Research And Application Of Neural Network And Its Combination Model In Metal Price Forecasting

Nonferrous metals are the basic materials and key strategic reserve resources for national economic development.The fluctuation of metal price seriously affects the development of modern industry and global economy.Therefore,accurate and robust nonferrous metal price forecasting is crucial and meaningful.At present,the price prediction methods of nonferrous metals mainly include single model prediction and mixed model prediction,but most of the research methods are too complex and the prediction results are relatively unstable.In this thesis,two novel hybrid models are constructed for metal price prediction through data preprocessing,model hyperparameter optimization and data post-processing.The research mainly designs single-step prediction and multi-step prediction experiments based on five different lengths of metal prices.Through model comparison and various evaluation indicators,a lot of analysis and experiments have been carried out on the prediction performance of the proposed model.The research work is mainly divided into :(1)This thesis introduces the relevant theories and concepts of metal price forecasting,summarizes the research background and significance,and focuses on the basic model theory and modeling process,data preprocessing methods,various evaluation indicators and methods for detecting model performance statistics.(2)A hybrid prediction model based on improved complementary ensemble empirical mode(ICEEMDAN),Bayesian hyperparameter optimization gated recurrent neural network(GRU)and autoregressive integrated moving average(ARIMA)model is proposed.Firstly,ICEEMDAN is used to decompose the original data and fuzzy entropy is used to judge the complexity of each subsequence after decomposition.The subsequence with the smallest fuzzy entropy value is predicted by the econometric ARIMA model,and the remaining subsequences are predicted by the GRU neural network optimized by Bayesian hyperparameters.Finally,the ARIMA model prediction results and the GRU-Bayesian prediction results are added to obtain the final prediction results.The proposed model is used to conduct single-step and multi-step prediction experiments on five metal price series with different lengths,and the prediction performance of the proposed model is analyzed by various evaluation indicators and model improvement percentages.(3)A hybrid model based on Prophet model,ICEEMDAN decomposition and multi-model preferred error correction is proposed.Compared with the neural network model,Prophet model has better interpretability in metal price prediction,which can divide the prediction results into trend items,holiday characteristics and seasonality(days,weeks and months).The Prophet model is used to predict the original price sequence,and the error sequence is decomposed by ICEEMDAN.The decomposed subsequence is imported into the constructed multi-model preferred error correction system for prediction.Finally,the final prediction results are obtained by adding the prediction results of each subsequence of Prophet model and multi-model preferred error correction.The DM test and model confidence set pruning are used to further test the overall prediction performance of the proposed model.The experimental results show that the hybrid model based on deep learning model shows good prediction performance in metal price prediction,which is of great significance to metal import and export countries and investment enterprises.