Can diffusion models generate market data?

In a post on its technical blog, Jane Street describes an attempt to generate synthetic order-book data with a causal transformer. A categorical head predicts the type of the next event, and a diffusion head predicts its price and timing.

Training with flow matching was stable where a standard denoising diffusion setup blew up, but rollouts over many steps still drift away from realistic behaviour. The post is a clear account of why data that mixes discrete events with continuous values resists generative modelling.