We constructed 5- and 10-yr smoothed-seismicity forecasts of moderate-to-large California earthquakes, and we examined the importance of several assumptions and choices. To do this, we divided the available catalog into learning and testing periods and optimized parameters to best predict earthquakes in the testing period. Fourteen different 5-yr testing periods were considered, in which the number of earthquakes varies from 18 to 63. We then compared the likelihood gain per target earthquake for the various choices. In this study, we assumed that the spatial, temporal, and magnitude distributions were independent of one another, so that the joint probability distribution could be factored into those three components. We compared several disjoint test periods of the same length to determine the variability of the likelihood gain. The variability is large enough to mask the effects of some modeling choices. Stochastic declustering of the learning catalog produced a significantly better forecast, and representing larger earthquakes by their rupture surfaces provided a slightly better result, all other choices being equal. Inclusion of historical earthquakes and the use of an anisotropic smoothing kernel based on focal mechanisms failed to improve the forecast consistently. We chose a lower threshold magnitude of 4.7 for our learning catalog so that our results could be compared in the future to other forecasts relying on shorter catalogs with a smaller magnitude threshold.