What’s unique about machine learning at Jane Street?

We break new ground in predictive modeling, developing approaches that have never been tried before. Here’s some of what we’re up against:

• Financial markets produce a torrent of data, mostly noise.
• We need ultra-low latency systems just to process it.
• Market data is “regime-y,” in that it undergoes frequent structural changes in reaction to things like pandemics, elections, new regulations—even swings in the collective mood.
• Identifying these distinct distributions can be tricky, let alone learning against them.
• And just when we think we’ve got it, our own actions influence the data we’re trying to model (itself a very fun problem to think about).

What does machine learning look like at Jane Street?

• No silos. Researchers, engineers, and traders sit a few feet away from each other and work together to train models, build systems, and run trading strategies.
• Depending on the day, we might dive deep into market data, tune hyperparameters, debug distributed training, or study our model’s trades in production.
• We build on the latest papers in LLMs, computer vision, RL, training libraries, cuda kernels, or whatever else we need to train good models.
• We invent our own set of architectures, tricks, and optimizations that work for trading.