Can you use autoregressive diffusion to generate market data?
Summary
Jane Street explores using autoregressive diffusion to generate market data, addressing whether market data should be modeled as continuous or discrete. The post covers Kavish's diffusion and flow-matching approach, issues with instability, and the novel atom smoothing technique to handle discontinuities in timing and price. Results show improvements with flow matching and smoothing, though the model is not yet a fully realistic generator and highlights directions for future work in representing mixed continuous-discrete market features.