Floods are some of the most damaging natural hazards, but they remain difficult to predict, especially the most extreme events. Machine learning models, like long short-term memory (LSTM) networks, are increasingly used to forecast streamflow, but they often miss the biggest floods because these events are rare and behave differently than typical flows. In this study, we test a variety of modeling strategies to help LSTM models better predict these rare, high-flow events without losing accuracy for more common conditions. Using data from hundreds of U.S. rivers and weather records, we compare techniques like rebalancing the training data, changing the way model errors are calculated, and reshaping the distribution of streamflow values before training. We find that some approaches—especially adjusting the distribution of target values and using weighted error functions—can significantly improve flood predictions. However, others, like removing low flows from training, can actually hurt model performance. These findings help researchers and practitioners design better machine learning models for flood forecasting, especially as climate change increases the risk of extreme weather.