The Study Of Combinational Forecasting Based On Neural Network And Principal Component Analysis

Stock market is financial market, which associated with the people's lives, social stability and economic development, How to effectively analyze and forecast the stock market's trend have been a problem of concern to the people, Now many theory methods and techniques have been put forward .In these methods, Artificial neural network, because of its unique structure, the method of processing information, and the features of highly Parallel, distributed Storage, have particularly been suited to dealing with uncertain fuzzy information and the question needed to considering conditions of many factors at the same time, Therefore, Artificial neural network have been widely attention in the stock market analysis.But, for complex issues and high-dimensional input variables, Artificial neural network's direct prediction will bring dramatic increase in network size, increase in computing time, decrease in the convergence of the network and the ability of generalization, On the other hand, predictors' relevance leads to information overlap, decrease Model prediction's accuracy. so we must Pre-treat the prediction factor.For forecasting closing price in stock market, we firstly divides samples which have characteristic of decentralization into different sub-classes with the aid of SOM neural network. BP neural network forecasting models will be built To the different sub-classes, Corresponding to the different sub-classes; For improving forecast accuracy and the efficiency of network training, we uses principle components analysis method to reduce the dimensionality of the feature space. In actual forecasting, we must classify real-time data, choice the corresponding model to predict network output. Presents a combinational prediction model based on neural network and principle components analysis(PCA),which is applied to real-time prediction of economic-technical indexes in stock prices. In order to verify effectiveness of the Presented model, the following three models will be built:(1) Samples without treatment will be put into BP neural network model to be trained and be forecasted.(2) Before training samples with BP network, we firstly divides samples which have characteristic of decentralization into different sub-classes with the aid of som neural network, and then BP neural network forecasting models will be built To the different sub-classes.(3) we uses principle components analysis method to handle sub-classes, then put the results into BP network model.Then we put the results into BP neural network model as importation. Through comparison of training time and step and forecasting accuracy among three models, the results show that the efficiency of network training and forecasting accuracy based on the Presented model is higher than others.