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

中国管理科学 ›› 2025, Vol. 33 ›› Issue (8): 1-13.doi: 10.16381/j.cnki.issn1003-207x.2024.1601

• •    

基于Transformer-LSTM分位数回归的全球股市极端风险溢出研究

姚银红, 王晓旭, 陈炜, 陈振松()   

  1. 首都经济贸易大学管理工程学院，北京 100070
  • 收稿日期:2024-09-13 修回日期:2024-10-22 出版日期:2025-08-25 发布日期:2025-09-10
  • 通讯作者: 陈振松 E-mail:chenzhensong@cueb.edu.cn
  • 基金资助:
    国家自然科学基金项目(72101166);国家自然科学基金项目(72071134);教育部人文社会科学研究青年基金项目(24YJCZH032);北京市教育委员会科技计划项目(KM202210038001);北京市属高校高水平科研创新团队建设支持计划项目(BPHR20220120)

Extreme Risk Spillover among Global Stock Markets Based on Transformer-LSTM Quantile Regression

Yinhong Yao, Xiaoxu Wang, Wei Chen, Zhensong Chen()   

  1. School of Management and Engineering，Capital University of Economics and Business，Beijing 100070，China
  • Received:2024-09-13 Revised:2024-10-22 Online:2025-08-25 Published:2025-09-10
  • Contact: Zhensong Chen E-mail:chenzhensong@cueb.edu.cn

RichHTML

0

PDF (PC)

24

摘要：

全球经济不确定性的增加和极端事件的频发使得精确测度全球股市极端风险溢出效应成为应对跨国金融冲击的重要途径。现有研究在综合考虑时间序列非线性、长期依赖性和多变量交互影响方面存在一定的局限性。因此，本文提出Transformer-LSTM分位数回归模型，在提取数据时序特征的同时，利用Transformer中的多头注意力机制并行处理多个注意力函数，以更精确地捕捉全球股市极端风险的时间演变特征，并构建溢出网络测度全样本和金融危机等危机事件发生期间的风险溢出效应。基于2001年12月—2024年3月共19个国家周度股指数据的实证结果表明：（1）本文提出的模型相较于多层感知机（MLP）、长短期记忆网络（LSTM）以及Transformer具有更强的预测能力。（2）各国股市在全样本期间的溢出效应具有非对称性，其中，美国股市存在显著的风险溢出效应，中国股市没有明显的风险溢出和接收效应。（3）危机事件发生时，极端风险溢出效应增加且非对称性增强。金融危机期间，美国股市风险溢出效应显著，多国存在显著的双向溢出效应；欧债危机期间，风险溢出效应主要集中在欧洲国家的股票市场；中美贸易摩擦期间，美国股市对中国股市的风险冲击明显增强；新冠疫情期间，美国、英国等发达国家股市仍是主要的风险溢出源。本文提出的模型为金融市场的极端风险溢出研究提供了新思路，研究结果对危机时期的全球股市风险管理具有重要的参考借鉴价值。

关键词: 全球股市, 极端风险溢出, CoVaR, Transformer-LSTM, 分位数回归

Abstract:

The increasing global economic uncertainty and the frequent occurrence of extreme events have made the precise measurement of extreme risk spillover effects in global stock markets a crucial approach for addressing cross-border financial shocks. Existing studies exhibit certain limitations in comprehensively considering the nonlinearities， long-term dependencies， and multivariable interactive effects of time series. Therefore， a Transformer-LSTM quantile regression model is proposed that leverages the multi-head attention mechanism in the Transformer to process multiple attention mechanisms in parallel， while extracting the temporal characteristics of the data. This approach aims to more accurately capture the temporal evolution of extreme risks in global stock markets and examine risk spillover effects during the full sample period and crisis periods such as a financial crisis through constructing spillover networks. Based on empirical results from weekly stock index data of 19 countries from December 2001 to March 2024， the findings are as follows： （i） The proposed model demonstrates superior predictive power compared to the Multilayer Perceptron （MLP）， Long Short-Term Memory （LSTM） network， and Transformer models. （ii） The spillover effects in cross-country stock markets exhibit asymmetry over the full sample period. Notably， there is a significant risk spillover effect in the U.S. stock market， while Chinese stock market shows no obvious risk spillover or receiving effects. （iii） During crisis events， extreme risk spillovers increase and asymmetry intensifies. During the financial crisis， the risk spillover effects from the U.S. are significant， with notable bidirectional spillover across multiple countries’ stock markets. During the European debt crisis， risk spillover effects are primarily concentrated in European countries’ stock markets. The risk impact from the U.S. stock market on China notably strengthens during the Sino-US trade friction. During the COVID-19 pandemic， stock markets of developed countries such as the U.S. and the U.K. remain the main sources of risk spillover. The proposed model offers new insights into capturing the extreme risk spillover in financial markets， which is important for risk management in global stock markets during times of crisis.

Key words: global stock markets, extreme risk spillover, CoVaR, Transformer-LSTM, quantile regression, spillover network

中图分类号: