Yield Forecasting Based On Long Short-Term Memory Neural Networks

China's stock market has developed for nearly 30 years,but the stock market is not as mature as America's because of the regulation.At the same time,investors'vicious investment events occur occasionally,which makes the securities market full of risks and the stock market instable.Investors face severe challenges now.In this context,how to accurately describe the future earnings of the stock market and related risks has become a hot issue that many scholars and investors have paid close attention to.Due to the highly self-learning,stability,and abstract simulation capabilities,neural networks are more advantageous for predicting financial time series than mathematical models in statistics and econometrics.As one of the classic models of deep learning technology,long short-term memory neural network is very advantageous in mining long-term dependence of sequence data,suitable for processing and predicting time series with relatively long interval and delay.For general recurrent neural networks,gradient disappearance and gradient explosion are prone to occur.The gradient disappearance problem can be prevented by optimizing the activation function or setting a changeable rate of learning,but it is not a good solution.The LSTM neural network has a self-loop and gate control mechanism,which can effectively avoid the gradient disappearing.And we can add regularized terms to avoid gradient explosion.In this paper,we first introduce the basic theory and fitting method of the traditional ARMA-GARCH model,and introduce the basic test methods before modeling with GARCH model.Then we introduce the LSTM neural network,deeply study the mathematical structure in its complex structure,and deduce the gradients of weight coefficients and bias term.In the empirical part of this paper,we forecast the rates of return of Shanghai Composite Index,Shenzhen Composite Index,SSE50,CSI 100 and CSI 300 with LSTM neural network,and then compare it with the traditional ARMA-GARCH model.The accuracy between different indices was compared and analyzed.When using the LSTM neural network for forecasting,the volatility variable is introduced.By predicting the volatility of the next trading day,the rate of return of the next trading day is calculated,which has certain compatibility with the forecasting step of the GARCH model.In order to facilitate the comparisons between the LSTM neural network and the GARCH model,the historical rate of return and the historical volatility are selected as the independent variables of each layer of neurons in the neural network.We choose RMSE and qualitative methods according to the three types of rise,stationary and fall to measure the accuracy of the predictions of the two models.From the prediction results of the five indices,the accuracy of the LSTM model is better than that of the GARCH model.In addition,when using the LSTM model for prediction,there is a certain difference between the predicted values of different indices.The accuracy of the CSI 300 is higher than other indexes,because the time of training set and test set of the five indexes selected in this paper is the same,the difference in forecasting accuracy may be relative to the sample stocks of each index.And the sample stocks of the Shanghai and Shenzhen 300 Index are large-scale,highly liquid,and good-quality corporate stocks.