Interpretable machine learning unveils nonlinear drivers of global energy risk spillovers: A TVP-VAR approach

This study examines global energy risk spillovers using a time-varying parameter vector autoregression (TVP-VAR) model and interpretable machine learning. Unlike previous studies constrained by single-factor analyses and linear assumptions, we resolve three key limitations: capturing multidimensional drivers, addressing multicollinearity, and modeling nonlinear dynamics. Our findings reveal that spillovers fluctuate temporally, driven by long-term components, with energy-rich and rapidly transforming economies as primary transmitters. Machine learning models outperform linear regression, identifying critical nonlinear interactions among economic development, energy structure, and balance of payments. Regional heterogeneity is pronounced: Europe and the U.S. prioritize economic growth, China focuses on capital flows, while Japan and Israel emphasize oil imports. By integrating interpretable ML with TVP-VAR, this study advances systemic risk analysis and provides policymakers with actionable, region-specific strategies for energy market stability.