As a new learning method, Support Vector Regression (SVR) has good generalization and prediction performance for time series modeling and predicting. In the course of SVR modeling, parameter choosing is very important to the accuracy of models. Aimed at problems in parameter optimization of SVR models, the paper proposes an SVR parameter choosing method for time series prediction, which improves the traditional cross-validation according to the features of time series prediction and sufficiently mines information included in numbered samples on the basis of maintaining the direction characteristic of time series. Furthermore, it is combined with (-weighted SVR in order to get good model parameters. Experimental results over typical time series show the validity of the parameter choosing method of SVR. The method gets good effect applied to time series prediction.