Application Of ε-SVR On Time Series Prediction

Support Vector Machine (SVM) is a new machine learning method based on Statistical Learning Theory which is developed by Vapnik and his co-workers, and can be use under small samples. Because the idea of Structural Risk Minimization is introduced, SVM has much better learning ability and solves many difficult problems such as dimension disaster, small samples, non-linear model, local extremum and so on, so it has already become popular in the field of data mining. So far SVM are mainly applied in classification and regression, for it was firstly used in the former, and receive excellent performance, research of the theories and practical applications in this field have being well developed, but there are still many different applications in regression worth our study.In this thesis, the main content of Statistical Learning Theory is firstly introduced briefly, based on this, the basic principle and process ofε-SVR (one algorithm of Support Vector Machine for Regression, SVR) is presented. Then using this method to model building and predict two different time series data (Shenzhen monthly tourist quantity data and Chengdu monthly price index data), to predict the former, two different kernel functions are employed, and the former's performance is evidently better than the latter's.ε-SVR's performance is also compared with that of traditional time series analysis method, and the former outperform the latter. Further more, comparing theε-SVR's performance of Chengdu monthly price index data prediction with that of two different existing algorithms to illustrateε-SVR's validity in the regression and prediction of time series data.