Application Of Bayesian Neural Network On Stock Forecasting

Stock prediction is an important subject in economic field. Because stock market has high noise and strong nonlinear characteristics, it is difficult to set up an accurate mathematical model. Currently we mainly adopt the method of Neural Network, however, it mainly has the following weakness: difficult to determine the dimension of input vectors, converging local minimum points, having too long training time, generalization ability weak and difficult to capture the stock market dark horse, etc. In these disadvantages, generalization ability weak and dark strands capture difficulty is major drawback. Aiming at these disadvantages, this paper has done several works as follows:1. Used phase space reconstruction to determine the Bayesian Neural Network input vector dimension. In the medium-term prediction, the length selection has always been a difficulty. This paper adopts the phase space reconstruction chaotic to determine the dimension network input vector and Bayesian Regularization Neural Network, which can effectively improve the network generalization ability. This method has comprehensive the chaos theory, neural network and Bayesian regularization methods, which not only takes into the chaos of the Shanghai Composite Index and the nonlinear relationship between phase space points, but also improving the network generalization ability. The forecasting analysis in second chapter shows that the prediction improved 8% around by this method than ordinary methods.2. Establish short-term forecast neural network model. In the stock market short-term prediction, we established two network models: Bayesian Neural Network can effectively improve the network generalization ability and automatically reduction the connected weights. Because the data contains original numerical and technical indexes, we established a parallel structure network, which can effectively reduce the coupling between two type data and simplify network parameters. Through the analysis of the predicted results in third chapter can be found: the Bayesian Network of general structure than ordinary BP Network increased by 10% or so, and the Bayesian of parallel structure neural network than ordinary structure of Bayesian Networks increased around 7% or so.3. Establish neural network model to capture dark horse shares. Through to dark horse shares research, using the dark horse's technical characteristics before starting, we abstract the characteristics of dark horse shares. Than we establish Bayesian Neural Network model, training this network using the data. We will put the nonlinear mapping relation of dark horse shares in the network model. The simulation results show that the prediction accuracy is 78% around.