EconoGNN: A graph neural network framework for temporal economic resilience insights

Abstract

Global economic shocks such as the 2008 financial crisis or recent trade escalations between the United States and China have exposed the complexity of interdependent economies and the need for systemic, multi-agent analysis. However, most regional economic resilience (RER) studies remain limited by localized datasets, inconsistent definitions, and static modeling approaches, restricting their ability to generalize insights across space and time. We introduce EconoGNN, a Graph Neural Network framework that integrates complexity theory, economic modeling, and machine learning to predict and explain regional economic resilience across 183 countries over 25 years. By combining over 81 million trade records (UN COMTRADE) and 500,000 macroeconomic observations (Penn World Table), and adopting an official resilience metric from the World Bank, our approach enables reproducible and interpretable global-scale analysis. EconoGNN achieves F1-scores of 0.750 with the temporal GNN architecture GConvGRU, AUC-ROC of 0.792, and PR-AUC of 0.757, demonstrating robust performance across different recovery threshold settings (τ = 0.90–1.00) with F1-scores ranging from 0.730 to 0.771, and yielding statistically significant improvements (p-value ≤0.05) over baselines. GNNExplainer validation confirms explanation reliability (Fidelity+ = 0.827, Characterization = 0.913), enabling country-customized interpretability of resilience drivers. Moreover the EconoGNN framework integrates key structural and welfare indicators to model both national and cross-border economic interactions, reducing omitted-variable bias and implicitly accounting for political, institutional, and cultural differences.