SC Crude Oil Futures Price Forecasting Based On Hybrid Multi-source Domain-adversarial LSTM Model

SC crude oil futures was listed on March 26,2018,which is an important milestone in the opening up of Chinas financial market and the process of RMB internationalization.Since its listing,SC crude oil futures has been running smoothly and gradually recognized by the market,becoming the third largest crude oil futures in the world.As playing the important role in providing asset allocation tools for financial institutions and hedging tools for industrial institutions,SC oil future price prediction seems to be of great significance.In this thesis,a hybrid multi-source DANN model based on LSTM is proposed to obtain a model with good performance in forecasting SC crude oil futures price and provide reference for the market.Today,the methods of crude oil price prediction in the academic community are mainly divided into econometrics method and machine learning method.Due to the complexity,high volatility and non-stationary characteristics of crude oil price series,the prediction power of traditional econometrics method is very pool.In recent years,machine learning method has been widely used in time series prediction.It can effectively extract the historical information of very complex series such as crude oil price,and make high accuracy prediction of future value.However,the current machine learning model needs large amounts of data,otherwise it is easy to cause the problem of over fitting.The SC crude oil futures series listed in 2018 is obviously difficult to meet the training requirements,and the traditional machine learning method will not work.Fortunately,the research and application of transfer learning in recent years provide a feasible plan for the prediction of Shanghai SC crude oil futures series.Transfer learning refers to the use of knowledge in related fields(source domain)to improve the learning performance of data sets in the target domain.In this thesis,using the domain adaptation method,The source domain estimator can be applied to the target domain with similar effects to achieve the purpose of transfer learning.The visualization results of t-SNE and UMAP both show that the proposed method can effectively narrow the data distribution gap.In addition,through comparative experimental analysis,it is found that the prediction of SC crude oil price using the transfer learning model is more accurate than the non-transferring model,and the transferring effect using Brent crude oil futures price is better than WTI crude oil futures price,mainly because the Brent crude oil futures price series is closer to the SC crude oil price series.In order to avoid the uncertainty caused by market differences,this thesis introduces the idea of multi-source domain migration learning,and uses the MK-MMD value of measuring distribution differences as the weight to weight the 2 groups of results,which improves the models robustness.