| As an important basic energy source in the world,crude oil is the basic guarantee for the normal operation of the national economy and the development of people’s livelihood,and an important support for the smooth operation of the entire modern society.As a strategic resource,it plays an important role in the global economy,political situation and military power.Accordingly,the fluctuation of crude oil prices will cause widespread concern around the world,and predicting changes in oil prices accurately and quickly is invaluable for decision makers.Due to the comprehensive influence of factors such as oil market supply and demand,global economic situation,stock futures market,speculative behavior,geopolitics and emergencies,oil prices often show high complexity,non-stationarity and nonlinear characteristics,which also brings huge challenges to oil price forecasting research.In this context,this article aims to improve the prediction accuracy and reduce the computational cost,based on the three typical data traits of complexity(including memory,chaotic and uncertainty)to carry out the research.The specific research content is composed of three aspects follows:(1)Memory-trait-driven decomposition-reconstruction-ensemble learning paradigm for oil price forecastingThis thesis proposes a new memory-trait-driven decompositionreconstruction-ensemble learning paradigm in terms of the memory trait of oil prices.This novel paradigm can improve the accuracy of the oil price prediction model while reducing the computational complexity of the prediction model.Specifically,each step of decomposition-ensemble forecasting,i.e.,data decomposition,components reconstruction,reconstructed components prediction,and ensemble prediction in this study is built based on memory traits of the oil price data.In terms of,the relationship between memory traits and various technologies,the technology selection rules of the decompositionensemble prediction framework are established,which solves the problems of unfounded technical selection,time-consuming calculation and cumulative deviation of the classic decomposition-ensemble framework.Experimental results show that compared with other six types of benchmark models,the proposed prediction model has the best prediction performance,which proves that this method can be used as a promising prediction method for oil price prediction.At the same time,comparing with the classic decompositionensemble model,it can be found that the reconstruction of the long-term memory trait components and the ultra-long-term memory trait components can improve the predictive performance effectively of the model to a certain extent.(2)Chaotic-trait-driven decomposition-ensemble learning paradigm for oil price forecastingIn order to improve the prediction accuracy and interpretability of the prediction model,this study proposes a chaotic-trait-driven decompositionensemble learning paradigm for oil price forecasting based on the ideas of"divide and conquer" and "data trait driven modeling".Specifically,this study introduces chaotic traits as important attributes into the study of oil price forecasting.In the research process,the mutual information method,the CAO method and the Lyapunov exponent are combined to test the chaotic traits of oil price data.Meantime,three steps of decomposition-ensemble are combined with the chaotic traits test results of each stage,and accordingly the technology selection rules for the decomposition-ensemble framework driven by the chaotic traits are established for the first time.This improvement to the classic decomposition-ensemble framework not only enhances the interpretability of the prediction framework,but also improves the prediction performance of the model.At the same time,it provides some new research ideas for oil price prediction research.Experimental results show that the proposed chaotic-traitsdriven decomposition-ensemble model has better prediction performance than the other four types of benchmark models,and has better interpretability and universal applicability.This proves the effectiveness of the learning paradigm proposed in this study,and also shows that the learning paradigm can be used as a promising predictive tool for predicting oil prices.At the same time,data decomposition,components prediction and ensemble prediction driven by chaotic traits are significantly different from the corresponding benchmark models.This shows that the proposed learning paradigm is effective for every step of the optimization of the classical decomposition ensemble framework,which also proves the effectiveness of chaotic trait driven technology selection in the decomposition-ensemble framework.(3)Uncertainty-trait-driven decomposition-reconstruction-ensemble learning paradigm for oil price forecastingThis thesis proposes an uncertainty-trait-driven decompositionreconstruction-ensemble learning paradigm for oil price forecasting.In fact,this learning paradigm is an improvement of the classic decomposition-ensemble framework.Specifically,this study introduces the uncertainty traits of oil prices into the classic decomposition-ensemble framework,and takes the uncertainty traits as the decisive factor for technology selection.In the research process,the mainstream technologies are selected from each stage and category.The relationship between the uncertain traits and the technologies of each stage is gradually established through experimental research,which solves the problem of unfounded technology choice of the decomposition-ensemble framework to a certain extent.The experimental results show that the uncertainty-trait-driven decomposition-reconstruction-ensemble prediction model is significantly better than the other five benchmark models,which proves that it can be selected as a promising prediction method.At the same time,the effectiveness of the uncertainty-trait-driven components reconstruction method based on fuzzy entropy and Fisher clustering is confirmed from both theoretical and experimental aspects.This method can not only effectively reduce the computational complexity,but also improve the prediction accuracy for oil price forecasting.In this thesis,three types of complexity-trait-driven learning paradigms for international oil price forecasting are constructed by combining the "data trait driven modeling" and "divide and conquer" ideas,which can improve the prediction performance and interpretability of the prediction model,while also greatly reducing computational costs.In a certain sense,the thesis further confirms the effectiveness of the combination of the "data trait driven modeling"idea and the "decomposition and ensemble" strategy. |
