Research On Prediction Of High Frequency Data Of Stock Index Based On LSTM And GRU Neural Network

Today,with the rapid development of many emerging technologies,the prediction of financial time series data continues to grow.This paper studies the application of neural networks in the prediction of stock index high-frequency data.Long-short term memory(LSTM)recurrent neural net-works and gated recurrent unit(GRU)recurrent neural networks are used to predict time series data.This paper discusses the forward propagation process and error back propagation process of LSTM recurrent neural network and GRU recurrent neural network in detail,and analyzes the feasibility of neural network in financial time series predictionThis article conducts three aspects of research:one is to conduct a comparative study of LSTM and GRU;the second is to verify the universality of LSTM and GRU single-step prediction in stock index high-frequency data;the third is to expand and explore the multi-step prediction of deep learning.In this paper,we use the five-minute high-frequency data of Shanghai Securities Composite Index,Shanghai Securities 50 Index,China Securities Small Cap 500 Index and China Securities 1000 Index for a total of 731 trading days for three years,and construct LSTM,GRU uni-factor single step models,Multi-factor single step models and multi-factor multi-step models,with root mean square error(RMSE),mean absolute error(MAE),and R2 score as evaluation indicators.The gradient descent and prediction results in the prediction process will be visualization as fig.In the LSTM uni-factor single step model and the GRU uni-factor single step model,the mean RMSE and MAE of the four data sets are less than 0.03,and the average R2 score is greater than 0.99;in the LSTM multi-factor single step model,the GRU multi-factor single step model and in the LSTM multi-step model,the RMSE and MAE values of the four data sets are less than 0.10,and the average R2 score is about 0.95;in the GRU multi-step model,the average RMSE of the four data sets is 0.8107,and the average value of MAE is 0.7189,R2_score appears negative,and the model fails Through the comparative study of LSTM and GRU in 3 loss functions,it is found that when the model is simple and the amount of data is small,the effect of GRU is better.As the amount of data increases and the complexity of the model increases,the performance of LSTM is more stable.The study also found that GRU takes less time in the training process.And the study found that LSTM and GRU neural networks are universal for single step prediction of high-frequency time series data of stock indexes.In addition,we found that he deep learning model built by two-layer neural network does not always work in multi-step prediction,the model built by LSTM performs well,and the model built by GRU is less effective.The research in this paper selects high-frequency data of stock index,so the research results provide a reference for high-frequency trading to a certain extent.This paper perfects the comparative study of LSTM and GRU,which has a certain significance in promoting the change of neural network principles.In addition,it has practical significance for broadening the scenarios of artificial intelligence algorithms,diversified financial time series data forecasting methods,and broadening multi-step forecasting ideas.