Research On Copper Spot Volatility Prediction Based On GARCH And Neural Network Hybrid Model

Copper has been widely used because of its good chemical and physical properties.China's copper consumption has accounted for one-half of the world's total copper consumption,and has become the world's largest copper consumer.The drastic fluctuation of copper price will increase the uncertainty of the market.It is helpful to predict the volatility of copper earnings correctly for risk management of market participants and the formulation of national policies.Based on the previous literature,we innovatively extend the ANN-GARCH model for copper market fluctuation prediction proposed by kristjanpoller and Hernández(2017),and the ANN medel used by Lin Jie and Gong Zheng(2018)to forecast shanghai zinc futures to ANN-LSTM,ANN-Bi LSTM,ANN-LSTM-GARCH and ANN-Bi LSTM-GARCH new models.ANN model is suitable for extracting the characteristics of spatial dimension,while LSTM and Bi LSTM model are better at capturing time dimension features,while GARCH model has strong financial interpretation power.Therefore,the mixed model can extract the fluctuation information of copper in the spot effectively,and improve the prediction ability of the model.In order to verify the validity of the model,we compare the hybrid model with the existing models of copper volatility prediction(GARCH,ANN and ANN-GARCH).We use the model to forecast the volatility of the spot copper in the next three weeks(15 days)in the future from 2008 to 2018,and use the prediction results of four error indicators MSE,MAE,MAPE and RMSE.In order to ensure the reliability of the prediction results,the robustness tests of the whole sample set and the sub sample set are carried out respectively.That is,the model is used to predict the volatility of copper spot in the next 2 weeks(10 days)and the next 4 weeks(10)days to carry out the whole sample set robustness test,and the six sub samples in bear and bull market,high and low liquidity market,and high and low exchange rate market are respectively used to carry out the sub sample test.Finally,the paper also compares the optimal model to analyze the predictability.The results show that:(1)the results of the prediction of copper volatility of ANN-LSTM and ANN-Bi LSTM are better than ANN,GARCH and ANNGARCH models.For example,when predicting the volatility in the next 15 days,the MSE error of ANN-LSTM(4,20)is 12.45%,47.29%,31.03% lower than ANN(4,20),GARCH and ANN-GARCH(4,20)respectively.The MSE error of ANN-Bi LSTM(4,20)models are reduced by 13.31%,47.81% and 31.70%.(2)the mixed model constructed by using GARCH prediction value as the influencing factors can further improve the prediction accuracy of the model.For example,when predicting the fluctuation rate of copper spot in the next 15 days,the ANN-LSTM-GARCH(5,10)model is compared with ANN-LSTM(5,10).The accuracy of the model was improved by 9.35%,and that of ANN-Bi LSTM-GARCH(5,10)model was 17.62% higher than that of ANN-Bi LSTM(5,10).(3)In the process of robustness test,it is further found that the longer the period of Volatility Prediction,the stronger the prediction ability of hybrid model is.That is,when predicting the volatility of copper spot in the next20 days,15 days and 10 days,the prediction ability of the mixed model is gradually decreasing.At the same time,the number of layers and neurons of different neural network structures has little influence on the prediction performance of copper volatility of mixed model.(4)The results show that compared with bear market,three optimal models(ANN-LSTM-GARCH(6,10),ANN-Bi LSTM-GARCH(4,10),ANN-LSTM-GARCH(5,20))are better than bear market,low liquidity period and high exchange rate period.