Mim Mimusa Azim, Tuhin Md Kamrul Hasan, Nobi Ashadun
Department of Computer Science and Telecommunication Engineering, Noakhali Science and Technology University, Noakhali, Bangladesh.
Department of Computer Science and Engineering, Kishoreganj University, Kishoreganj, Bangladesh.
PLoS One. 2025 Jul 14;20(7):e0326947. doi: 10.1371/journal.pone.0326947. eCollection 2025.
In this study, we present a novel approach to analyzing financial crises of the global stock market by leveraging a modified Autoencoder model based on Recurrent Neural Network (RNN-AE). We analyze time series data from 24 global stock markets between 2007 and 2024, covering multiple financial crises, including the Global Financial Crisis (GFC), the European Sovereign Debt Crisis (ESD), and the COVID-19 pandemic. By training the RNN-AE with normalized stock returns, we derive correlations embedded in the model's weight matrices. To explore the network structure, we construct threshold networks based on the middle-layer weights for each year and examine key topological metrics, such as entropy, average clustering coefficient, and average shortest path length, providing new insights into the dynamic evolution of global stock market interconnections. Our method effectively captures the major financial crises. Our analysis indicates that interactions among American indices were significantly higher during the GFC in 2008 and the COVID-19 pandemic in 2020. In contrast, interactions among European indices were more prominent during the 2022 Russia-Ukraine conflict. In examining net inter-continental interactions, the influence was stronger between Europe and America during the GFC and the ESD crisis while, the influence between America and Asia was more powerful during the COVID-19 pandemic. Finally, we determine the structural entropy of the constructed networks, which effectively monitors the states of the market. Overall, our RNN-AE based network construction method provides valuable insights into market dynamic and uncovering financial crises, offering a powerful tool for investors and policymakers.
在本研究中，我们提出了一种新颖的方法，通过利用基于递归神经网络的改进自动编码器模型（RNN-AE）来分析全球股票市场的金融危机。我们分析了2007年至2024年期间来自24个全球股票市场的时间序列数据，涵盖了多次金融危机，包括全球金融危机（GFC）、欧洲主权债务危机（ESD）和新冠疫情。通过使用标准化股票回报训练RNN-AE，我们得出了模型权重矩阵中嵌入的相关性。为了探索网络结构，我们基于每年的中间层权重构建阈值网络，并检查关键的拓扑指标，如熵、平均聚类系数和平均最短路径长度，从而为全球股票市场互联的动态演变提供新的见解。我们的方法有效地捕捉了主要的金融危机。我们的分析表明，2008年全球金融危机和2020年新冠疫情期间，美国指数之间的相互作用显著更高。相比之下，2022年俄乌冲突期间，欧洲指数之间的相互作用更为突出。在研究洲际净相互作用时，全球金融危机和欧洲主权债务危机期间欧美之间的影响更强，而新冠疫情期间美亚之间的影响更大。最后，我们确定了所构建网络的结构熵，它有效地监测了市场状态。总体而言，我们基于RNN-AE的网络构建方法为市场动态和金融危机揭示提供了有价值的见解，为投资者和政策制定者提供了一个强大的工具。
