Can you use autoregressive diffusion to generate market data?

A research intern explored using autoregressive diffusion models to generate synthetic market data, treating order book events as continuous data similar to video or audio. While the fully continuous approach proved inadequate for capturing the discrete, "jagged" nature of real market data—such as the clustering of orders at round numbers and "pennying" behavior—the experiment provided valuable insights into what a practical generative market data model would require, including the finding that flow matching outperformed traditional DDPM for this application.