Support vector regression (SVR) has often been applied in the prediction of time series with many characteristics such as large sample sizes, high noise, non-stationary, non-linearity. Due to much time consumption of global SVR, local methods come into existence as the situation requires. We introduce grey local SVR (GL-SVR) combined grey relational grade regarded as neighbourhood function with local support vector regression. To optimize the machine, based on leave-one-out errors, pattern search method is adopted for model selection. Experiments are carried out on a real stock price change forecasting with GL-SVR and the results demonstrate that Our approach can not only speed up the computing speed, but also improve the prediction accuracy.