Abstract
As financial markets grow increasingly complex in the big data era, accuratestock prediction has become more critical. Traditional time series models, suchas GRUs, have been widely used but often struggle to capture the intricatenonlinear dynamics of markets, particularly in the flexible selection andeffective utilization of key historical information. Recently, methods likeGraph Neural Networks and Reinforcement Learning have shown promise in stockprediction but require high data quality and quantity, and they tend to exhibitinstability when dealing with data sparsity and noise. Moreover, the trainingand inference processes for these models are typically complex andcomputationally expensive, limiting their broad deployment in practicalapplications. Existing approaches also generally struggle to captureunobservable latent market states effectively, such as market sentiment andexpectations, microstructural factors, and participant behavior patterns,leading to an inadequate understanding of market dynamics and subsequentlyimpact prediction accuracy. To address these challenges, this paper proposes astock prediction model, MCI-GRU, based on a multi-head cross-attentionmechanism and an improved GRU. First, we enhance the GRU model by replacing thereset gate with an attention mechanism, thereby increasing the model'sflexibility in selecting and utilizing historical information. Second, wedesign a multi-head cross-attention mechanism for learning unobservable latentmarket state representations, which are further enriched through interactionswith both temporal features and cross-sectional features. Finally, extensiveexperiments on four main stock markets show that the proposed methodoutperforms SOTA techniques across multiple metrics. Additionally, itssuccessful application in real-world fund management operations confirms itseffectiveness and practicality.