Stock Price Index Research Based On PCA And NARX Neural Network

The stock market in the financial market occupies an irreplaceable role,and the overall trend of the stock market is the stock index.The trend of the stock index is the basis for the regulation and control of the stock market,where the investors can understand and grasp the chaotic law of China’s stock market volatility more accurately,This can provide some important issues such as portfolio selection.Therefore,how to accurately predict the stock index has a very important significance.Due to many factors which influence the stock index,the fluctuation range is large,the factors have high noise and non-linearity,and the traditional stock price forecast method can not eliminate the high noise of the data.Then in order to improve the accuracy of stock forecasting,based on previous studies,we propse the method as follows:First,we review the relevant factors that affect the trend of stock index.On this basis,we choose 21 macroeconomic indicators and six public opinion indicators,which constitute 27 indicators of variables,Then we study the relution between the closing price of CSI 300 index and these variables.we analyzed these 27 variables by principal component analysis,and the number of variables were reduced to 6.where the redundancy between the factors were elimimated.Next,based on the research of stock price forecasting with neural network,We apply the NARX neural network to forecast the stock index.Based on principal component analysis A NARX neural network stock forecasting model(PCA-NARX)is formulated,which by Levenberg-Marquardt algorithm,Bayesian-Regularization algorithm and Scaled Conjugate Gradient algorithm.Finally,we train the PCA-NARX model by the economic data of Shanghai and HS300 Index,and compare the results of PCA-NARX model and NARX model.The conclusion of the proposed method is that there exict differences of different algorithms,all the R values of the solution are greater than 0.85.Furthermore,the higher the complexity of the training algorithm,the longer the time is required for the solution,and the better the fitting is.After the number of variables is recluced by principle compant analysis,the efficiency of the training network is improved,the model over-fitting optimization is roduced,and the model generalization ability is better.This is also a good way to verifies the effectiveness of predictably the stock market by PCA-NARX model with news and public opinion combined with macroeconomic indicators.