Resharch On Stock Forecasting Based On Deep Learning

With the continuous improvement of China's economic system,the continuous improvement of the financial system,and the gradual maturity of the stock market,more and more people are engaged in the way of financial investment in stock trading.The volatility of stock prices is not only affected by investor awareness,but also affected by the national economy,policies,and other aspects.Therefore,exploring the development trend of the stock market not only can provide certain benefits for the broad masses of the people,but also can provide a strong reference value for the construction of the country's economic foundation and the formulation of relevant policies.In recent years,due to the rapid development of neural networks,the development of deep learning theory has been promoted,and it has been widely used in various practical applications,which has spurred an upsurge of deep learning.Based on deep learning theory,this paper explores the feasibility of LSTM(Long Short-Term Memory)prediction of stock returns in China's stock market.Then on the basis of LSTM,combined with the relevant theories of deep learning,the corresponding model structure improvement and optimization are proposed,and corresponding model comparisons are made.The main work and innovation of this article include the following points:(1)Proposed to use LSTM model to predict stock returns.First of all,it theoretically analyzes the advantages of LSTM over other models.Then described in detail the LSTM model construction step method and related theoretical basis.Finally,using the constructed LSTM model,the influence of the data of different sequence segment length T and the yield of different prediction horizon ? on the prediction effect of the model is explored.(2)AdamL(Adaptive Moment Estimation Log)model parameter optimization algorithm is proposed.The superiority of the Adam algorithm compared to other optimization algorithms is analyzed,and it is pointed out that the Adam algorithm has drawbacks in the processing of the learning rate.A new learning rate attenuation formula is proposed and applied to Adam algorithm.The AdamL algorithm is obtained and used in the parameter optimization of the post-order model.(3)The model of LSTM-Ds(LSTM-Denses)is proposed and applied to the prediction of stock returns.Firstly,two network layers are added after the LSTM model output.then,a Dropout layer is added between the two network layers to prevent the model from overfitting,and the LSTM-Ds model is obtained.Finally,the LSTM-Ds model was used to forecast stock returns,and compared with the LSTM model prediction results.The experimental prediction results show that the LSTM-Ds model has better prediction performance than the LSTM model.(4)A SAE-LSTM-Ds(Stack AutoEncoder LSTM Denses)stock prediction model was proposed.Firstly,the feature extraction ability of the Stack AutoEncoder is used to extract the features of the original input data;then the extracted features are taken as the input of the LSTM-Ds model,thereby obtaining the SAE-LSTM-Ds prediction model;and finally the SAE-LSTM-Ds It is used to forecast stock returns and compared with the LSTM-Ds model.The experimental results show that the SAE-LSTM-Ds model has greater predictive power than the LSTM-Ds model.