Research On BiLSTM-GRU Stock Price Prediction Based On Two-stage Feature Extractio

The stock market is not only affected by many factors,but also has complex nonlinear dynamic interaction among the influencing factors,which makes the time series data become a complex system with the comprehensive characteristics of sequence correlation,nonstationarity and nonlinearity.Therefore,the establishment of scientific and effective methods to achieve accurate prediction of the stock market has become a very challenging work with important academic and application value.Traditional econometric methods and shallow machine learning methods have some important defects,such as insufficient processing ability of original data,over-fitting and slow convergence.Feature extraction and prediction modeling of stock price series have become a key problem in the field of financial data modeling.Taking Shanghai-Shenzhen 300 index and China Securities 500 index as research targets,this paper proposes an integrated framework of Bil STM-GRU stock price prediction based on two-stage feature extraction.Feature extraction and deep learning are combined for daily stock market closing price prediction.Considering the numerous factors affecting the stock market,several exogenous variables are introduced as the input variables of the model.After data pretreatment,principal component analysis(PCA)is used to reduce the dimension of data input in the first stage feature extraction,which can avoid redundancy between data and improve the operation efficiency of the algorithm.Considering the complexity of stock price changes,it is far from enough to process data by principal component analysis alone.Therefore,hierarchical clustering is adopted in the second stage of feature extraction to classify the data after dimensionality reduction.Hierarchical clustering mines and rearranges the main features of exogenous variables in training data,divides the strongly correlated data into a class,forms different training subsets,and avoids the problem that the model lacks generalization ability due to the feature dispersion of training data.Feature extraction lays a foundation for the subsequent establishment of bi-directional LSTM neural network(Bi LSTM)for prediction.For each training subset,Bi LSTM neural network was established respectively to obtain the prediction results for each class.Finally,the prediction results of Bi LSTM are input to the gated recurrent unit(GRU),which integrates the prediction results of multiple Bi LSTM.The neural network’s advantage in fitting datasets can be used to further fit the errors and optimize the prediction results.In order to verify the effectiveness of the proposed hybrid model,the benchmark model,such as machine learning,a single deep learning do contrast experiment,the results show that the hybrid model effectively combines the advantages of each model can be fully learn stock time series data of multi-scale,complex dynamic characteristics,such as a higher prediction precision and better generalization ability.This paper provides a feasible way of thinking and method for China A share price prediction.