Jane Street explores autoregressive diffusion for generating market data

On October 9, Hacker News covered a market data generation experiment by Jane Street intern Kavish. The work builds an event-level generative model with autoregressive diffusion, pairing a causal-masked Transformer encoder with a diffusion head, and generates continuous features such as price and time from four years of US equity data. The report frames it as a test of whether autoregressive diffusion can generate market data, and its title says the experiment exposed limits of continuous diffusion, but the available material lists no specific limits, evaluation metrics or generation results.

Written by AI from the coverage · updated 58 minutes ago

Not enough continuous observations to draw a trend yet.