Autoregressive Diffusion for Market Data Generation

Quantitative finance typically relies on models that predict future prices based on market data events. Generative models, which can synthesize detailed event sequences, are less common but could provide richer predictions. Market data exhibits both continuous and discrete characteristics, making it challenging to model effectively. This project explored using autoregressive diffusion models to generate market data, inspired by techniques from image generation. The researcher, Kavish, built a diffusion model to handle the unique properties of market data, including discrete event types and continuous features like price and timing. The study found that treating market data as fully continuous led to significant issues, particularly with the divergence of denoising trajectories. Flow matching proved to be a more stable alternative to DDPM. The research also highlighted the need to address discontinuities in market data, such as clustered order timings and price changes, through categorical modeling and targeted diffusion targets.