As one of the most advanced ensemble learning technologies, Stacking is an important way to improve the performance of runoff prediction. The existing researches of Stacking-based runoff prediction mostly focus on the accuracy evaluation under few basins and few lead times. Applicability evaluation and influencing factors analysis under multiple basins and multiple lead times remains unexplored. In this study, Support Vector Regression (SVR) and Random Forest (RF) were used as individual learners, and Ridge Regression was used as a meta-learner, and runoff prediction models based on Stacking were constructed. Taking 200 basins in CAMELS dataset as the study area, and taking 1~7 days as the lead times, the accuracy, stability and applicability of Stacking-based runoff prediction were systematically evaluated, and the correlation between the effectiveness of Stacking and the characteristics of basins and the accuracy of individual learners were analyzed. The main results are as follows: (1) The overall accuracy and stability of Stacking are higher than those of the individual learners. (2) Stacking can improve the accuracy of runoff prediction in most basins in the continental United States. The improvement effect is more significant in the basins with heavy precipitation and high temperature, but the effect is relatively limited in the basins with light precipitation and low temperature. (3) Stacking tends to improve prediction accuracy in the basins with low accuracy of the individual learners, but it is difficult to improve the prediction accuracy in the basins with high accuracy of the individual learners. This study can provide a reference for the application of Stacking in runoff prediction.