Attention-based reading, highlighting, and forecasting of the limit order book

Managing high-frequency data in a limit order book (LOB) is complex due to its high dimensionality, irregular timing, and complex spatiotemporal dependencies across price levels. These challenges often exceed the capabilities of conventional time-series models. Accurate prediction of the multi-level LOB, not just the mid-price, is crucial for understanding market dynamics but is difficult due to the interdependencies among attributes like order types, features, and levels. This study introduces advanced sequence-to-sequence models to forecast the entire multi-level LOB, including prices and volumes. Our key contribution is a compound multivariate embedding method that captures spatiotemporal relationships. Empirical results show that our method outperforms others, achieving the lowest forecasting error while maintaining LOB structure.