Modeling And Forecasting Seasonal Time Series Using Seasonal Support Vector Regression

Support Vector Machine (SVM) is a new kind of machine learning algorithm based on Statistical Learning Theory. Because the idea of Structrual Risk Minimization is introduced, SVM has much better learning and generalization abilitiy than troditional Empirical Risk Minimization based methods. What's more, SVM has the ability to model small-sample and no-linear problems. Support Vector Regression (SVR) was introduced by Vapnik at the base ofε-non-sensitive loss function. Because of the same theoretic advantages as SVM, SVR has aroused researchers' interest since it was first introduced. So far, SVR has provieded promising results in many forecasting problems. However, the topic of forecasting seasonal time series using SVR has not yet carefully discussed.This article focuses on building seasonal SVR models and their application on real data sets. Evaluating the value of data preprocessing in improving SVR's forecasting performance is one of the article's most important issues. Specifically, this article is arranged as follows:First part of the article is the introduction, literature review, research purpose, contents and innovation points. Second, this study reviews two traditional seasonal time series forecasting methods-Exponential Smoothing and SARIMA. In the third part of the article, three seasonal SVR models (Raw Data-SVR, Seasaonal Differencing-SVR and Seasonal-SVR) and their modeling steps are discussed. After that, empirical findings are presented. The data sets used in this part are collected from real world. Finnally, this article draws the conclusion as follows:⑴The forecasting performance of Raw Data-SVR is not very good;⑵Through proper data preprocessing, SVR can achieve better forecasting performance;⑶The two data preprocessing approaches introduced in this article, which are seasonal differencing and seasonal adjustment combining proper data transformation, empirical findings show that the latter is more efficient than the former.;⑷Seasonal Differencing-SVR and Seasonal-SVR have better forecasting performance than traditional Exponential Smoothing and SARIMA models;⑸SVR has the ability to forecast under small-sample conditions. The Seasonal Differencing-SVR and Seasonal-SVR presented in this article can both achieve excellent out-of-sample forecasting performance.