Improvements in modelling power and input data have vastly improved the precision of physical flood models, but translation into economic outputs requires depth–damage functions that are inadequately verified.
In particular, flood damage is widely assumed to increase monotonically with water depth. Here, we assess flood vulnerability in the US using >2 million claims from the National Flood Insurance Program (NFIP). NFIP claims data are messy, but the size of the dataset provides powerful empirical tests of damage patterns and modelling approaches. We show that current depth–damage functions consist of disparate relationships that match poorly with observations. Observed flood losses are not monotonic functions of depth, but instead better follow a beta function, with bimodal distributions for different water depths. Uncertainty in flood losses has been called the main bottleneck in flood risk studies, an obstacle that may be remedied using large-scale empirical flood damage data.