Predicting Shock Propagation and Uncovering Heterogeneity with Graph Neural Networks

Recent research has made substantial progress in studying shock propagation through inter-firm transaction networks, and empirical studies have directly documented firm-level shock propagation. Despite these advances in both theory and empirics, no method has yet been established to accurately predict the effects of large-scale future shocks, such as natural disasters, financial crises, or pandemics. A central challenge is the heterogeneity inherent in firms and transaction relationships, which makes it difficult to identify which firms are important for shock propagation and which links amplify it. To address this issue, this study uses firm-level data and a graph neural network (GNN) to predict firm growth rates with a model that explicitly incorporates network structure. In particular, by analyzing the trained GNN model, we quantitatively identify the firms and transaction links that are important for shock propagation. Using the global financial crisis, specifically the sharp decline in exports, as a case study, we show that incorporating network structure significantly improves predictive performance and enables us to identify specific firms and links that are important for propagation.