Application Research Of LSTM Neural Network In Time-Series Model

With the advent of the era of big data,information has penetrated into various industries and become an important production factor.With the passage of time,People are mining huge amounts of data and analysis and prediction of time series data gradually increase.Nowadays,with the globalization of economy and finance,the scale of economic market is increasing.The stock market has become an important part of the economic market and a representative of people's analysis and prediction of time series data.Traditional time series analysis is difficult to fit the non-stationary and nonlinear time series data such as stock.Neural network is gradually applied to analyze different types of data with its powerful information processing ability.For predicting time series data such as stocks,the most suitable model is Recurrent Neural Network(RNN).Scholars use Long Short-Term Memory(LSTM)to improve it and solve the problem of long-term dependence.In this thesis,LSTM network model is taken as the basic model,and three groups of17-dimensional index individual stock data and 24-dimensional index data are selected to clean and standardize the data so as to eliminate the dimensional influence.Firstly,RNN network and LSTM network are introduced in detail,and a prediction model based on LSTM network is established.Secondly,two feature extraction methods are introduced,namely Principal Component Analysis(PCA)and Denoising Autoencoder(DAE).In this thesis,dimension reduction results of PCA are used as a reference for dimension reduction of DAE.Finally,the Attention Mechanism(AM)is introduced into the DAE-LSTM network model,and the prediction model based on AM-DAE-LSTM network is established.Three evaluation indexes were selected to compare the prediction results.The results show that feature extraction and attention mechanism can effectively improve the prediction accuracy.After the stability analysis of THE AM-DAE-LSTM network model,the prediction accuracy is higher than 0.9,which verifies the reliability of the model in the field of stock prediction.