Research On Prediction Of Real Estate Stock Price Index Based On SSA-GRU Recurrent Neural Network

The stock market is a complex system with complex non-linear dynamics and price fluctuations characterised by instability and chaos.In order to achieve accurate forecasting of stock price index time series and to guide investors in making sound investment decisions.This paper proposes an improved Gated Recurrent Unit(GRU)network prediction model,which on the one hand overcomes the limitations of traditional mathematical and statistical methods that cannot predict accurately,and on the other hand,the GRU network has stronger generalization ability compared with traditional networks due to the closer horizontal connection of the units between implicit layers and stronger network memory.In order to obtain the predictability of the property index closing price data,this paper first conducted a long memory analysis for the property stock price index,using the modified R/S test and the GPH test to verify the existence of long memory in the property index time series,laying the foundation for the subsequent forecasting study.Secondly,the model is improved as follows: firstly,68 indicator variables are selected in this paper,including 8 fundamental indicators,35 technical indicators,10 style factor indicators,11 macro-financial indicators and 4 policy indicators for constructing the variable indicator set,which is more convincing than the traditional indicator set.The paper then uses the Singular Spectrum Analysis algorithm(SSA)to noise reduce the original time series data,removing the trend term,seasonal term and noise term from the original data to obtain a smoother series for the construction of the network model,and the empirical evidence shows that the denoised series can yield a prediction model with a better fit.Finally,for the advantages of GRU networks with less gating and fewer training parameters,this paper uses the GridSearchCV algorithm to find the optimal combination of hyperparameters in the network to achieve a global search to make the best training structure for the whole network.Finally,this paper uses the daily trading data of the real estate index from March 2011 to March 2021 as the training set and the daily closing price data of the real estate index from June 2020 to March 2021 as the test set,inputs the denoised time series data into the GRU neural network prediction model,and then obtains the prediction set data after tuning the parameters according to optimal hyperparameter combination,and combined with the real data of the real estate index closing prices,the quantitative indicators of mean square error(MSE),root mean square error(RMSE)and mean absolute error(MAE)are used as model evaluation indicators to select a more reliable model for industry research.Finally the real estate stock price index forecast results are used as a reference for industry stock investment strategies.The empirical results show that the GRU neural network model,which is sufficiently trained,has strong reference significance in predicting the trend of the property share price index,and provides a basis for formulating investment strategies for the property sector as well as for individual stocks with high correlation.