Statistical models valid for each month May to September were developed to predict the probability of lightning occurrence in three-hour intervals for an area encompassing much of Canada and the northern United States. Lightning is an episodic phenomenon that generally occurs over only a small fraction of this large region in a three-hour period even on significant days in the warm season. Predictors were derived from output of the GEM numerical weather prediction model at the Canadian Meteorological Center. Statistical models were built for 5º latitude by 5º longitude sectors using tree-structured regression with data from 2000 and 2001 and run in real time in 2003. Reduction of initial predictand variance achieved by most of the models was in the range .4 to .7. Many predictors were required to successfully model lightning occurrence for this area. The highest ranked predictors overall were the Showalter index, mean sea-level pressure, and troposphere precipitable water. Three-hour changes of 500 hPa geopotential height, thickness, and mean sea level pressure were highly ranked predictors in most areas, likely since they are representative of frontal motion in temperate latitudes. The three-hour average of basic predictors was more important overall for prediction than the maximum value (or minimum when appropriate), probably since the predictand was derived from total lightning in three hours. CAPE, while necessary for convection, was outranked by several related predictors, indicating that it must appear with other predictors in successful statistical lightning prediction models. Verification of forecasts for 2003 in three-hour projection time intervals from 0-3 hours to 45-48 hours showed the statistical models predicted lightning occurrence with a good degree of success on most days.