Abstract: This study introduces Support Vector Regression (SVR) model and predicts on time series based on SVR, proposing several new approaches which improve traditional SVR in order to enhance prediction accuracy. In terms of the feature of time series, the approaches give different weights to different history data at different times and make the punishing coefficient C and non-sensitive loss ε in optimization objective function of SVR adjustable along with different sample data. The experimental results show that the proposed approaches improve the prediction precision and testify the validity of these approaches.