Research On Stock Price Prediction Based On Long And Short-Term Memory Network And Fused Attention

In the financial field,the stock market has always been one of the hot topics that people pay a attention to.Accurate prediction of the future trend of the stock market can help enterprises make reasonable financing plans,and can also develop more stable investment strategies for individual investors and institutions to reduce investment risks.Based on this need,many relevant researchers describe and evaluate the correlation between different stock variables by statistical inference,attempting to use past stock characteristics to predict future prices.However,the data derived from the stock market is essentially a nonlinear,non-stationary and multi-scale interval time series.In this case,the traditional statistical methods have certain limitations,which can not only assume a constant data,but also take into account the characteristics of high noise,time varying and dynamic.With the continuous development of deep learning,more and more scholars use time series prediction models based on deep neural networks to simulate the changing trend of stock market prices.Therefore,this paper conducts in-depth research on the existing Long and short term memory network(LSTM),proposes an improved model,and conducts experiments on real stock data sets.Specific work is as follows:(1)A multi-scale convolution attention model(MCA-LSTM)is proposed.Firstly,the model splits the multiple features of stock time series into two sequences,with the daily closing price as the target sequence and the rest as the exogenous sequence.Secondly,the encoder uses time convolution to extract the feature of exogenous sequence,and obtains the feature information of the exogenous sequence on multiple time scales by setting convolution kernels of different size.Then,combined with the hidden layer state of the decoder of the previous moment,these exogenous features are attention-weighted fusion to generate the context vector of the corresponding moment.The value of the target sequence of each time and the context vector of the corresponding time are combined as the input of the decoder,and the final prediction result is obtained.In this paper,the three stock data sets of Shanghai Stock Exchange50,Shenzhen Component Index and Shanghai Shenzhen 300 are compared with the current mainstream models RNN,LSTM,TCN,CNN-AM-LSTM,AE-AM-LSTM.The experimental results show that the proposed MCA-LSTM model is superior to all comparison models in many prediction error indicators.Therefore,the performance of stock price prediction model by integrating the attention mechanism of exogenous sequences can be greatly improved.(2)A stock prediction model(TA-Conv LSTM)integrating temporal attention is proposed.Firstly,the model introduces Conv LSTM module to better capture long-term dependencies of stock time series.Secondly,in order to extract the feature of the exogenous sequence adaptivelt at each time,and fully consider the influence of the feature on target sequence at different time,this paper introduces temporal attention module.Then,in order to eliminate the effect of time series prediction lag,the nonlinear transformation of target sequence is carried out.Finally,the TA-Conv LSTM model is tested on five datasets of Shanghai Stock Exchange 50,Shenzhen Component Index,Shanghai Shenzhen 300,Sinopec and Bank of China,and compared with the existing mainstream methods,the experimental results show that the TA-Conv LSTM model has superior prediction performance and strong generalization ability.