Long short-term memory autoencoder based network of financial indices

We present a novel approach for analyzing financial time series data using a Long Short-Term Memory Autoencoder (LSTMAE), a deep learning method. Our primary objective is to uncover intricate relationships among different stock indices, leading to the extraction of stock networks. We examine time series data spanning from 2000 to 2022, encompassing multiple financial crises within the S&P 500 stock indices. By training a modified LSTMAE with normalized stock index returns, we extract the inherent correlations embedded in the model weights. We create directional threshold networks by applying a fixed threshold, calculated as the sum of the mean and standard deviation of matrices from various years. Our investigation explores the topological characteristics of these threshold networks across different years. Notably, the observed network properties exhibit unique responses to the various financial crises that occurred between 2000 and 2022. Furthermore, our sector analysis reveals substantial sectoral influences during times of crisis. For example, during global financial crises, the financial sector assumes a prominent role, exerting significant influence on other sectors, particularly during the European Sovereign Debt (ESD) crisis. During the COVID-19 pandemic, the health care and consumer discretionary sectors are predominantly impacted by other sectors. Our proposed method effectively captures the underlying network structure of financial markets and is validated by a comprehensive analysis of network metrics, demonstrating its ability to identify significant financial crises over time.