Probabilistic Limit Order Book Forecasting with Conditional Diffusion Models
The study introduces LOBGrad, a conditional DDPM framework combining an autoregressive RNN encoder with a diffusion generator for probabilistic limit order book forecasting, establishes an excess-risk oracle inequality for the one-step DDPM noise-prediction objective, and shows improvements over an encoder-free ablation on BTC/USDT, GOOGL, and META.
01 ABSTRACT
Fact: The authors propose LOBGrad, which couples an RNN encoder with a diffusion generator, and prove an excess-risk oracle inequality for the one-step DDPM noise-prediction objective under exponential mixing and regularity assumptions. Experiments on three datasets show LOBGrad improves over its encoder-free diffusion ablation on all reported queue metrics and is competitive with other baselines. Author view: Under contractive dynamics, the dependence on conditioning-window length can be reduced to dependence on latent-state dimension, and the generated forecasts show potential value in a downstream task.
02 KEY FINDINGS
- LOBGrad combines an autoregressive RNN encoder with a conditional DDPM for probabilistic limit order book forecasting.
- An excess-risk oracle inequality is established for the one-step DDPM noise-prediction objective under exponential mixing and regularity assumptions.
- On BTC/USDT, GOOGL, and META, it improves over the encoder-free ablation on all reported queue metrics and reproduces stylized facts of LOB microstructure.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG · ARXIV Q-FIN