This paper addresses the overlooked issue of limit order book (LOB) data representation for deep learning models. The authors identify that the commonly used compressed representation, which organizes LOB data by price levels, violates key assumptions of neural networks such as spatial homogeneity and source consistency across input channels. This leads to fragility under adversarial perturbations, where minimal market orders cause drastic shifts in the representation. To overcome these issues, the paper proposes two alternative representations: a moving-window representation that encodes volumes in a fixed grid centered at the mid-price, and a market-depth representation that accumulates volumes for improved spatial smoothness. Both representations are designed to be compatible with convolutional and recurrent architectures.

Through experiments on the FI-2010 dataset, the authors demonstrate that the proposed representations significantly enhance model robustness and accuracy. Under adversarial perturbations simulating small limit orders, models using compressed representations suffer substantial performance drops (e.g., DeepLOB accuracy falls from 77.30% to 69.02%), while models using moving-window or market-depth representations maintain stable performance. Moreover, even simple models like LSTM with market-depth representation achieve state-of-the-art results (77.46% accuracy) without complex architectures. The work emphasizes that data representation is as critical as model design in quantitative finance, and provides practical guidelines for constructing robust LOB representations that align with deep learning principles.