Research On Prediction Of Stock Price Based On Neural Network

How to make correct judgments about the future trend of stock prices becomes one of the core issues that investors are concerned about.This not only helps investors make reasonable investment decisions,but also helps to increase the market's effectiveness and promote the effective and rational distribution of capital.As the stock market is a highly complex nonlinear trading system,there are many factors affecting the stock price,and the impact mechanism is complex.Traditional statistical models are difficult to effectively predict the stock price.The artificial neural network has a high degree of self-learning,self-adjusting ability,and strong robustness.It can consider a large number of influencing factors and fit the predicted target through complex nonlinear structures.So the paper explores the application of artificial neural network in stock price forecasting.The main work of the paper is:Firstly,this paper summarizes papers about the prediction of stock prices using artificial neural networks at home and abroad,and summarizes artificial neural network models such feedforward neural network model trained based on error back propagation algorithm(BP neural network).Secondly,for the recurrent neural network(RNN neural network)in deep learning that is rarely involved by domestic and foreign scholars,this paper discusses the variant long short-term memory network(LSTM neural network)of RNN neural network in stock price forecasting from the perspective of improving the gradient disappearance and gradient explosion defects.In the empirical analysis,the data selects the squall line data of all trading days from May 1,2005 to September 30,2017 of the CSI 300 Index,including the opening price,closing price,lowest price,highest price,volume,and amount.In which a total of 3022 daily historical transaction data of 6 attributes were included.The mean square error and mean absolute error of different model prediction results were compared and analyzed.The empirical results show that the prediction effect of LSTM neural network is better than that of RNN neural network.The prediction effect of RNN neural network is better than that of BP neural network.Thirdly,in order to avoid the defect that the commonly used stochastic gradient descent method is difficult to converge to the optimal value due to the fixed learning rate,the paper chooses a dynamic update learning rate and an adaptive parameter updating method Adam(Adaptive Moment Estimation).The mean square error and the mean absolute error of the prediction results of LSTM neural network based on the stochastic gradient descent method and that based on Adam are compared and analyzed.The empirical results show that the prediction effect of the LSTM neural network based on Adam is better than that based on the stochastic gradient descent methodFourthly,in order to improve the prediction ability of LSTM neural network,the paper adopts a stacked LSTM neural network and combines the stacked LSTM neural network and Adam algorithm to achieve better prediction results.The mean square error and the mean absolute error of the prediction results of stacked LSTM neural network based on Adam and LSTM neural network based on Adam are compared and analyzed.The empirical results show that the prediction effect of the stacked LSTM neural network based on Adam is better than LSTM neural network based on Adam.