Research On Expected Stock Return Based On LSTM

The study of expected return is a major vane of risk management and hedging.The improvement of the ability to predict the expected return is helpful to the research on the applicability of risk factors affecting the expected return under the characteristics of the domestic financial market.This paper mainly includes the following four parts:The first part USES the LSTM has obvious advantage in dealing with financial time series neural network model,the object of study on model building and diagnosis,design under the different hidden layer,the fitting effect of different number of neurons on the model,will receive the LSTM optimization neural network model,to avoid over fitting,owe fitting the training effect.Conclusions can be drawn:1.The non-stationarity of data is one of the reasons for reducing the prediction performance of LSTM neural network model;2.Overall,the prediction accuracy of LSTM changes in the same direction with the number of hidden layers and in the opposite direction with the input span,while the influence on the number of neurons in the layer fluctuates greatly and does not change regularly.The second part firstly forecasts the price and yield of CSI300 through the preliminary LSTM model.It is found that the LSTM model with multi-variable input has better prediction effect than the LSTM model with single-variable input which can predict the closing price.Secondly,the LSTM model with multivariable input is adopted to predict and analyze some components of CSI300,and the prediction results of different stock characteristics are different.Thirdly,a buy-and-sell strategy portfolio based on the LSTM model to predict the trend of individual stocks is constructed for backtesting.The results show that the LSTM model can bring the excess return of the strategy portfolio higher than the CSI300.Finally,comparative analysis the LSTM model,convolution neural network(CNN)model and the advantages and disadvantages of random forest model prediction effect found that LSTM model on the prediction precision is better than the other two,CNN model prediction accuracy of the worst,but has a good advantage in the direction of the trend,random forest model on the prediction precision is better than CNN model,but insufficient stability prediction.The third part is the empirical analysis of China's a-share market from May 2016 to April 2019 based on the fama-french three-factor model.The results show that the traditional three-factor model is not capable of explaining the risk premium of China's a-share market.Then,by constructing A multi-factor model based on the fama-french five-factor model and adopting the fama-macbeth cross-sectional regression method,the author searched for new effective factors to better explain the economic operation characteristics of China's a-share market.The conclusion of the above research shows that the six-factor model constructed by the market factor,scale factor,book-to-market ratio factor,profit factor,the profit-to-share ratio factor and liquidity factor among the valuation factors has more advantages in explaining the risk premium in China's a-share market.A factor model,the fourth part is based on the research conclusion,first of all factors of timeliness,in this paper,a variety of characteristics of statistical analysis of the factors of IC value,observation factors effective continuous interval of that,the factors of effective continuous interval as important basis of frequency factor to choose the sort of storehouse,the empirical results show that the effective factors of continuous interval can improve build by multi-factor model expected earnings,have LSTM prediction model is better than no LSTM prediction model of factor when choosing a higher expected return.