|
The stock market has a high-yield and high-risk characteristics coexist on stock market analysis and prediction research has been paying attention. But the stock market influenced by political, economic, and many other factors, its internal law is very complex, resulting in many traditional forecasting methods are not satisfactory. In recent years, the rapid development of artificial neural networks for the modeling and forecasting the stock market provides a new technology and methods. Artificial neural network is a nonlinear science, it has a strong fault tolerance, self-adaptive and nonlinear mapping ability of artificial intelligence methods to overcome the traditional shortcomings in information processing, so that experts in the nervous system, pattern recognition , intelligent control, combinatorial optimization, forecasting and other fields has been successfully applied. This article aims to analyze the characteristics of the stock market, grasp its variation, and make better prediction for the stock market regulator regulate the stock market, investors investment strategy and government departments regulate the development of the stock market provide a reliable basis. BP network is a multilayer feedforward neural network capable of complex nonlinear mappings. But its weakness is slow convergence, network vulnerable to fall into local extreme point, and the initial weights, thresholds, and no basis for the choice of the network structure, has great randomness, which are affecting the BP neural network generalization capability. Paper introduces the momentum method and adaptive learning rate method, discussed the network topology, the number of hidden nodes in the principles, the sample data selection and pretreatment, the initial determination of parameters and so on. To avoid the network falling into local minimum point and improve network convergence speed genetic algorithm proposed by the GA-BP neural network algorithm. Optimize BP neural network initial weights, threshold, determining a better search space, instead of randomly selected initial weights, threshold, and then in the solution space to train the network to learn to converge, search out the optimal solution. RBF neural network is a novel and effective feedforward neural network, which has the best approximation of the performance and characteristics of the global optimum, and the structure is simple, fast training. Therefore, in the stock prediction has its unique advantages. In the RBF neural network, the number of hidden layers and the position of the center of the selection key to the performance of the entire network, the network directly IMPACT approximation ability. GRNN is a branch of RBF neural network, the network input layer node only passes the input signal to the hidden layer, hidden layer nodes by a Gaussian function such as the role of radial functions constitute the output layer node is usually a simple linear function. The relationship between its data acquisition method, unlike interpolation and fitting, in the same structure can be directly sampled or calculated to modify data on the network, no need to recalculate parameters. So, GRNN has fast calculation, the results are stable, less artificially selected parameters and other characteristics. It also determines the GRNN to maximize avoid human subjective assumptions on the prediction results. This article describes the various existing stock market forecasting methods and neural network structures and algorithms to select the most representative stocks on the Shanghai Composite Index and hops, respectively, with BP neural network, GA-BP network, RBF network and GRNN four models were rolling forecast. Numerical results show that the predicted value and the actual value of the basic agreement, the neural network is used to predict the stock market is feasible and effective, with good prospects.
|