To keep the electric power system reliable, grid operators procure reserve generation capacity to protect against generator failures and significant deviation from the load forecast. Current methods for determining reserve requirements use historical generator availability data (recorded as failure events) to compute the fraction of the time each unit in the power system was unavailable unexpectedly. These values are then combined using analytical or simulation methods to yield a distribution of available capacity. From this distribution, the reserve capacity needed to maintain a particular reliability target may be determined. Such an approach implicitly assumes that generator failures occur independently of one another and that generator availability is not seasonal. To test these assumptions, we process the more than two million event records reported to the Generating Availability Data System (GADS) database maintained by the North American Electric Reliability Corporation (NERC) between January 1, 2012 and December 31, 2015. This allows us to construct complete availability histories (hourly time series) for each of the approximately 8000 generating units reporting to GADS during this period. Using these time series, we find strong evidence of correlated failures in most regions, even when removing Hurricane Sandy and the exceptionally cold month of January 2014 from the data. We find that correlated failures occur in all seasons. We do not find evidence of seasonality but note that seasonal structure may emerge with more data. In addition we determine the distribution of unscheduled unavailable capacity, unscheduled derating magnitudes, event durations, event arrival probabilities, and mean time between failure (MTBF) and mean time to recovery (MTTR) values. In each case, we report fit parameters to facilitate use by practitioners. The distributions of unscheduled unavailable capacity in each region are reasonably well modeled by Weibull and lognormal distributions. We find statistically significant differences in mean time between failure for small and large units for three unit types when aggregating over regions. Finally we present time series of unavailable capacity from unscheduled, maintenance, and scheduled events. These may be used in conjunction with load data to directly study resource adequacy risks without assuming independent failures or constant availability. Our findings suggest that power system resource planners should consider correlated outages as they identify reliability and reserve capacity requirements.