500 billion - 1 trillion parameters which could be slow on less power hardware. But, their clusters support distributes training, allowing for quick retraining of models to adjust to these shifts.

  • Latency constraints are critical in trading, where decisions must be made in milliseconds. Clusters provide ultra-low latency systems, as mentioned in their job postings for Machine Learning Researchers, which discussed the need for extremely low latency constraints.

  • Trading actions can create complex feedback loops, influenced the very data being modelled. Clusters enable Researchers to run sophisticated simulations to understand these very loops, enhancing model robustness.