Deep-learning streamflow forecasters often use ensemble spread as a proxy for uncertainty. In a new arXiv preprint, Justin Gallant and Tyler Wilson test whether that proxy stays reliable as flows rise, using LSTM ensembles and observed streamflow data. They find that predictive error is roughly twice the corrected spread at ordinary flows and up to four times the spread at the highest flows, with the mismatch growing as a power law (log-log slope 1.30, 95% CI [1.20, 1.41]).
Most striking is behavior beyond the training record: for observed extreme flows, the discrepancy reaches 18 times. The authors say this under-dispersion is pervasive and not explained by conditional mean bias. The practical danger is silent failure: the ensemble ranks errors almost as well as an oracle (Area Under the Sparsification Error 0.071), and simply ordering by predicted discharge comes close to matching that. So the spread still tells you when the model is likely wrong, but it badly understates how wrong.
Because error ranking cannot be verified in deployment, operators may rely on the available spread as the uncertainty signal under increasingly extreme conditions. The paper argues this systematically understates risk, most severely for the extreme events that matter most.