Multisource Information Fusion as a Decision Support System: Optimizing Deep Learning Ensembles for Financial Risk in Crude Oil Forecasting

Extreme volatility and geopolitical shocks in crude oil markets compromise strategic planning in capital-intensive energy projects. This study aims to forecast crude oil prices under severe uncertainty and to develop a volatility-grounded financial risk index that converts predictive accuracy into actionable decision support. To this end, we propose a Multisource Information Fusion Framework that couples daily Brent, WTI, and OPEC benchmarks with the Geopolitical Risk (GPR) Index, six deep learning architectures tuned by the Tree-structured Parzen Estimator, and an adaptive ensemble strategy (EnsResidChamp) that frames model aggregation as a dynamic optimization problem over forecast residuals. Fusing the GPR signal reduced errors in most architectures, and EnsResidChamp attained SMAPE values of 2.36% (Brent), 2.45% (WTI), and 2.27% (OPEC), outperforming univariate baselines and standalone models with significance confirmed by Friedman and Nemenyi tests. Value-at-Risk backtesting at the 99\% confidence level yielded violation rates of 0.92%-1.38% with no exception clustering (Kupiec and Christoffersen, p > 0.05), and the proposed Financial Risk Index (FRI) was minimized by EnsResidChamp for Brent (2.72) and WTI (2.84), whereas for the OPEC basket, a simpler ensemble achieved the lowest FRI (2.32) due to a more favorable Expected Shortfall-to-VaR ratio. We conclude that combining geopolitical fusion with residual-based ensembling delivers not only sharper forecasts, but a Decision Support System whose tail-risk behavior is statistically validated for hedging and capacity planning.