A dynamic ensemble learning with multi-objective optimization for oil prices prediction

Accurately predicting oil prices is a challenging task since its complex fluctuation characteristics. This paper innovatively introduces the “metabolism” mechanism and sliding window technology and proposes a dynamic time-varying weight ensemble prediction model with multi-objective programming to ameliorate the oil price's prediction performance. This paper first adopts the random forest to select and generate the best feature sets. Second, different individual models are selected to build a heterogeneous ensemble prediction framework. Then, a multi-objective weight generation model is established by considering horizontal and directional accuracy. Moreover, the nondominated sorting genetic algorithm-II is utilized to compute the prediction errors of a single model at different stages and achieve model optimization selection and ensemble weight generation. Finally, we take Brent and WTI oil prices as the prediction objects to verify the effectiveness and superiority of the proposed model. The experimental results reveal that the dynamic time-varying weight ensemble forecasting model has excellent prediction capability for oil prices and can become an effective forecasting tool.