Research On Stock Time Series Prediction Based On Neural Network Model

Stock time series is a common non-linear time series.The existing stock time series prediction researches are mainly focusing on the prediction research of single variables in multiple technical indicators of stocks.Most scholars only use one specific method to explore the prediction application which is one of the stock technical indicators,and do not consider the joint influence of multiple indicators on a single output,let along constructing a systematic and feasible stock price time series prediction modeling system.With the development of machine learning algorithms and the wide application of neural network models,considering the characteristics of high noise,non-linearity and complex influencing factors of stock time series data,this dissertation analyzes the problems of feedforward neural network prediction algorithm widely used in nonlinear time series prediction and LSTM multi-scale nonlinear time series prediction model based on feedback neural network with the help of new theory and method is put forward.And it is used in the prediction of stock time series.The experiment explores the inherent laws in stock time series and provides some practical value for the application of new neural network in stock time series prediction analysis.In this dissertation,the researches on establishment of prediction model based on feedforward BP neural network are proposed firstly,and the influence of its internal network parameters on prediction performance is analyzed.The single scale of the closing price is selected as the only input of the network to establish a BP neural network prediction model.The experimental results show that when the input dimension of the historical data set is fixed,the setting of the hidden layer node has a significant impact on network's prediction performance.When the historical training data set is large enough,the network error under the same input dimension is not sensitive to the setting of the hidden layer node.On this basis,this dissertation uses LM algorithm and Bayesian Regulation(BR)to improve the training rules of neural network.Based on this experience,this dissertation uses LM algorithm and Bayesian Regulation(BR)to improve the training rules of neural network.In addition to the five basic indicators: daily closing price,opening price,highest price,lowest price and closing price of the previous day,the input part of the network also includes three volatility indicators:the 10-day average price,the 20-day average price,and the historical volatility.The experimental results show that the prediction performance of the multi-scale prediction model is better than that of the single-scale prediction model,and the multi-scale BR-BP prediction model has better prediction performance.Based on this basis,a new feedback neural network LSTM network with timing concept is inducted for the time series characteristics of stocktime series.The stock time series is predicted and analyzed using the above prediction model,and the sample space still uses the above multi-scale training data set and drop out is used to optimize the network structure during training.Compared with the prediction results of the feedforward neural network model,it is feasible and efficient for the LSTM network to predict nonlinear time series.The prediction experiment analysis under different sample data shows that the model has certain robustness.