Scholarlink

Overnight stock movement prediction is vital for pre-market decision-making, enabling refined trading strategies and improved risk management. Traditional approaches often neglect sector-level interdependencies and fail to model the dynamic nature of stock networks, particularly under volatile market conditions. To overcome these limitations, we propose DGCN-DGRU, a novel framework that integrates Dynamic Graph Convolutional Networks (DGCN) with Dynamic Gated Recurrent Units (DGRU). DGCN captures sector-specific relationships and latent structural patterns, while DGRU fuses multi-source temporal data to enhance predictive accuracy and robustness. Extensive evaluation on real-world datasets from three countries demonstrates that our model outperforms existing methods, offering a robust and adaptable solution for overnight stock trend forecasting. This study highlights the effectiveness of combining dynamic graph learning with recurrent architectures in financial time series prediction.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Overnight Stock Movement Prediction Based on Dynamic Graph Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Siyu Sun,
  • Xinyi Zhou

摘要

Overnight stock movement prediction is vital for pre-market decision-making, enabling refined trading strategies and improved risk management. Traditional approaches often neglect sector-level interdependencies and fail to model the dynamic nature of stock networks, particularly under volatile market conditions. To overcome these limitations, we propose DGCN-DGRU, a novel framework that integrates Dynamic Graph Convolutional Networks (DGCN) with Dynamic Gated Recurrent Units (DGRU). DGCN captures sector-specific relationships and latent structural patterns, while DGRU fuses multi-source temporal data to enhance predictive accuracy and robustness. Extensive evaluation on real-world datasets from three countries demonstrates that our model outperforms existing methods, offering a robust and adaptable solution for overnight stock trend forecasting. This study highlights the effectiveness of combining dynamic graph learning with recurrent architectures in financial time series prediction.