# Understanding stock market instability via graph auto-encoders

**Authors:** Dragos Gorduza, Xiaowen Dong, Stefan Zohren  
**arXiv:** [2212.04974](https://arxiv.org/abs/2212.04974)  
**Date:** 2022-12-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.04974v1/

## Key points

- Understanding stock market instability is crucial in financial management
- Co-movements can be represented as a graph with nodes and edges
- Learning timely indicators of co-movement breakdowns is central to understanding financial stability and volatility forecasting
- Graph auto-encoder (GAE) edge reconstruction accuracy can serve as an indicator for how spatially homogeneous connections between assets are, which serves as a proxy to infer market volatility
- Higher GAE reconstruction error values are more correlated with higher volatility
- The approach has been tested on S&P 500 data from 2015-2022 with promising results
- Minute-level frequency data was used for six years of stock market prices for stocks in the S&P 500 index
- Log return matrices were calculated based on price movements and Pearson correlation matrices were derived based on rolling windows of length S days
- An adjacency matrix representing connected companies was constructed using a threshold of 0.7 for correlations
- Return volatility (RV) was used as a proxy for instability according to Eq. (2), which measures variance over time periods t+∆t: RVt+∆t/t = (r(t,n)2)
- The proposed methodology contributes to the literature of machine learning in finance, particularly in understanding stock market instability

## Summary

Understanding stock market instability is a crucial question in financial management, as it enables practitioners to forecast potential breakdowns in asset co-movements that could lead to rapid and devastating collapses in portfolio value. To better understand financial stability and volatility forecasting, the structure of co-movements can be represented as a graph, where companies are nodes and edges capture correlations between their price movements. Learning a timely indicator of co-movement breakdowns, manifested as modifications in the graph structure, is central to this understanding. To address this challenge, researchers have proposed using the edge reconstruction accuracy of a graph auto-encoder (GAE) as an indicator for how spatially homogeneous connections between assets are. Based on financial network literature, this measure serves as a proxy to infer market volatility. The higher the GAE reconstruction error values, the more correlated they are with higher volatility. This approach has been tested on S&P 500 data over the period from 2015-2022 and has shown promising results. The research team used minute-level frequency data for six years of stock market prices for stocks in the S&P 500 index. They calculated log return matrices based on these price movements and then derived Pearson correlation matrices based on rolling windows of length S days. These correlation matrices measured linear relationships between all stock pairs and were used to construct an adjacency matrix representing connected companies. The team defined an adjacency matrix A such that Auv = 1 if the (u,v) return pair had a correlation higher than 0.7; otherwise, Auv was set to 0. This threshold ensured only strongly connected pairs were kept while discounting correlations with insufficient strengths. Return volatility (RV) was used throughout subsequent sections as a proxy for instability according to Eq. (2), which measures variance over time periods t+∆t: RVt+∆t/t = (r(t,n)2). Volatility was measured as the natural logarithm of RV and termed log-RV due to the power law distribution of market volatility with orders of magnitude larger swings on high-volatility days. The proposed methodology contributes to the literature of machine learning in finance, particularly in understanding stock market instability.