A Combined Framework Based on Feature Selection and Multivariate Mixed-Frequency for Crude Oil Prices Point and Interval Forecasting

Effective crude oil price forecasting is essential for energy supply stabilization, investment decisions, policy formulation, and economic impact assessment. However, previous studies of crude oil futures price forecasting either consider only a single variable without adequately accounting for the influence of various factors on crude oil prices, or they consider only the same frequency of influences and ignore the importance of mixed frequencies, resulting in forecasts that do not achieve the desired results. To overcome these problems, this paper proposes a new multivariate combinatorial mixed frequency forecasting system to predict the weekly closing price of crude oil futures. The system is divided into four modules: Data denoising, Feature selection, Combined forecasting, and Performance evaluation module. To obtain smooth data, ICEEMDAN is used to denoise the original data. Furthermore, to select appropriate variables and reduce model redundancy, recursive feature elimination is used to select appropriate variables with low frequency. Considering the importance of mixed-frequency data, the mutual information method was used to select appropriate high-frequency variables for modeling the crude oil price forecast model. To overcome the shortcomings of Back Propagation Neural Network, Gate Recurrent Unit, and Radial Basis Function Neural Network models, integrate their advantages, and obtain accurate and stable prediction results, a combined forecasting mechanism based on a Multi-Objective Sparrow Search algorithm was developed to obtain both point and interval forecasting results, and finally, two data sets were selected for empirical analysis. The results show that the mean absolute percentage errors of the point forecast of this model are 1.96% and 1.84%, respectively, about 31% and 15% higher than those of the competing models (mean absolute percentage errors 2.57% and 2.13%, respectively). For interval forecasting, the accumulated width deviation is 0.0037 and 0.002, respectively, about 35% and 25% higher than those of the competing models (accumulated width deviation 0.005 and 0.0025, respectively). Thus, the proposed forecasting framework outperforms all comparative models and can be used effectively for forecasting crude oil prices.