Dynamic Connectedness and Spillover-Based Machine Learning for Energy-Market Risk Identification: Evidence from U.S. Energy Markets

Cross-market risk transmission in U.S. energy markets has become increasingly complex as fossil fuel prices, electricity markets, and clean energy financial exposure respond differently to stress episodes. Identifying whether dynamic spillover information contains forward-looking diagnostic value is therefore important for energy market risk monitoring. This study examines a daily six-market U.S. energy return panel covering WTI crude oil, Henry Hub natural gas, Brent crude oil, RBOB gasoline, PJM West electricity, and CELS clean-energy equity exposure from 2016 to 2025. We first estimate time-varying total, directional, and net connectedness using a TVP-VAR-DY framework and then transform the resulting connectedness measures into spillover-based features for supervised high-DSV20-state classification. The results show that energy-market connectedness is clearly time-varying, with crude oil benchmarks occupying central positions and market-level net spillover roles changing across market conditions. Under the retained label-80 Random Forest specification, connectedness-based features provide moderate diagnostic value for identifying future high-DSV20 states. Net WTI, Net Henry Hub, and Net CELS are the most informative spillover-role variables. Additional validation checks indicate that the evidence is best interpreted as support for diagnostic risk monitoring rather than as a high-accuracy forecasting system. The findings highlight the usefulness of dynamic connectedness measures as transparent inputs for energy-market risk assessment.