Gold Futures Price Prediction Based On CEEMDAN-PCA-LSTM Model

Gold has good physical and chemical properties,making it difficult to change due to time and environment.Therefore,the history of gold change is the history of human civilization evolution.In recent years,due to the influence of gold supply and demand and the international situation,the international gold market has changed rapidly.As a special gold investment product,gold futures can maintain market stability,diversify transfer price risks,and stabilize the national economy for the macro economy;for individuals,it can be hedged.Therefore,the research and prediction of gold futures price fluctuations are of extremely important significance to the country or individuals.Even if the accuracy of the prediction is slightly improved,it may have a significant impact on trading decisions.However,compared with other financial time series,there are still relatively few studies on gold futures,and most of them are based on traditional econometrics and statistical models,the combination with deep learning also lacks innovative forms.Therefore,there is still a lot to do for research on the forecast of gold futures prices.This article aims to establish a new gold futures price prediction model,and at the same time provide a new idea for the future financial time series prediction problem.In the past forecasts of gold futures prices,most of the methods used were to find other economic variables related to it as much as possible,and then fit the equations of gold futures prices on these variables,including linear and non-linear ones.However,considering the complex formation mechanism of gold futures prices,there are many factors that affect it,and it is difficult to find all the characteristics related to it.Therefore,this article starts from the gold futures price itself,based on the idea of multidisciplinary integration,introduces the CEEMDAN method for signal processing,fully extracts the volatility information implicit in the price sequence itself,and decomposes a series of new features.Then through principal component analysis,the obtained intrinsic mode functions and residual sequence are further reduced in dimensionality and feature reconstruction to improve the calculation efficiency and prediction accuracy of subsequent models.Finally,the long short term memory LSTM with memory is used to make full use of the long-term dependence in the sequence to predict the gold futures price sequence.In terms of empirical research,this article uses the daily closing price data of the main continuous contract of COMEX gold futures from the Wind database for model training and testing.The comparison shows the advantages of CEEMDAN compared with EMD and EEMD in signal decomposition;at the same time,compared with the CEEMDAN-LSTM model that does not use principal component analysis,it reflects the role of principal component analysis in dimensionality reduction and feature reconstruction;and finally the comparison with LSTM and GARCH shows the advantages of this model compared with the traditional single model.The experimental results show that the various evaluation indicators of the CEEMDAN-PCA-LSTM model on the test set are:RMSE=9.38,MAE=7.82,MAPE=0.60%,DS=70.10%,which is better than other comparison models,which proves the superiority and effectiveness of this model.