Accurate forecasting of daily streamflow is essential for water resource planning and management. As a typical non-stationary time series, it is difficult to avoid the effects of noise in the hydrological data. In this study, the wavelet threshold de-noising method was applied to pre-process daily flow data from a small forested basin. The key factors influencing the de-noising results, such as the mother wavelet type, decomposition level, and threshold functions, were examined and determined according to the signal to noise ratio and mean square error. Then, three mathematical techniques, including an optimized back-propagation neural network (BPNN), optimized support vector regression (SVR), and adaptive neuro-fuzzy inference system (ANFIS), were used to predict the daily streamflow based on raw data and wavelet de-noising data. The performance of the three models indicated that a wavelet de-noised time series could improve the forecasting accuracy. The SVR showed a better overall performance than BPNN and ANFIS during both the training and validating periods. However, the estimation of low flow and peak flow indicated that ANFIS performed best in the prediction of low flow and that SVR was slightly superior to the others for forecasting peak flow.