Analyzing crises in global financial indices using Recurrent Neural Network based Autoencoder

Area of Science:

• Quantitative Finance
• Network Science
• Machine Learning

Background:

• Global stock markets exhibit complex interdependencies that evolve during financial crises.
• Understanding these dynamics is crucial for investors and policymakers to navigate economic volatility.

Purpose of the Study:

• To develop and apply a novel Recurrent Neural Network Autoencoder (RNN-AE) model for analyzing global stock market interconnections.
• To identify and characterize network structures and topological metrics during major financial crises from 2007-2024.

Main Methods:

• Utilized time series data from 24 global stock markets (2007-2024).
• Employed a modified RNN-AE to derive correlations from normalized stock returns.
• Constructed threshold networks using middle-layer weights and analyzed topological metrics (entropy, clustering coefficient, shortest path length).

Main Results:

• The RNN-AE model successfully captured major financial crises, including the Global Financial Crisis (GFC), European Sovereign Debt Crisis (ESD), and COVID-19 pandemic.
• Identified increased interactions among American indices during the GFC and COVID-19, and European indices during the Russia-Ukraine conflict.
• Revealed distinct inter-continental interaction patterns: Europe-America during GFC/ESD, and America-Asia during COVID-19.

Conclusions:

• The RNN-AE based network construction method offers valuable insights into market dynamics and financial crisis detection.
• Structural entropy effectively monitors market states, providing a tool for investors and policymakers.
• The study highlights the evolving nature of global stock market interconnections during crises.