State-space models for limit order books
Mamba and other selective state-space models are among the most frequent new architectures in the AI journals, built for long sequences. Limit order books produce exactly such sequences. Yet only 4 of 329 recent Mamba and state-space papers carry a finance keyword, all of them in AI journals. Can these models forecast short-term price moves better, and is the gain tradable after costs?
What prompted this
Since October 2025, "Mamba" appears on 293 papers, 290 of them in AI journals, and 329 papers carry a Mamba or state-space-model keyword; only 4 of them carry a finance, stock, trading, market or asset keyword. "Limit order book" appears on 13 papers and "high-frequency trading" on 21, mostly in finance journals.
Questions
- On standard limit order book data, do state-space models predict mid-price moves better than convolutional and Transformer models at the same computing cost?
- Do the gains survive realistic trading costs and latency?
- Do they capture order-flow patterns that last longer than other models can see?
Data that could help
Public limit order book benchmarks (for example the FI-2010 dataset or LOBSTER samples) and published baseline models.
Next step
Train a Mamba model on a public limit order book benchmark, and compare its forecasts and simulated trading results with published baselines.