Research Explores Autoregressive Diffusion Models for Market Data Generation

Generative Models for Market Data

Quantitative finance typically uses models that predict future prices based on market data events like order additions, cancellations, and executions. This research explored generative models that could synthesize entire order book events, including their timing, offering more detailed market simulations than point estimates.

Challenges in Modeling Market Data

Market data presents challenges for generative modeling due to its hybrid nature. While order book actions are discrete, parameters like price can appear continuous due to high cardinality, and timing is a continuous variable. However, real-world distributions are spiky, with phenomena like "pennying" and bursts of orders, making a purely continuous approach difficult.

Autoregressive Diffusion Model Implementation

A research intern developed an event-level generative model for market data using autoregressive diffusion, drawing inspiration from autoregressive image generation techniques. This approach aimed to represent sequential, multimodal data similar to its application in image, video, and audio domains.

Research Findings and Limitations

The research found that DDPM, a common diffusion model technique, diverged in this application, while flow matching performed better. The primary conclusion was that fully continuous diffusion models were not effective for capturing the jagged and discrete characteristics of actual market data.