On Capturing Multi-Scale Market Dynamics for High-Frequency Stock Price Forecasting Using a Hybrid Attention-Based Deep Learning Model

Abstract

Forecasting high-frequency stock prices is a significant challenge due to inherent noise, non-stationarity, and complex market dynamics. Conventional models often struggle to effectively extract meaningful signals from raw price and volume data. To address this, we propose an innovative hybrid framework, EWOA-VMD-ATT-BiGRU, which introduces several key novelties to enhance prediction accuracy. First, our model uniquely decomposes both price and volume sequences as an innovative feature engineering, which allows for the unveiling of multi-scale market characteristics and effectively mitigates signal interference. Second, we employ an enhanced whale optimization algorithm (EWOA) to adaptively optimize the variational mode decomposition (VMD) parameters, ensuring a more precise and data-driven signal separation. Finally, a Bidirectional GRU network integrated with an Attention mechanism (ATT-BiGRU) is utilized to dynamically weigh the importance of the decomposed features for superior prediction. Empirical results on the high-frequency SSE index demonstrate our model's superior performance. Compared to the second-ranked model, it achieves reductions in MSE, RMSE, MAE, MSLE, MAPE, and SMAPE by 6.23%, 3.17%, 3.22%, 5.87%, 3.16%, and 3.15%, respectively. Notably, our strategy of decomposing both price and volume yields a substantial improvement, reducing key error metrics by over 40% compared to an equivalent non-decomposed model. Our proposed decomposition-based hybrid model offers a more precise and robust approach for forecasting of high-frequency stock prices.