Flood-loss estimates are needed for floodplain development and mitigation projects, for setting fair insurance rates, and for guiding climate adaptation policy. Currently, flood-loss models, including depth-damage functions (DDFs) widely used in the U.S., lack empirical validation commensurate with the geographic extent and diversity of structures and flood exposure over which these predictions are needed. Using data from 845,776 U.S. National Flood Insurance Program claims, we validate DDFs and create alternative models grounded in empirical data and validation. These alternative models more accurately predict average observed damages for many types of structures and hazard compared to current DDFs which omit important variables and interactions that drive observed losses. We find that a major bottleneck in flood-loss estimation is the development and validation of flood-loss models for both damaged and undamaged homes, a gap FEMA could help close.