The talk follows the development of our research on forecasting limit order book dynamics with machine learning, from dense deep networks to architectures built on higher-order network structure.  It covers both established results and more recent, preliminary ones.

Early comparative studies on NASDAQ data showed that deep AI models learn mostly from the cross-sectional structure of the book, with performance varying strongly across stocks.  This led to the design of Homological Neural Networks, which are wired by Information Filtering Networks.  Their order-book version, HLOB, outperforms nine state-of-the-art AI models, including Transformers, on 15 NASDAQ stocks over three years.  More recent work folds time into the network and constrains Transformer attention with the same structure.  In controlled benchmarks, these models match HLOB on the order book and outperform the TabPFN foundation model when data are scarce.

Background reading

Briola, Turiel, Aste, Deep learning modeling of limit order book: a comparative perspective, arXiv:2007.07319 (2020) Briola, Bartolucci, Aste, Deep limit order book forecasting, arXiv:2403.09267 (2024); code: github.com/FinancialComputingUCL/LOBFrame
Briola, Bartolucci, Aste, HLOB: Information persistence and structure in limit order books, Expert Systems with Applications 266 (2025) 126078 Briola, Wang, Bartolucci, Aste, Homological convolutional neural networks, arXiv:2308.13816 (2023) Lin, Briola, Wang, Aste, Compositional sparsity as an inductive bias for neural architecture design, arXiv:2605.14764 (2026)

Zoom Meeting link:  texastech.zoom.us/j/3067000354

Meeting ID:  306 700 0354

Passcode:      TTUMF