(with Jonathan Brogaard)

Credit Default Swap (CDS) spreads exhibit network effects. This paper employs Graph Neural Networks (GNNs) to predict firm CDS spreads by incorporating inter-firm network effects. GNNs treat firms as nodes and idiosyncratic volatility spillover effects as directed edges, effectively capturing the dynamics of inter-firm contagion. Out-of-sample predictions show that GNNs improve prediction accuracy by over 50% compared to classic machine learning algorithms that cannot incorporate inter-firm edge characteristics. We further enhance GNN with node- and edge-attention layers to elucidate its mechanism. These attention layers highlight the importance of specific nodes, including manufacturing and intermediary firms, and crucial edges, such as those between intermediary, retail trade, or information firms and others, in predicting CDS spreads. Our findings enrich CDS pricing models with insightful financial networks and advanced machine learning methodologies.

(with Tim Johnson)

This paper investigates the amplification of idiosyncratic shocks into macroeconomic tail risk under the dynamic evolution of input-output networks. We demonstrate that "normal" microeconomic shocks, without "tail" occurrences at the firm level, can be amplified into macroeconomic tail risk through dynamic networks. The dynamic changes in networks disrupt the generalized central limit theorem, inhibiting the rapid decay of aggregate volatility and fostering the clustering of large macroeconomic tails. This study underscores the critical role of input substitution elasticity. Firms' inability to quickly substitute inputs, particularly when all inputs are complementary, enhances the network effect, escalating macroeconomic tail risks.

(with Mahyar Kargar, Sébastien Plante, John Shim, and Karamfil Todorov)