Study On Forecasting Crude Oil Prices With ICEEMDAN And Machine Learning

As a national strategic resource,Crude Oil has the property of commodity and financial product.There are many factors influencing the price fluctuation of crude oil,such as market supply and demand,interest rate,exchange rate,national policy,political conflict,short-term flow of funds between international capital market,climate anomaly and so on.These factors are complex and data is difficult to be collected.Some of them are restricted by frequency of updating and even has lag.Scholars believe that no matter what the factors are,their effects on the Crude Oil prices will be reflected in the fluctuation of Crude Oil prices.Therefore,we can predict Crude Oil prices in the future through the Crude Oil prices in the past.This can ease the process of collecting the data of factors.However,it arises another problem.The time series of international Crude Oil prices are often non-linear and non-stationary,which brings difficulties for statistical models to forecast.Owing to the great efforts made by many scholars,they made a major breakthrough.Scholars proposed that it was easier and more accurate for models to forecast the Crude Oil prices after the sequence of prices is decomposed into linear or stationary pieces containing local physical characteristics of raw data.This thesis mainly proposes two new models for predicting crude oil prices—ICEEMDAN-XGBOOST and ICEEMDAN-TCN.The author of this thesis has published relative researches results on ICEEMDAN-XGBOOST-based approach on forecasting crude oil prices before.The previous thesis used CEEMDAN(complete ensemble empirical mode decomposition with adaptive noise)algorithm to decompose Crude Oil prices into several intrinsic mode functions(IMFs)and a residue,and then used XGBOOST to predict them respectively.Finally,the previous thesis added up the predicted results of these components to obtain the final prediction of Crude Oil prices.According to the empirical results,the previous thesis found that this method performed better than other models on experiments.Another new model for predicting crude oil prices proposed in this thesis,ICEEMDAN-TCN,also decomposes crude oil price data into several intrinsic mode functions(IMFs)and a residue.Then,it uses another model to replace XGBOOST.The model is called as TCN,which is proposed in 2018.It is said that this deep learning model performs better than the LSTM and GRU.The experiments of this thesis focuses on these two models.Since the author of this thesis published the corresponding experimental results on the ICEEMDANXGBOOST,the relative experiments on ICEEMDAN-XGBOOST in this thesis is simple.This thesis makes more experiments on ICEEMDAN-TCN.Specifically,this thesis emphasizes the performance of forecasting crude oil prices by the comparison of different model results.This thesis compares the experimental results of different models without the signal decomposition algorithm,and also compares the experimental results of different models with the signal decomposition algorithm.In the relative experiments on ICEEMDAN-XGBOOST,the benchmark model is set as the SVR model,which is determined based on previous research results.When it comes to the experiments on ICEEMDAN-TCN,this thesis set the SVR,XGBOOST,LSTM,GRU as the benchmark.By comparing the experimental results,this thesis finds that the ICEEMDAN-XGBOOST and the ICEEMDAN-TCN have a good effect in predicting the crude oil prices.Among them,the ICEEMDAN-TCN model performs better than the ICEEDMAN-XGBOOST model.