Spillover-based information set selection: evidence from oil spot price forecasting
提出一种基于溢出效应的信息集选择方法，通过系统内部互联性而非单个变量筛选来预测原油现货价格，发现包含全球市场指标的混合模型在中期和长期预测中表现更优。
Accurately identifying the determinants of oil spot prices remains a persistent challenge. This paper proposes a spillover-based approach to information-set selection, in which candidate system specifications are ranked based on their internal interconnectedness rather than via individual variable screening. Conceiving markets as dynamically evolving information networks, we implement the architecture of Total Spillover Index (TSI) to quantify the transmission of shocks across variables within candidate systems. We then construct, within a cointegration framework, alternative models representing Brent and WTI markets from both isolated and globally integrated perspectives. Spillover analysis shows that systems that incorporate global market indicators exhibit very strong interconnectedness and respond more sensitively to macroeconomic shocks, as reflected in their co-movement with the Global Economic Policy Uncertainty Index. Out-of-sample forecasts using both Fractional Cointegration Vector Autoregressive (FCVAR) models and Long Short-Term Memory networks show that a hybrid global specification consistently outperforms models that are restricted to isolated markets, particularly at medium and longer horizons. These results suggest that information coherence, capturing persistent cross-variable transmission within the system, provides a useful criterion for identifying forecasting-relevant information-sets in complex market environments such as global oil markets.