A Study On The Prediction Of The Price In Stock Market Based On SVR

With the fast development of economic and society, as well as with the growth of the investment consciousness, more and more people are concerned with the trade of stock. The investment in stock market has become an important part of people's life. But the high return in stock market always comes with high-risk. So the prediction of the stock market has an extremely important value. But in the stock market, there are a large number of indicators and lots of the information. How to find the critical factors in stock market which has a complicated structure and many factors? So it is difficult to find its inherent pattern of the stock market. The existing forecasting methods usually have not a very good result. Support Vector Machine is an emerging technology, by means of optimization theory, duality theory, using structural risk minimization, nuclear function and other methods to solve the classification problem, has a better regression and generalization performance. Recently it makes a rapid progress in the research of regression problems. So SVR (Support Vector Regression) is emerged which is based on SVM. In recent years some scholars make an application of it in stock market to predict the price of the stock for its good performance. They all make a good result.This paper thinks about how to connect the characteristics of the stock with SVR and how to use of the information more effectively, and makes a model based on SVR. Also we make a test with the data of China Unicom (600050) and NEW WELLFUL (600975). This paper analyzes the existing forecasting methods firstly and discusses their advantages and disadvantages. Then propose the method based on SVR to predict the stock price. In order to make a full use of information, we collect data as much as possible, and make a linear correlation analysis by using correlation coefficient; then make a simple non-linear analysis by using scatter plots. Then use sensitivity analysis to identify the factors which have little correlation with the target, in order to achieve the proposal of finding the critical information. Secondly process the data by the means of PCA (principal component analysis), which can help to diminish the dimensions and remove the correlation among the factors. In order to make the model adapt the characteristics of the stock market, use SOFM (self-organizing feature map) neural network to cluster the samples in order to expand the impact of those non-frequent modes and weaken the impact of the long-term trend. Finally make three groups of tests based on the history data of China Unicom and NEW WELLFUL. The first model is based on the original data. The second model is based on the factors which are analyzed by the correlation analysis and sensitivity analysis. The third model is based on the samples clustered by SOFM. The results show that the third model has the best forecasting performance and the first model has the worst performance. These prove that the improved model has achieved the target of increasing the forecasting accuracy.Experiments show that SVR can predict stock prices effectively. And optimizing the choice of critical variables makes a better use of information. Also the cluster based on SOFM has improved the defects that the impact of long-term trend mode weakens the impact of non-frequent modes. The final prediction has also been significantly improved.