Deep Learning‐Based Network Relationship Construction Method and Its Impact on Futures Risk Premiums

This study proposes a deep learning framework, MIDAS–TGCN, to model the volatility spillover networks in futures markets and examines their impact on risk premiums. Traditional approaches typically rely on variance decomposition methods or VAR models, facing limitations in capturing high‐dimensional nonlinear dependencies and integrating macroeconomic factors. Our framework combines mixed‐data sampling (MIDAS) with Temporal Graph Convolutional Networks (TGCNs) to process high‐frequency market data and low‐frequency macroeconomic indicators through dual pathways, generating distinct short‐term (market‐driven) and long‐term (macro‐driven) volatility networks. The volatility network spillover effects derived through our proposed modeling framework not only capture structural responses to systemic events but also demonstrate enhanced robustness with reduced sensitivity to tail events compared with conventional approaches. Importantly, network spillover dynamics constructed via MIDAS–TGCN methodology exhibit significant explanatory power in decoding term structure risk premia in futures markets, which can be seen that the volatility network spillovers have asset pricing effects. This empirical validation aligns with emerging literature on high‐frequency risk transmission while extending the analytical frontier through temporal graph convolutional architectures.