The maxi-min margin machine (M4) algorithm, contrast to the traditional support vector machine (SVM) algorithm, gives a more robust solution and gets better generalization performance. In this paper we extend the M4 classification algorithm to deal with regression problem, and propose a novel regression method. This method inherits the characteristics of M4 such as good robustness and generalization performance. In this paper we discuss the linear and nonlinear case of the proposed method, and experimental results indicate its effectiveness and better robustness and generalization performance compared with the traditional SVR algorithm.