For years, generative AI has been split into two camps: the 'next-token' predictors (autoregressive models like GPT) and the 'noise-clearers' (diffusion models like Midjourney). But in the high-stakes world of quantitative finance, neither approach alone quite hits the mark. The real question is: what happens when you smash them together to simulate the chaos of the stock market?
The Identity Crisis of Market Data
To build a model, you first have to define what market data is. Is a price stream a sequence of words (text), or is it a fluid, continuous wave (video)? Jane Street has explored this exact dilemma. If you treat market data as a stream, you can use a diffusion model to generate the features of the next event, which is then autoregressively appended to the sequence. This hybrid approach attempts to capture both the granular noise of a tick-by-tick trade and the long-term trend of the market.
Breaking the Discrete-Continuous Wall
Traditionally, autoregressive models loved discrete tokens, while diffusion thrived in continuous spaces. However, new research—including work on Autoregressive Diffusion Models (ARDMs)—is blurring these lines. We're seeing a shift where these aren't separate families, but complementary capabilities. In finance, this means a model could potentially use diffusion to create a 'rough draft' of a market scenario and then use autoregressive logic to ensure the data remains coherent and mathematically sound over time.
The Future of Synthetic Finance
While we aren't yet at the point of perfectly simulating a flash crash, the convergence of these architectures is promising. By combining the iterative refinement of diffusion with the sequential logic of autoregression, researchers are moving closer to high-fidelity synthetic data that can be used to stress-test trading algorithms without risking real capital.
