Research On Stock Price Forecasting Based On DBLSTM-ARIMA Model

With the rapid development of the national economy,people's material needs have been basically met.With the improvement of living standards,people have more energy to participate in the activities of stock investment and become a part of the stock market.In recent years,deep learning,especially recurrent neural network,has been very popular in the field of time series analysis.The classic model of recurrent neural network LSTM model is the most widely used.Stock data is the most typical research data in time series analysis,and with a series of ideas such as algorithmic trading and quantitative investment,it is the general trend to apply the method of recurrent neural network to the research of stock data.In this paper,DBLSTM-ARIMA model is used to predict the closing price of stock to verify the applicability of machine learning fusion statistical method model.In the process of modeling,specific solutions are put forward for the over fitting problem.From the empirical point of view,this paper provides some valuable information for the construction of stock forecasting model based on machine learning and statistical learning.The data set selected in the experiment is the closing price of Shanghai 50 index,and the evaluation criteria of the model are average absolute error,mean square error and root mean square error.The method used in the experiment is bidirectional LSTM network,which has the following advantages: first,it avoids the problem of gradient vanishing or gradient exploding in the recurrent neural network;Second,it is conducive to the learning of some information that has a long-term dependence on time;Thirdly,it is conducive to the use of the context of two time directions(forward and backward)in time series analysis.In addition,multiple bidirectional LSTM layers are superimposed in the experiment.This multi-layer neural network structure is more conducive to mining the deep features of time series data.Because the connection mode of the whole neural network structure is full connection and the number of network layers is high,the training difficulty of the model may become greater and the convergence speed will become slower.Therefore,it is necessary to use dropout strategy in the experiment process to avoid these problems.The empirical results show that the performance of deep bidirectional long short memory neural network is better than that of single layer long short memory neural network in mean absolute error,mean square error,root mean square error and coefficient of determination,and the performance of single layer long short memory neural network is better than that of deep bidirectional long short memory neural network in explaining variance score.Dblstm model is better than LSTM model.The contributions of this paper are as follows: firstly,dblstm-arima model is proposed;Secondly,HP filtering method is used to decompose the stock data into two parts: trend term and cycle term;Third,the ARIMA model and machine learning model are used to fit the two parts of the stock data,and the stock price is predicted by the two results.The experimental results show that the prediction effect of machine learning fusion ARIMA model is better than that of ordinary machine learning model.