Uncertainty-based Crude Oil Price Forecasting

As the blood of the industry,crude oil is competed among countries.Its price volatility not only influences the domestic economy,but also plays an important role in international economy and politics.The price fluctuations mainly lies in supply-demand dynamic equilibrium,global economy,international politics and emergency events,which results in the high nonlinear characteristic and enhances the prediction difficulty.To improve the forecasting accuracy and reduce the computational complexity,this thesis focuses on crude oil price volatility analysis and forecasting model optimization,which is based on formulating the interpretation framework via uncertain factors,crude oil price prediction via uncertain factors ensemble learning and LSSVR ensemble learning with uncertain parameters.For this purpose,the main work for this thesis are described as follows.(1)Crude oil price volatility analysis based on uncertain factors investigationThe first work combines empirical ensemble mode decomposition(EEMD)and correlation analysis(Granger causality test and principal component analysis or linear regression)to formulate the interpretation framework.That model investigates the linkage between three categories factors(i.e.,supply and demand,economic factors and Google trends)and price related series(i.e.,decomposed components,reconstruction components and origin price series).Consequently,some conclusions can be summarized below.First of all,diverse frequency series contains different signals and the high frequency series is influenced by economic factors.Second,this framework is limited on low frequency series explanation.Finally,Google trends shows strongly explanatory capability on crude oil price volatility.(2)Crude oil price prediction based on uncertain factors ensemble learningThe second work considers the ensemble factor series as an external variable obtained from the former study.Those ensemble series are used to formulate the novel prediction approach to complete this research.In this study,the typical prediction methods are employed to verify the direction accuracy.Then,the experimental study shows that the prediction results mostly depends on ensemble tools.Prominently,the ensemble factor series obtained via linear approach(i.e.,PCA or LR)can promote linear prediction accuracy but limit the non-linear tools.Thus,the diverse factors ensemble method can provide the various ensemble results.(3)LSSVR ensemble learning with uncertain parameters for crude oil price forecastingThis work considers model predefined parameters as uncertain(or random)variables to formulate uncertain LSSVR ensemble learning model.The model includes four steps:probability distributions design,individual member formulation,individual results generation and ensemble results output.Compared with LSSVR models based on different searching approaches(i.e.,PSO,GA,FA and grid),it reveals that the proposed model with random parameters outperform the benchmarks.Furthermore,the proposed model diversifies the parameter searching method.The main contribution of this thesis presents an important interpretation framework and improve the time-series prediction method by uncertain processing.The experimental results prove their effectiveness of the uncertain processing methods.