Interval decomposition ensemble approach for crude oil price forecasting

Crude oil is one of the most important energy sources in the world, and it is very important for policymakers, enterprises and investors to forecast the price of crude oil accurately. This paper proposes an interval decomposition ensemble (IDE) learning approach to forecast interval-valued crude oil price by integrating bivariate empirical mode decomposition (BEMD), interval MLP (MLP^I) and interval exponential smoothing method (Holt^I). Firstly, the original interval-valued crude oil price is transformed into a complex-valued signal. Secondly, BEMD is used to decompose the constructed complex-valued signal into a finite number of complex-valued intrinsic mode functions (IMFs) components and one complex-valued residual component. Thirdly, MLP^I is used to simultaneously forecast the lower and the upper bounds of each IMF (non-linear patterns), and Holt^I is used for modeling the residual component (linear pattern). Finally, the forecasting results of the lower and upper bounds of all the components are combined to generate the aggregated interval-valued output by employing another MLP^I as the ensemble tool. The empirical results show that our proposed IDE learning approach with different forecasting horizons and different data frequencies significantly outperforms some other benchmark models by means of forecasting accuracy and hypothesis tests.