<p>Regional runoff prediction plays a crucial role in enhancing water resource management and ensuring water security in ungauged basins. This study selects 40 typical watersheds in major river systems across China. We evaluate the performance of a traditional hydrological model (HMETS) and the Long Short-Term Memory (LSTM) network for locally calibrated modeling and runoff prediction in ungauged basins. Furthermore, we investigate training strategies and optimal approaches for regional LSTM models. Key findings include: (1) Both HMETS and LSTM show similar performance in locally calibrated models, yet both fail to capture runoff generation mechanisms under significant anthropogenic disturbances; (2) Among regional parameterization approaches, the spatial proximity-result averaging method performs well in undisturbed basins, but generally fails in human-dominated basins; (3) LSTM models demonstrate strong capabilities in identifying the underlying relationships between precipitation and runoff, achieving superior prediction accuracy in ungauged basins compared to traditional hydrological methods. The proposed training strategy further enhances cross-basin generalization performance; 4) Overall, model performance shows a nonlinear trend, initially increasing and then decreasing, as the number of donor basins increases. This highlights the importance of basin similarity, especially the alignment of human activity patterns between human-dominated basins. These findings provide valuable insights for runoff prediction in ungauged regions across China.</p>