Forecasting of Time Series Data Using Support Vector Resgression based on Kernel Types

Abstract

Support vector regression (SVR) is a nonlinear method that uses the principle of risk minimization. SVR model uses the kernel function with regularization constant (C and ε). The kernel function which is commonly used in the SVR are linear, polynomial and radial basis function (RBF). Determination of kernel with its constants is required in SVR model. Grid search algorithm determines the best regularization constant combination based on the smallest error rate. SVR is able to replace the case of linear regression and forecasting. The accuracy of forecasting is very important. The type of time series data is divided into several models, i.e. constant, trend, polynomial, and seasonal model. The study simulated various model of time series data using SVR. Model that had been tested was evaluated by Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). The simulation found that SVR model which used RBF kernel on the four types of time series data performed the best forecasting accuracy. In the application data, the RBF kernel with C of 10 and ε of 0.04 was implemented to build the SVR model. In the training data, the best SVR model produced MAE of 7.15436, RMSE of 10.03607 and MAPE of 0.86904%, while in the testing data the model produced MAE of 6.67387, RMSE of 9.41013 and MAPE of 0.865755%. Even if the testing data produced smaller error compared to training data, the difference of errors in the training and testing data was not statistically significant.