As part of an effort to develop a systematic methodology for earthquake forecasts, we use a simple model of seismicity based on interacting events which may trigger a cascade of earthquakes, well-known as the Epidemic-Type Aftershock Sequence model). The ETAS model is constructed on a (bare) Omori's law, the Gutenberg-Richter law and the idea that large events trigger more numerous aftershocks. We offer an analytical approach to account for the yet unobserved triggered seismicity adapted to the problem of forecasting future seismic rates at varying horizons from the present. Tests presented on synthetic catalogs validate strongly the importance of taking into account all the cascades of still unobserved triggered events in order to predict correctly the future level of seismicity beyond a few minutes. We find a very strong predictability gain if one accepts to predict less than typically 25% of the large-magnitude targets. However, the probability gains degrade fast when one attempts to predict more than 30% of the targets. This results from the fact that a significant fraction of events remain uncorrelated from past seismicity. This delineates the fundamental limits underlying forecast skills, stemming from an intrinsic stochastic component in these interacting triggered seismicity models.