This paper investigates the predictability of high-frequency limit order book (LOB) mid-price changes using deep learning, focusing on the link between market microstructure and forecasting performance. The authors analyze 15 heterogeneous stocks from the NASDAQ exchange over three years (2017-2019), employing the DeepLOB model to predict the direction of mid-price movements at horizons of 10, 50, and 100 LOB events. They introduce LOBFrame, an open-source codebase for efficient LOB data processing and model evaluation, addressing reproducibility issues in the field.

The study reveals that stocks' microstructural properties, such as tick size, spread, and liquidity, strongly influence the effectiveness of deep learning forecasts. However, high predictive accuracy does not guarantee actionable trading signals, highlighting a simulation-to-reality gap. The authors argue that traditional machine learning metrics fail to capture the practical utility of forecasts and propose an alternative operational framework that evaluates predictions based on the probability of correctly forecasting complete transactions. This work provides a comprehensive guide for academics and practitioners on the scope and limitations of deep learning in LOB forecasting, emphasizing the importance of microstructural priors and realistic performance assessment.