No Free Lunch? Improving LSTM Flood Predictions with Minimal Loss in Overall Skill

Extreme runoff events are among the most impactful hydrologic phenomena, yet remain difficult to predict due to their rarity, the skewed distribution of streamflow, dependence on complex antecedent conditions, and uncertainty under future climate pathways. While deep learning models, particularly long short-term memory (LSTM) networks, have shown promise in hydrologic prediction, they often underpredict peak flows. Recent research has also identified a theoretical maximum prediction limit that precludes these models from extrapolating to flood events outside the range of training data. This study evaluates a suite of modeling strategies designed to improve LSTM performance on extreme runoff events, with an emphasis on understanding the tradeoffs between extreme value predictive skill and overall model performance. We train LSTM models on a large sample of US stream gauges using gridded meteorological forcing and static watershed attributes. We implement and compare several techniques, including undersampling of frequent flows, oversamping of infrequent flows, the use of different loss functions, target value distribution transformation, and combined pre-training approaches. We find that certain techniques—particularly transforming the distribution of target streamflow values and the use of weighted loss functions—have the potential to improve the prediction of high flows with minimal degradation to overall model performance. We also identify situations where well-intentioned modifications (e.g., undersampling) can degrade performance across nearly all metrics. These results offer practical guidance for hydrologists and modelers seeking to enhance the fidelity of machine learning models in predicting floods, and provide a foundation for future research into increasing the performance of flood forecast models for rare events.