Improving System Generalisation & Forecasting via Spillover-Based Variable Selection Approach - Evidence From the Energy Market.

This paper develops and validates a spillover-based variable selection approach to improve the forecast precision of oil market spot prices. Leveraging the Total Spillover Index (TSI) as a quantitative measure of dynamic interdependencies, we construct six models within a cointegration framework to capture the behaviour of Brent and West Texas Intermediate (WTI) markets from both isolated and global perspectives. A dynamic spillover analysis encompassing static, frequency-domain, and rolling window techniques shows that models incorporating global market indicators exhibit higher TSI values and are more sensitive to macroeconomic shocks, as evidenced by their comovement with the Global Economic Policy Uncertainty (GEPU) Index. Out-of-sample forecasting experiments using both Fractional Cointegration Vector Autoregressive (FCVAR) framework and Long Short-Term Memory (LSTM) networks demonstrate that the hybrid global model, which integrates Brent and WTI spot prices with global market variables, consistently outperforms isolated market models, particularly over medium and long term horizons. These findings highlights the importance of incorporating global dynamics in variable selection, thus enhancing the endogenous explanatory power of forecasting models in complex, interdependent markets.