Ascertaining price formation in financial markets with machine learning: Evidence from Chinese stocks

This paper examines stock price prediction in the Chinese market using deep learning techniques applied to tick-level high-frequency data. We demonstrate that the high-frequency features constructed in this study capture significantly more predictive information than raw price and volume inputs, leading to measurable improvements in model performance. Beyond accuracy, we provide empirical evidence of three structural characteristics embedded in order flow data: nonlinearity, intertemporal stability, and path dependence properties that help explain the dynamics of the price formation process. Furthermore, we show that a universal model trained on data from multiple stocks consistently outperforms stock-specific models in out-of-sample prediction tasks. These findings offer both practical insights for developing robust trading strategies and theoretical implications for understanding the microstructural mechanisms of financial markets.