Deep Learning Architectures for High-Frequency Stock Price Prediction: An Empirical Evaluation

Abstract

High-frequency trading (HFT) generates vast amounts of financial data at millisecond intervals, presenting both opportunities and challenges for accurate stock price prediction. Traditional econometric and statistical models, while useful for low-frequency data, often fail to capture the nonlinear dependencies, temporal correlations, and microstructural patterns inherent in high-frequency financial markets. In this study, we present an empirical evaluation of multiple deep learning architectures including Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), Temporal Convolutional Networks (TCNs), and Transformer-based models for high-frequency stock price forecasting. Using a dataset of tick-level order book information and intraday price movements from major U.S. exchanges, we assess each model’s predictive power, robustness, and computational efficiency. Evaluation metrics include root mean squared error (RMSE), mean absolute percentage error (MAPE), and directional accuracy, alongside financial performance measures such as cumulative returns and Sharpe ratio in a simulated trading environment. Results reveal that attention-based architectures, particularly Transformers, consistently outperform recurrent and convolutional counterparts in capturing complex temporal dependencies, while TCNs demonstrate superior efficiency in low-latency scenarios. The findings highlight the trade-offs between accuracy, interpretability, and latency, providing actionable insights for practitioners in algorithmic trading, risk management, and market microstructure research.