Interview
Finding Signal in the Noise: Machine Learning and the Markets with In Young Cho
- Jane Street shifted in 2013 from anonymous electronic exchanges to a model involving bilateral agreements and financial advice via phone trading with end clients, a trajectory expected to continue with voracious data consumption of tens of terabytes daily to support model scaling.
- The organization plans to expand its hardware fleet to over 5,000 high-end GPUs, intending to learn weekly about machine learning capabilities and hardware requirements while navigating a tradeoff between robust production systems and flexible exploratory environments.
- Complex deep learning models are expected to remain the dominant predictive trend through at least the remainder of the current year, with multimodal data including images, text, and market data targeted to improve price predictions.
- New developments in large language models (LLM) and high-compute reinforcement learning are anticipated to integrate with trading models over the next couple of years to cover more of the trading process.
- Machine learning capabilities will expand via transfer learning to apply understanding from established asset classes to data-scarce domains, addressing the necessity of using priors when massive datasets are unavailable.
- Internal assistants and agents leveraging ML are planned to improve employee efficiency, though the specific methodologies for leveraging foundation models, such as fine-tuning in-house versus partnering with vendors, remain unclear.
- Jane Street intends to reshape its modeling approach by prioritizing hardware limitations and potentially designing hardware to fit modeling needs, moving beyond a sole focus on model optimality.
- The company anticipates addressing "regime changes" caused by external events like pandemics through robust training methods that adapt to shifting market distributions.
- Significant challenges regarding research reproducibility are expected due to the friction between evaluation tools like Python notebooks and production system requirements.
- New researchers and traders will be trained using fundamentals-based introductions to distill complex tooling into intuitive knowledge before exposure to the full modern data stream.
- Jane Street expects a high learning rate regarding machine learning capabilities, acknowledging past incompetence with hardware and ongoing uncertainties in the field.
- Predictions regarding the organization's capabilities in 2026 are expected to be "absolutely wrong" given the tumultuous developments in AI over the next couple of years.
- The path to deploying the most powerful AI toolkits to traders via assistants is viewed as unclear, involving complex decisions on priority building and data requirements.