An Auxiliary Decision-making Model For Precious Metals Futures Operations Based On Deep Learning Multi-step Forecasting

As a relatively mature futures in the financial futures market,precious metal futures plays a very important role in the current capital market because of its hedging function,price discovery function,resource allocation function and venture capital function.In order to give full play to the function of precious metal futures,it is necessary for enterprises in precious metal industry chain to predict the future changes of precious metal futures price.It is difficult to extract effective information from complex financial market based on a single forecasting model for effective forecasting.To solve this problem,a lot of researches combined various algorithms of various disciplines to construct hybrid forecasting models with the aim of improving prediction accuracy.In this study,a hybrid financial time series multi-step-ahead prediction model,which combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise and Long Short-Term Memory,is proposed.Complete Ensemble Empirical Mode Decomposition with Adaptive Noise in this prediction model is used to decompose the non-stationary original time series into several stationary time series components with different frequencies(Intrinsic Mode Functions).Then,part of Intrinsic Mode Functions which were related to prediction targets are selected as input features to feed Long Short-Term Memory to construct regression prediction model.To verified the validity and universality of the proposed prediction model,the gold futures and silver futures price indices from Shanghai Futures Exchange and New York Commodity Exchange were selected as the original data sets,and the daily closing prices of the four precious metal futures were utilized as the prediction targets;six baseline prediction models,Support Vector Machines,Multi-layer Perception Machines,Recurrent Neural Networks,two Long Short-Term Memory with different input features,were utilized to compared with the proposed prediction model,CEEMDAN-LSTM.The experimental results of this study show that:(i)Based on the criteria of R Squared,Root Mean Square Error,Mean Square Error,Mean Absolute Percent Error,and Mean Absolute Error,the CEEMDAN-LSTM hybrid prediction model proposed in this study is obviously superior to the other five baseline prediction models.(ii)In order to improve the prediction performance of the proposed prediction model,when the prediction step was increased,the number of selected eigen modulus functions should also be increased at the same time.(iii)With the increase of prediction steps,the CEEMDAN-LSTM prediction model proposed in this paper can restrain the attenuation of prediction performance to a certain extent on the basis of maintaining high accuracy of numerical prediction results.(iv)Transactions based on the predicted results of proposed CEEMDAN-LSTM model can yield more benefits than those under the "Buy and Hold" strategy.