We analyze the patterns of seismic activity which precede large events in a mechanical model of a fault. The model generates a statistical distribution of events similar to that observed for a single fault, with a scaling region consistent with the Gutenberg-Richter law at small and moderate magnitudes, and an excess of events at large magnitudes. We find only slight variation in the scaling behavior during a loading cycle. However, we do observe systematic variations in space and time of the overall rate of activity. In the model, the activity accelerates dramatically preceding a large event and is usually a maximum in the neighborhood of the future epicenter. These results are compared to California siesmicity data, where we find that activity patterns vary regionally. Looking at patterns of activity in the San Francisco Bay Area since 1948, we find an increase of activity on the Calaveras fault near San Jose beginning in the 1980s which, if our model is relevant, would forecast a large earthquake in the region. The 1989 Loma Prieta earthquatke occurred on the San Andreas fault within 30 km of the section of the Calaveras fault showing increased activity. ¿ American Geophysical Union 1992