A clustering-based selective support vector machine ensemble forecasting model is presented. For improving the generalization ability of support vector machine ensemble, a hybrid clustering algorithm which combines the SOM and K-means algorithm is used to select the most accurate individual support vector machine from every cluster for ensembling, which ensures accuracy of individual support vector machines and improves the diversity of the individual support vector machines. This method can improve support vector machine ensemble generalization ability effectively with low cost. To illustrate the performance of the proposed forecasting model, simulations on chaotic time series prediction of the Mackey-Glass time series and the time series generated by the Lorenz systems are performed. The results show that the chaotic time series are accurately predicted, which demonstrates the effectiveness of this method.