Forecasting the full distribution of the number of earthquakes is revealed to be inherently superior to forecasting their mean. Forecasting the full distribution of earthquake numbers is also shown to yield robust projections in the presence of surprise large earthquakes, which in the past have strongly deteriorated the scores of existing models. We show this with pseudoprospective experiments on synthetic as well as real data from the Advanced National Seismic System database for California, with earthquakes with magnitude larger than 2.95 that occurred between the period 1971 and 2016. Our results call in question the testing methodology of the Collaboratory for the Study of Earthquake Predictability (CSEP), which amounts to assuming a Poisson distribution of earthquake numbers, which is known to be a poor representation of the heavy‐tailed distribution of earthquake numbers. Using a spatially varying epidemic‐type aftershock sequence (ETAS) model, we demonstrate a remarkable stability of the forecasting performance, when using the full distribution of earthquake numbers for the forecasts, even in the presence of large earthquakes such as 7.1 Hector Mine, 7.2 El Mayor–Cucapah, 6.6 Sam Simeon earthquakes, or in the presence of intense swarm activity in northwest Nevada in 2014. Although our results have been derived for ETAS‐type models, we propose that all earthquake forecasting models of any type should embrace the full distribution of earthquake numbers, such that their true forecasting potential is revealed.