Five National Fire Danger Rating System indices (including the Ignition Component, Energy Release Component, Burning Index, Spread Component, and the Keetch–Byram Drought Index) and the Fosberg Fire Weather Index are used to characterise US fire danger. These fire danger indices and input meteorological variables, including temperature, relative humidity, precipitation, cloud cover and wind speed, can be skilfully predicted at weekly to seasonal time scales by a global to regional dynamical prediction system modified from the National Centers for Environmental Prediction’s Coupled Forecast System. The System generates global and regional spectral model ensemble forecasts, which in turn provide required input meteorological variables for fire danger. Seven-month US regional forecasts were generated every month from 1982 to 2007. This study shows that coarse-scale global predictions were more skilful than persistence, and fine-scale regional model predictions were more skilful than global predictions. The fire indices were better related to fire counts and area burned than meteorological variables, although relative humidity and temperature were useful predictors of fire characteristics.