Research On Fluctuation Analysis And Prediction Of Crude Oil Price Time Series Based On Machine Learning

Crude oil is an essential national strategic reserve resource as a primary energy product.Crude oil and its extracts are involved in various industries in the manufacturing sector and permeate the general public clothing,food,shelter,and means of travel.Crude oil price is the cornerstone of many commodity price movements and a barometer of the changing international economic environment.Meanwhile,sharp fluctuations in oil price seriously affect the stability of national economic development,the stability of national life,and the sustainable development of society in all aspects.Analyzing and predicting crude oil price fluctuation can help policymakers,enterprise managers,and relevant investors deal with the risk of oil price fluctuation.The relevant research has important scientific significance and potential application value.It is well known that a variety of factors influence crude oil price.For example,it is affected by supply,inventory,demand,policy uncertainty,and geopolitics in the trade market and by the stock market index,exchange rate,and commodity price in the financial market.As a result,crude oil price exhibits non-stationary and non-linearities dynamic characteristics,which pose significant challenges to the fluctuation analysis and forecasting of crude oil price.In such a complex background,this paper attempts to delve into the following aspects: How to detect the dynamic relationship between relevant factors and crude oil price? How do we improve the generalization ability of oil price forecasting models? How do we provide an early warning signal for the dramatic fluctuations in oil price? The main research contents and innovative research results of this paper are in the following four aspects:(1)This study proposed a multi-scale lag linkage analysis method to investigate the linkage characteristics between related variables and oil price.The economic policy uncertainty(EPU)index and the price of West Texas Intermediate crude oil(WTI)are used as examples for empirical analysis.Firstly,the crude oil price’ long-term,mediumterm,and short-term trend series are obtained using the decomposition and reconstruction algorithm.At the same time,to capture the change of linkage between EPU index and oil price,the sample data is divided into different intervals according to the fluctuation state of oil price.Then,the linkage effect between EPU index and different trends is analyzed from three aspects by using time-delay cross-correlation analysis and its derivation method.Finally,the study results are used to improve the oil price decomposition integration prediction model.The empirical results show significant differences in the lag and correlation intensity between EPU index and crude oil price on different scales.The multi-scale lag linkage analysis between EPU index and oil price effectively improves the accuracy of oil price decomposition integration prediction model.(2)Most existing studies have used regression forecasting models to forecast the values of crude oil price,return or volatility.This research offers a machine learning classification hybrid prediction model based on modal fusion data features to improve the generalization capacity of forecasting models.First,from the technical analysis of securities price,this paper constructs five modal fusion data features using the variational modal decomposition algorithm: primary trend,relative position,deviation position,variance ratio,and fluctuation range change.Meanwhile,this paper symbolizes the crude oil price volatility series by using the symbolic time series analysis method to represent the fluctuation trend of crude oil price with symbols.It uses the trend symbols as labels for machine learning classification algorithms.Finally,three machine learning multi-classification models are trained to forecast the trend of crude oil price.The results of the empirical analysis based on WTI futures price show that modal fusion data features can help improve the predictability of crude oil price,and the classification forecasting algorithm can effectively improve the generalization ability of the forecasting model.(3)The existing research on price mutation prediction mainly uses econometric model to identify the structural mutation of price series.However,the econometric model cannot effectively predict the nonlinear time series.The structural mutation cannot directly measure the direction and amplitude of the series change in a specific time.This paper proposes a price mutation prediction model based on a deep learning algorithm.First,considering the difference in individual risk tolerance,three types of mutation points are defined according to the direction and magnitude of price fluctuation: mixed mutation point,up mutation point,and down mutation point.The mutation prediction model is based on long short-term memory networks,convolutional neural networks,and variational modal decomposition algorithm.The advantage of this model is that it can extract both temporal and different modes spatial features in the price series.The empirical analysis results based on WTI futures price show that the price mutation prediction model has good ability for three mutation points,especially for mixed mutation points.(4)Given that the combined forecasting model has better stability than the single forecasting model,this paper proposes a combination selection method based on machine learning classification algorithm for price trend forecasting.Considering the difference in price trend states in different intervals have significant effects on forecasting,this paper first constructs three state variables that can characterize the long or short-term trends and fluctuation states of price series,to improve the sub-models forecasting ability.Screening suitable sub-models is one of the challenges addressed by combined forecasting approaches.The screening method adopted by the existing combining model based on the regression algorithm is not suitable for this paper.For this reason,this paper proposes a new optimal-model set iterative selection algorithm.The proposed combined forecasting technique significantly outperforms the single forecasting model in comparative experiments using WTI futures price,which is related to the fact that the combined forecasting method enhances the forecasting accuracy of the trend category with high volatility.