Accurate and reliable streamflow forecasting is important in hydrology and water resources planning and management. In the present work, wavelet-based direct (DF) and multi-component (MF) forecast methods performed by the à trous algorithm (AT) are proposed for both deterministic and stochastic monthly streamflow prediction improvement. They are developed in the case of the one-month lead streamflow prediction of the East River basin in China, and then compared with the benchmarks that are implemented without wavelet transform so as to evaluate the effectiveness for forecasting accuracy improvement. An existing blueprint that is flexible and practical to incorporate various sources of forecast uncertainty is extended to generate the stochastic probability prediction of streamflow. Partial mutual information is adopted for predictors selection, and six kinds of Extreme learning machine (i.e. one linear ELM and five common nonlinear kinds) are separately used as the learning algorithms coupled with the wavelet-based forecast methods to conduct a comprehensive performance evaluation. The comparison results indicate that both DF and MF can effectively increase the point prediction accuracy of monthly streamflow under deterministic and stochastic forecasting conditions, while MF performs better than DF. For stochastic prediction, it is much more reasonable to consider both parameter and model error uncertainties than just to consider only parameter uncertainty, and with the reasonable setting MF method can significantly improve the probabilistic interval prediction by greatly improving the forecast sharpness. It can be concluded that the approach using AT wavelet-based DF or MF could provide a feasible way for streamflow prediction improvement.