Score-based diffusion models typically perturb data with Gaussian noise, corresponding to a light-tailed forward stochastic differential equation. According to a new arXiv paper, this assumption can be a poor abstraction for market time series, particularly limit order book data, which exhibit extreme fluctuations and heavy tails.
The paper introduces StochLOB, a method that replaces the standard Gaussian perturbation with heavy-tailed score matching. By aligning the diffusion process with the actual statistical properties of limit order book data, StochLOB aims to improve the robustness and accuracy of trend prediction in high-frequency trading contexts.
The abstract does not provide experimental details or quantitative results, so the claimed improvements are not yet verified beyond the methodological proposal. The work highlights a growing recognition that standard diffusion assumptions may need adjustment for financial time series.