Graph neural networks for financial contagion
Financial contagion and systemic risk are rising in the finance and economics journals, and graph neural networks are among the largest topics in the AI journals. The two have met in only five papers. Can a graph model learn the channels of contagion that network models fix in advance?
What prompted this#
"Financial contagion" is among the fastest-growing keywords across all journals on the trends page (31 papers in July–September 2026 against 16 expected), and "systemic risk" is growing in economics. Since October 2025, 392 papers carry a contagion or systemic-risk keyword; 66 of them also carry a network keyword, but only 5 carry a graph-neural-network keyword. Over the same period "graph neural networks" appears on 578 papers, 555 of them in AI journals, and graph convolution is near the top of the equation ranking.
Questions#
- Given exposure networks between banks or firms, does a graph neural network forecast joint tail losses (CoVaR, marginal expected shortfall) better than cascade models with fixed rules?
- Which links carry stress: lending, common holdings, supply chains or news co-mentions?
- Can the model's attributions tell supervisors which institutions transmit stress, and are they stable enough to act on?
Data that could help#
- Common-holdings networks from 13F filings, supply-chain links, and bank exposures where supervisors publish them.
- Daily returns for tail-risk measures such as CoVaR.
Next step#
Rebuild one published common-holdings contagion network from 13F data and compare a simple graph convolution model with a CoVaR benchmark.