主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (4): 47-62.doi: 10.16381/j.cnki.issn1003-207x.2024.0896cstr: 32146.14.j.cnki.issn1003-207x.2024.0896

• • 上一篇    下一篇

上市公司多层关系网络中混频信息交互与传播能够提升资产定价性能吗?——基于图神经网络的资产定价研究

王泽舟1, 许启发1,2(), 蒋翠侠1   

  1. 1.合肥工业大学管理学院,安徽 合肥 230009
    2.合肥工业大学过程优化与智能决策教育部重点实验室,安徽 合肥 230009
  • 收稿日期:2024-06-04 修回日期:2024-10-19 出版日期:2026-04-25 发布日期:2026-03-27
  • 通讯作者: 许启发 E-mail:xuqifa1975@163.com
  • 基金资助:
    国家自然科学基金面上项目(72171070);国家社会科学基金一般项目(21BJY255)

Can Interaction and Dissemination of Mixed-Frequency Information in Companies' Multi-Layer Relationship Networks Improve Asset Pricing? Asset Pricing Study Based on Graph Neural Networks

Zezhou Wang1, Qifa Xu1,2(), Cuixia Jiang1   

  1. 1.School of Management,Hefei University of Technology,Hefei 230009,China
    2.Key Laboratory of Process Optimization and Intelligent Decision-making,Ministry of Education,Hefei 230009,China
  • Received:2024-06-04 Revised:2024-10-19 Online:2026-04-25 Published:2026-03-27
  • Contact: Qifa Xu E-mail:xuqifa1975@163.com

摘要:

本文建立上市公司间供应链、股权和行业网络,融合低频公司特征和宏观经济信息以及高频市场、交易和情绪信息等作为网络节点特征,构建混频跨期注意力多层图卷积(MF-IAMGCN)模型,检验混频(MF)定价信息在上市公司多层关系网络中的同期交互和跨期传播对资产价格的预测能力。空间维度上,MF-IAMGCN模型基于注意力多层图卷积网络框架,引入混频数据抽样方法,刻画混频定价信息在多层网络中的同期交互,包括节点间的高阶依赖和非线性关系;时间维度上,借助门控机制,刻画连续时间步骤中节点之间的跨期信息传播和层间交互。选取2003年1月—2022年12月期间中国A股市场为研究样本,从个股定价、测试资产定价和投资组合绩效三个方面,实证检验模型性能。研究结果表明:(1)MF-IAMGCN模型在个股和测试资产层面的定价表现均优于四个竞争性模型,具备较强的综合定价性能。(2)MF-IAMGCN模型构造的投资组合取得了最优的风险调整绩效,兼顾高回报和低波动。(3)MF-IAMGCN模型刻画定价信息的动态传播模式,充分挖掘高频定价信号,显著提升了定价性能。(4)股权网络隐含金融市场资本活动信息,能够显著降低ST股票预测误差。

关键词: 资产定价, 多层关系网络, 同期交互, 跨期传播, 图注意力网络, 混频数据抽样

Abstract:

The interconnections among listed companies are intricate and constitute multi-layer relationship networks. The contemporary interaction and intertemporal dissemination of mixed-frequency information in multi-layer relationship networks will impact asset prices and returns. To this end, supply chain, equity, and industry networks among listed companies are established. Then low-frequency company characteristic information, low-frequency macroeconomic information, as well as high-frequency market, trade, and sentiment information are integrated as network node features. As a result, a novel MF-IAMGCN model is constructed to investigate how mixed-frequency information interaction and dissemination in multi-layer networks predict asset prices. In the spatial dimension, the mixed data sampling (MIDAS) method is incorporated into the attention multi-layer graph convolutional network (AMGCN) framework. The MF-IAMGCN model can capture the contemporary interaction of mixed-frequency information in multi-layer networks, including high-order dependencies and nonlinear relationships between nodes. In the temporal dimension, the MF-IAMGCN model leverages the gate mechanism to capture the intertemporal information dissemination between nodes in consecutive time steps. All listed companies in the Chinese A-share market from January 2003 to December 2022 are choosed and the proposed MF-IAMGCN model’s pricing power from individual stock pricing, test asset pricing, and portfolio performance is examined. The empirical results show that: (1) The pricing performance of the MF-IAMGCN model outperforms four competitive models on both individual stock and test asset levels. (2) The portfolios constructed by the MF-IAMGCN model achieve optimal risk-adjusted performance in terms of high returns and low volatility. (3) The mixed-frequency data processing module, network information contemporary interaction module, and network information intertemporal dissemination module within the MF-IAMGCN model jointly learn dynamic dissemination patterns of pricing information. The mixed-frequency data processing module contributes greatly to improving pricing performance by exploiting high-frequency pricing information. (4) Equity networks contain information about capital activities (such as strategic investments, mergers and acquisitions, insider trading, and “backdoor listings”), significantly reducing pricing errors of ST stocks and playing a vital role in asset pricing.

Key words: asset pricing, multi-layer relationship networks, contemporary interaction, intertemporal dissemination, graph attention network, mixed data sampling

中图分类号: