Research Of Stock Price Prediction Based On Features Autoencoder And Time Convolution Network

Nowadays,with the rapid development of China's economy,capital market,especially the security market,is increasingly interesting.Stock price prediction has always been the focus of both institutional and individual investors,which is of great significance in the field of capital market.Traditionally,investors make stock price prediction based on statistical analysis or simple machine learning methods,but these methods have great limitation because stock market is a complex nonlinear dynamic system.In recent years,neural network,relying on its powerful ability of nonlinear modeling,in the domains of computer vision,natural language processing,automatic driving and so on,has made very great progress.More and more people are considering using neural network to predicting price.LSTM,with its powerful sequence modeling ability,gets more favour.People also unceasingly through theoretical innovation,model innovation promote price prediction accuracy.Recently,time convolutional network(TCN)theory has achieved better results than LSTM and other recurrent neural networks in series tasks by virtue of its better sequence modeling ability.At the same time,as an unsupervised learning algorithm,autoencoder network has a unique effect on feature engineering and extracting advanced semantic features.Based on the above two aspects,this paper innovates the stock price prediction algorithm.This thesis deeply reviewed the theory significance and the necessity of stock price forecasting,and systematically expounds the scholars at home and abroad in recent years the progress made in the stock price prediction model innovation.In particular,in light of the neural network model of the performance,in-depth studies are based on constructing innovation model for forecasting stock prices--based on features autoencoding and time convolution network share price forecast model.,and USES historical data quantity and price to empirical analysis of the superiority of the model.The main innovations are as follows:(1)Based on the mature autoencoding algorithm and LSTM stock prediction algorithm theory,through combining the above two algorithms for the first time,we build "Feature Autoencoding Network+ LSTM" combinating stock price prediction algorithm.An empirical analysis comparing the above algorithm with "Pure LSTM"algorithm is as follows:select the historical quantity and price data of pingan bank,one of shenzhen component index constituents,extract high-level semantic feature through the autoencoding network,combine high and low level semantic features,input them into LSTM stock price forecasting model.After model training and testing,we will compare indicators between the above algorithm and pure LSTM stock price prediction model.Through the above model construction and comparative empirical analysis,the following conclusions are drawn:the prediction effect of stock price is partially enhanced by the advanced semantic features obtained through the self-learning of the feature autoencoding network;The date features of timestep length before time t predicts the stock price at time t+m,and the prediction effect of medium size m(roughly 20 in this case)is more stable and has better effect.(2)Based on the existing autoencoding algorithm and the latest time convolution network(TCN)algorithm,for the first time we use TCN algorithm in the stock price prediction task,combine it with "Feature Autoencoding Network" that is proved the super performance of stock price forecasting in(1)and build "Feature Autoencoding Network + TCN" combining stock prediction algorithm.An empirical analysis is made by comparing the algorithm with the "Feature Autoencoding Network +LSTM"combination stock price prediction algorithm in(1):the same historical price data of ping'an bank are selected and the same operation is carried out as in(1).Through the above model construction and comparative empirical analysis,the following conclusions are drawn:replacing LSTM with time convolution network can partially improve the prediction effect of stock price;The date feature of timestep length before time t is used to predict the stock price at time t+m.The prediction effect of medium size m(roughly 20 in this case)is more stable and has better effect.With the extension of the time step of the predicted data,the ability of TCN to process long strings of data is released,and the performance of the model is partly improved,but there is an upper limit of improvement.Finally,we prove the superiority of the "Feature Autoencoding Network + TCN"model in the task of stock price prediction through the above analysis.