Prediction model for high-volatile time series based on SVM regression approach

This paper deals with modelling high-volatile time series using modern machine learning technique called Support Vector Regression. After discussing the basic principles of Support Vector Machines (SVM), we construct SVM Regression Prediction Model. Afterwards, this prediction SVR model is applied to oil prices. Due to high-volatile and dynamic character these data are very difficult to model. Experiments are performed on three different time periods of crude oil data in order to find out the best conclusions for high-volatile series. Prediction accuracy of our SVR (Support Vector Regression) models, which is measured by MAPE numerical characteristic, is compared to standard approach which is represented by statistical Box-Jenkins models. Evaluating and discussing the results we find out that SVM models are more accurate than statistical Box-Jenkins models. Moreover, linear kernel function is best for modelling short-time high-volatile series and radial functions are much more effective in modelling long-time high-volatile series than kernel SVM functions.