Neural Network Stock Prediction Research Based On Wavelet Anlysis

With the rapid development of China's economy, the stock market of China is also growing. More and more investors put money into the stock market. High returns in the stock market are attracting investors, but high risk also comes. So, how to bear the smallest risk in the case of a larger income became a target for investors. Not only that, if we can grasp the stock market situation, it's also advantageous for the national macroeconomic regulation and control. So, the stock prediction became a hot research topic at home and abroad.However, the stock market is a nonlinear dynamic system, the traditional time series forecasting method is difficult to reveal its inherent law. So the prediction effect is difficult to have a breakthrough. This paper chose the neural network model which has good nonlinear fitting effect, combined with the theory of wavelet analysis, proposed a dynamic prediction model of neural network based on wavelet analysis, to model and forecast the part of the closing price of the CSI 300 index. The main research contents of this paper are as follows:First of all, this paper introduces the wavelet analysis, through the wavelet decomposition and reconstruction, the original time series is decomposed into different scales, according to the characteristic of different scale sequences, modeling respectively. And in the process of wavelet analysis, the selection of wavelet basis function and decomposition scale are determined, all of which are analyzed in detail to provide for the follow-up work.Through the wavelet analysis, a low frequency sequence and several high frequency sequences are generated. In this paper, we will use the Elman dynamic neural network model to analyze the low frequency sequence which is sensitive to historical data, and use the traditional BP neural network model to analyze the high frequency sequence of the random characteristic. Finally, integration of the data, get the forecast results.In the process of the neural network modeling and analysis. In order to improve the training efficiency of the network and avoid the local minimum, we use the additive momentum method and the gradient descent method with adaptive rate adjustment to improve the network convergence slow and easy to fall into local minimum in the traditional learning algorithm of neural network. Compared with the experimental results, the improved neural network learning algorithm is significantly improved in the training effect and learning speed.Finally, compare the dynamic neural network model based on wavelet analysis proposed by the author with the static neural network based on wavelet analysis and the traditional BP network of these two prediction models multi-angularly. It is proved that the model proposed in this paper is not only effective and feasible, but also significantly better than the other two models in terms of the prediction accuracy and the training rate.