<p>Runoff prediction is crucial for flood forecasting, and water resources planning and management. In recent years, Long Short-Term Memory (LSTM) models have been widely used in runoff prediction. There are different strategies that LSTM models can employ to predict runoff with multiple lead times. However, the effects of different multi-step-ahead prediction strategies on LSTM-based runoff prediction remain insufficiently explored. In this study, a systematic evaluation of the accuracy of LSTM models under three commonly used strategies, i.e., the multi-output, direct, and recursive strategies was carried out across multiple lead times and basins for the first time. An experimental study was conducted utilizing 48 basins, which were randomly chosen from the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset. The main conclusions are as follows: (1) The accuracy of the LSTM models under all three strategies generally declines as lead time increases. Specifically, the recursive strategy exhibits the most pronounced deterioration in accuracy, while the multi-output strategy demonstrates the least deterioration. (2) For lead times ranging from 1 to 7&#xa0;days, the LSTM models under the direct strategy generally achieve superior prediction accuracy compared to the models under the other two strategies. For 1- to 3-day lead times, the recursive strategy generally outperforms the multi-output strategy, while whereas the opposite trend emerges for 4- to 7-day lead times. (3) The difference in accuracy among the LSTM models under the three strategies generally shows an increasing trend as lead time extends. It indicates that the effects of LSTM model strategy selection on runoff prediction accuracy become more profound as lead time increases. These findings offer practical insights for selecting appropriate strategies when applying LSTM models to runoff prediction.</p>