In this study, a triple-stage support vector regression (SVR) based neural network ensemble forecasting model is proposed for foreign exchange rates forecasting. In the first stage, multiple single neural predictors are generated in terms of diversification. In the second stage, an appropriate number of neural predictors are selected as ensemble members from the considerable number of candidate predictors generated by the previous phase. In the final stage, the selected neural predictors are combined into an aggregated output in a nonlinear way based on the support vector regression principle. For further illustration, four typical foreign exchange rate series are used for testing. Empirical results obtained reveal that the proposed nonlinear neural network ensemble model can improve the performance of foreign exchange rates forecasting.