A systematic method based on Taguchi's experimental design approach is proposed for selecting input variables for an inferential predictor. Several implementation difficulties arising from dynamic variation and correlation among process variables are addressed. The predictor is developed using support vector regression (SVR) in order to capture the nonlinearity in the process. The prediction performance of the proposed Taguchi-SVR is compared with the existing variable importance in projection (VIP)'SVR method. The industrial case study clearly indicates that the proposed methodology can be a valuable tool for process variable selection and it can improve the prediction performance of the inferential predictor.