Kernel Comparison for Time Series Forecasting Using SVR

Abstract

Forecasting is considered a crucial topic in time series analysis due to its role in supporting future decision-making through the use of statistical models to predict patterns and trends. This study aimed to employ the Support Vector Regression (SVR) model to forecast nonlinear time series using five kernel functions: Linear, Polynomial, Radial Basis Function (RBF), and Sigmoid, by generating simulated data that mimic air pollution data. This approach was chosen because such environmental time series are often nonlinear and contain outliers, making traditional models less efficient for prediction. Therefore, the performance of the five kernels within the SVR model was compared using statistical evaluation metrics: MSE, MAE, MAPE, and R². Simulation results showed the superiority of the RBF kernel over the other kernels across different sample sizes (2000, 4000, 6000, 8000), indicating its better capability to represent nonlinear relationships and improve prediction accuracy. The study concluded that the SVR model can efficiently handle nonlinear data compared to traditional models, with a recommendation to use SVR with the RBF kernel in environmental applications such as air pollutants and energy production.