Researches On Stock Index Prediction Model Based On Several Improved Dynamic Neural Network

People in 2015 has not only experienced a magnificent bull market, but also experienced an unprecedented crash, which result in a kind of roller-coaster change in most Chinese people. With the development of economy and the accumulation of wealth, stock market will play an important role in undertaking the asset allocation and financing for small and midsize enterprises. How well the stock market will do may directly affect people's livelihood, hence it's important to predict the trend of the market.Through literature review, we found that the traditional method of forecasting stock market is wavelet de-noising when in data de-noising. Then, unreasonable threshold may lead to excessive noise in high frequency part, which the primary data details could not be preserved well. Also, most of the articles reuse the methods, which lack of inventiveness.In this article, we extracted the closing rate of CSI 300 Index from April 6,2005 to April 3,2015 as the primary research database via Wind, consisting of 2428 groups of real data. When denoising the primary data, we put forward a method that mixed wavelet de-noising with partial differential de-noising, as the partial differential de-noising can better preserve the characteristics of the data. We selected the following three types of data:the primary data, wavelet de-noising data, wavelet partial differential equation de-noising data. We extracted 542 groups as the input variables of the forecasting models for comparative study.Due to the nonlinearity of stock data, this article introduced a GRNN Neural Network model, which has advantages in dealing with nonlinear problems. Then we compared it with the traditional stock forecasting:BP Neural Network. Based on the dynamic neural network theory, we added sliding time windows and feedback connection in the BP Neural Network, and then we got the'BP-Dynamic Neural Network Model'. Also, we added sliding time windows and feedback connection in the GRNN Neural Network, and then we got'GRNN-Dynamic Neural Network Model'.Finally we took three types of data into those two models for training and forecasting, and we got four types of combined models for forecasting stock data. The conclusion are as follows:1) By the de-noised index graph we found that the method of wavelet mixed with partial differential de-noising not only removed noise but also preserved the characteristics of the data.2) In the context of the same model, wavelet de-noising data has the best prediction result, followed by the wavelet partial differential equation de-noising data, primary data is the worst.3) In the context of the same data, the GRNN-DNN model is better than the BP-DNN model.4) In the context of the same Neural Network, wavelet de-noising data has the best prediction result, but this conclusion is built on the fact that the data which compared with the predicted values has been dealt with the wavelet de-noising method. When comparing predicted values with primary data, the method that combinated wavelet partial differential equation de-noising model and GRNN-DNN model has the best result.