The Research And Improvement Of Probabilistic Neural Network And Their Applications On Stock Forecasting

Using the prior probabilities of the samples-Bayes Theory and Bayes decision rule based on risk minimum principle could classify the new samples, and then bring out Probabilistic Neural Network(PNN). This Neural Network, based on statistics theories, has shorter training time and is not often to be trapped in local maximum. However, its network architecture limits it only to be able to output the classification result.Focus on it, this paper connects PNN with Radial Basis Function Neural Network (RBFNN), and forms a new network architecture-Radial Basis Function-Probabilistic Mixed Neural Network(RBF-PMNN). It not only has the sorter characteristic of PNN, but also has the function fitting characteristic of RBFNN. On the other hand, because the RBF-PMNN network performance is fluenced by the network parameters, this paper separately uses EM algorithm and Genetic algorithm to optimize parameters: EM algorithm uses the characteristic of the second level which cannot be observed to constructe EM-RBF-PMNN; while Genetic algorithm picks up the center vectors and smoothing factor through global optimization searching to constructe GA-RBF-PMNN.Moreover, in order to confirm the network performance, the paper separately utilizes five models (PNN, RBFNN, RBF-PMNN, EM-RBF-PMNN and GA-RBf-PMNN) to build up stock forecasting model which based on volatility, and uses the training data and the testing data to confirm these network models are effective. The result is Satisfying.