This paper provides a comprehensive survey of deep learning techniques applied to limit order book (LOB) data, which is fundamental in high-frequency trading. The authors systematically review various LOB data representations, including raw order flow, aggregated features, and 2D spatial structures, and map them to appropriate deep learning architectures such as CNNs, RNNs, attention mechanisms, and generative models. They also present an extensive empirical comparison of state-of-the-art models on a unified benchmark dataset, evaluating performance on tasks like mid-price movement prediction and volatility forecasting.

The results demonstrate that deep learning models consistently outperform traditional econometric approaches, with attention-based models excelling in long-horizon predictions and CNNs performing best with spatial representations. The study also highlights the importance of data representation and the potential of generative models for data augmentation. The paper concludes by identifying open challenges, including interpretability, computational efficiency, and the need for standardized benchmarks, and outlines promising future research directions such as reinforcement learning for trading and multi-agent modeling.