Support Vector Machine's AOSVR And It's Application To Stock Market Prediction

Ever since 45 years ago, when F.Rosenblatt pioneered the perception model, the theory of machine learning has been developed significantly. After entering 80's, a so-called Back-Propagation method even more attracted researcher's attention on machine learning. The invention of Support Vector Machine in 90's made people understand more about machine learning. In this article, the development of machine learning is reviewed in terms of methodology, which is often ignored by researchers. In addition, a brief introduction of Support Vector Machine is given.The computational problems in Support Vector Machine is mainly on its quadratic optimization problem. Since its large scale of data source, it differs from other common problems. In the second part of this article, five kinds of solutions on SVM's quadratic optimization are introduced. Some important issues are also proposed for further investigation.AOSVR is a newly developed online support vector regression algorithm, which is quite useful when dealing with time-series problems. However, AOSVR does not have a specific consideration on the SVR parameters and the parameter of the kernel function. Based on a profound research of the statistical properties of SVR, V. Cherkassy developed a simple parameter-selection method which has proved to be useful in the batch mode SVR. In the third part of this article, an algorithm combining AOSVR and V. Cherkassy's method is proposed , in order to enrich the ability of online SVR by online adapting the parameters according to the movement of the time series. Meanwhile, based on the characteristic of stock market, we use a "forgetting" bias to ignore to some extent the earlier data and to concentrate on the recent data.