arXiv q-fin· Valentin Mohl, Alexandre Bismuth, Aramis Marti-Shahandeh, George Nigmatulin, Kang Li, Stefan Zohren, Jakob Foerster, Mihai Cucuringu, Anisoara Calinescu·· 5 hr agoAI Score44
Understanding the Internal Mechanisms of State Space Foundation Models for Limit Order Books
Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books
AI summary
A study on the 8.2M-parameter Mamba-3 state space model shows that after autoregressive training on limit order book message streams of 8 large-cap US stocks, the model can learn general internal representations of market microstructure. The model achieves an R² of approximately 0.59 for order flow imbalance decoding, significantly higher than control groups and baselines. The representations can generalize to unseen stocks and exhibit causal controllability. This paper has been accepted by the 7th ACM International Conference on AI in Finance (ICAIF '26).
Source:arXiv q-fin · arxiv.org
Academic papers#Papers / Research#Machine Learning#High-Frequency / Market Microstructure#US Stocks#arXiv