Graph neural networks model systemic financial risk and volatility forecasting in banking systems and cryptocurrency markets, suggesting time-varying dynamics are essential.
Key Points
This research aims to forecast systemic financial risk by modeling cross-market contagion between banking systems and cryptocurrency markets.
Utilized a dynamic graph neural network framework for modeling financial networks.
Nodes represent financial entities, and edges are constructed with rolling-window dependency measures.
Implemented stress regimes to assess performance under market turmoil.
Conducted experiments including static-graph ablations and cross-market removal to analyze contributions of network dynamics.
Dynamic GNNs significantly outperform static-graph models during stress periods.
A strong LSTM baseline showed superior overall volatility forecasting accuracy.
Cryptocurrency assets add limited incremental information for bank-specific forecasts but are useful for system-level characterizations.