In machine learning, selecting useful features and rejecting redundant features is the prerequisite for better modeling and prediction. In this paper, we first study representative feature selection methods based on correlation analysis, and demonstrate that they do not work well for time series though they can work well for static systems. Then, theoretical analysis for linear time series is carried out to show why they fail. Based on these observations, we propose a new correlation-based feature selection method. Our main idea is that the features highly correlated with progressive response while lowly correlated with other features should be selected, and for groups of selected features with similar residuals, the one with a smaller number of features should be selected. For linear and nonlinear time series, the proposed method yields high accuracy in both feature selection and feature rejection.