Support vector regression (SVR) adapts the support-vector framework to continuous outcomes. It fits a function inside an epsilon-insensitive tube: prediction errors within the tolerance ε incur no penalty, while observations on or beyond the tube’s boundaries may become support vectors. section 8.8.1
SVR favors a restrained function shaped by consequential observations rather than responding equally to every training error. It seeks simplicity while keeping most errors within the tolerance; kernels can represent nonlinear relationships without explicitly constructing transformed features. Vapnik extended the framework to regression. section 8.8.1

