Research On Stock Price Prediction Based On Improved LSTM Model

With the rapid development of China's economy and the enhancement of people's investment consciousness,the stock has become one of the important investment ways in daily life.However,the high-risk nature of stock will bring huge hidden danger to investors,so it is of great significance for investors to predict the trend of the stock market.As we all know,the stock is a complex system with many factors.When using technical methods to build a predictive model of stock,the training speed of the model is slow and it can not learn the implied rules of the data because of the problems of input data variables,overlapping data information,a large impact of outliers on training and so on.Moreover,the stock prediction model itself has some problems,such as the parameters are difficult to determine and the specificity in the field of stock prediction is not enough,which often leads to poor generalization and poor prediction effect of the trained stock prediction model.Because of the above problems,this paper proposes to use Z-score standardized method to deal with the basic data,to eliminate the influence of large value,different dimension and abnormal value in stock data on training.Principal component analysis is used to reduce the dimension of basic data,reduce the dimension of data while retaining the original information and eliminate the correlation between indicators,so as to improve the analysis efficiency and prediction accuracy of the model.Combined with KDJ and MACD,we can get the technical index which can express the deeper hidden law in the data.Taking all index values as input data,we can reduce the learning difficulty of the model and improve the learning efficiency and learning ability from the perspective of sample quality.At the same time,in the LSTM neural network,according to the characteristics of the stock,we adjust the structure of the neural network and the parameters of the model,improve the specificity and prediction stability of the model for the stock,and introduce the dynamic learning rate to optimize the convergence curve,solve the problem that can not be fitted and further improve the training efficiency of the model.At last,the idea of stochastic volatility model is combined to add the measure of volatility to neural network.By controlling the volatility to the output gate,the response speed of the model to emergencies is improved,and the prediction ability and universality of the model are enhanced.To make the experiment representative,the real data of Ping An Bank stock in A50 which is the most representative in the Chinese stock market is selected for training and prediction.The improved LSTM model based on the principal component reduces the average error of prediction,minimizes the running time,improves the stability of prediction,and finally predicts Ping An Bank's closing price more accurately,judging from results in a randomized controlled trial.Also,the training model is used to predict other stocks in A50,and the prediction result of the overall trend of other stocks is accurately,which shows that the model has certain generality.Finally,according to the Elliott wave principle on the buy and sell point obtained in the forecast curve to operate,and using real data to compare,the results show that the overall situation is profitable.So it is proved that it has a certain application value to predict stocks using the LSTM neural network model based on the principal component.