| Since the opening of the Shanghai and Shenzhen Stock Exchange in 1990,China’s capital market reform has been advancing in depth,the basic system construction has been gradually improved,venture capital,industrial capital,financial capital and resident capital have increasingly recognized the stability and value orientation of the domestic capital market,and their enthusiasm for participation has gradually increased.At the same time,the research on how to use measurement means to reasonably predict the stock price also rises.By revealing the trend of stock price and guiding rational investment and value investment,it has important application value for activating the capital market and stabilizing the overall economic situation.Among the existing mainstream methods,the quantitative prediction methods of stock price mainly include time series analysis method and neural network analysis method.The time series analysis method,which takes advantage of the strong periodicity of the stock market and only needs the recent price data,can simulate the future trend.It is simple to apply and accurate for short-term prediction,but there may be some deviation in the long term;Neural network analysis method regards the stock market as a nonlinear whole system,which has the advantages of strong self-learning ability,self-adaptive ability and high fault tolerance rate.However,there are some problems in practical application,such as slow convergence speed and so on.Therefore,it is very difficult to use a single method to predict the stock price trend.Therefore,this paper constructs ARIMA model,LSTM neural network model and BP neural network model at the same time to predict the Shanghai Stock Exchange 50 index and explore the use of different methods.In the specific experiment,this paper uses the closing price of Shanghai stock index from January 1,2019 to June 1,2021 and three kinds of influencing factors related to stock market,consumer sentiment and market environment to predict the index price from June 1,2021 to September 30,2021 and compare it with the actual value.During the construction of ARIMA model,firstly,the value range of P,D and Q is determined by network search parameter optimization method,and then the specific order value is determined by AIC and BIC rules.Then,LB Test is selected to determine the lag order and trend simulation is carried out;When constructing LSTM neural network model,Adam optimization algorithm is selected to determine the number of hidden layers and the number of input and output layers in the neural network model;When building the BP neural network model,the BP neural network model is built by determining the number of neurons,the number of neural network layers and the number of hidden layer nodesThe empirical results show that in terms of prediction accuracy,the prediction accuracy of ARIMA model,LSTM model and BP model are 45.45%,58.46%and 55.36%respectively.The prediction of neural network is more accurate than that of time series,and the accuracy of LSTM model is the highest;In terms of application,the three models fit well in predicting the rise and fall trend,but the accuracy is slightly poor in the period of large rise and fall;In terms of parameter setting,ARIMA model order selection(1,2,3)and neural network model hidden layer node selection 64 have the best effect. |
