Flood-Forecasting AI Underestimates the Biggest Floods, Swiss Test Finds

Researchers at the Karlsruhe Institute of Technology report that a standard machine-learning model for turning rainfall into river flow underestimated 97% of the highest flows in a Swiss test set, the top 0.05% of the record. A version of the same model that predicts a range of possible flows rather than a single number placed 83% of those events inside its 99th-percentile range.

The models at issue are LSTMs, networks that learn from sequences, which the authors describe as the current benchmark for turning rainfall into runoff. The point of the paper is that this standard form is weakest at exactly the flows that cause floods, and that the reason sits in how it is trained: asked for one number, it returns something near the middle of the flows a storm could plausibly produce, while the real spread runs out in a long tail toward rare, large events.

The study, published Oct. 16, 2026, in Hydrology and Earth System Sciences, drew on daily records for 196 Swiss catchments in the public CAMELS-CH dataset. Sanika Baste and Ralf Loritz of KIT, with colleagues at Marburg University and the Interdisciplinary Transformation University Austria, set the single-number model against several that output a full range. The paper found that those were imperfect too, with each one's spread off in its own way.

A forecast range still has to become a decision, the authors write, because a warning is either issued at some level or it is not. So the team built that step as well: a reinforcement-learning agent it calls FRiCA, which reads a week of forecast ranges and issues a graded warning, from a flood expected every two years to one expected every 300. The authors report it beating fixed rules, such as warning off the single-number forecast. They trained and tested it on historical Swiss data, and it is not in operational use.

Their conclusion is that the limit on extreme-flood prediction may lie less in the design of the network than in what the training data contains. The paper is open access, and the catchment dataset is public.