The rapid development of the futures market and deep learning technology has attracted the attention of many researchers.Deep learning models in growing number are used in futures market researches,among which forecasting futures trends is a hot research issues.A futures trend prediction model based on attention hybrid neural network is proposed after analyzing the researches of futures markets at home and abroad.The data set used is the 5-minute futures trading data from the main rebar contract and the Shanghai and Shenzhen 300 stock index futures contract from March 26,2010 to March 26,2020.In order to improve the quality of futures trading data,the wavelet transform method is used for data denoising,after which the noise is easier to distinguish.The triplebarrier method is adopted as the data labeling method,which combines the direction of futures price changes to get closer to the real futures trading scenario.By selecting different futures technical indicators and time periods to increase the richness of data features,and using random forest method to filter the data features.In terms of network architecture,a new attention-based hybrid neural network is designed with a combination of the two neural networks and attention mechanism algorithms after analyzing the advantages of both convolutional neural networks and recurrent neural networks in spatial dimension features and temporal dimension features extraction.Finally,the performance experiment of attention-based hybrid neural network is conducted compared with the existing futures forecasting networks like CNN-TA,CNN-T,Bi LSTM,Att-Bi LSTM.The experiment results show that the attention-based hybrid neural network is superior to the other four networks in both network learning ability and network generalization ability. |