This paper presents a comprehensive comparative study of deep learning models for Limit Order Book (LOB) modeling in high-frequency trading. The research systematically evaluates multiple state-of-the-art architectures including Random models, Logistic Regressions, LSTMs, Attention-equipped LSTMs, CNN-LSTMs, and MLPs under identical experimental conditions. By using the same tasks, feature space, and dataset for all models, the study ensures fair comparison and meaningful insights into model performance and characteristics.

The key finding reveals that simpler Multilayer Perceptron (MLP) architectures perform comparably to or even better than more complex CNN-LSTM models, suggesting that dynamic spatial and temporal dimensions—while providing a good approximation—may not represent the true underlying dimensions of LOB dynamics. Through clustering analysis based on pairwise similarity and performance metrics, the paper investigates the intrinsic dimensions of modeling techniques and their relationship to LOB dynamics. This work contributes to both theoretical understanding of LOB modeling and practical guidance for model selection in high-frequency trading applications.