Knowledge Graphs of Bank Volatility in the COVID-19 Crisis

This study applies the knowledge graph framework of Chen and Zhang (2024), From Liquidity Risk to Systemic Risk: A Use of Knowledge Graph, to 70 banks listed in the United States over the COVID-19 episode.

Findings

At a cut-off $c=0.035$, we found that none of the 25 largest banks are isolated during the crisis, and the knowledge graph holds a dense core. This aligns with the expectation that firms tend to become more interconnected during crisis periods when common macroeconomic shocks overshadow idiosyncratic risks. Nevertheless, this is a claim about weak links only. A large share of strong conditional dependence in this sample occurs in the non-crisis period. At $c = 0.050$, there are 8 edges inside the crisis against 84 outside it, and the surviving structure matches the geographic split of the non-crisis graph. This matches the volatility result of Chen and Zhang (2024), which they describe as counterintuitive relative to their liquidity result.

Layout

data/     committed input prices
src/      analysis modules, every parameter declared once in config.py
report.pdf

Data

data/bank_prices.csv is the 70 bank stocks daily close from Yahoo Finance, from May 2017 to December 2021. The universe is the ranking by market capitalization published at stockanalysis.com.

Reproducing

Python 3.13.

pip install -r requirements.txt
python src/plots.py

This prints every number quoted in the report and rewrites the three figures it uses to latex/figures/, which is created on the first run and is not tracked.