The Canadian Lightning Detection Network provides continuous detection to about 65°N in the west and 55°N in the east. Coverage in the south is melded with the U.S. network to 35°N east of 100°W and to 40°N west of 100°W for use in Canadian weather offices. Studies of 1998-2000 lightning "climatology" by Burrows et. al. (2002) and Orville et al. (2002) revealed strong latitudinal, seasonal, and diurnal dependencies, and significant influence by topography and land-water boundaries. Thus, lightning is a highly non-linear phenomena and will require complex models to predict its occurrence. Statistical models to predict lightning occurrence probability in 3-hour intervals to 48 hours for the north portion of the NALDN were developed and implemented in experimental mode May 01 2003 at the Canadian Meteorological Center (CMC) for potential use in the national automated public forecast production system there. Graphical depiction of forecasts were put up in real time on an internal website for use by Meteorological Service of Canada forecasters. Models for 5 degree latitude-longitude boxes were built for each of the four 3-hour intervals from 0000 UTC and 1200 UTC, giving a model for each of the 8 three-hour intervals in a 24-hour period. The same models are used in any 24-hr period. Grid point resolution is currently 22 km. To make a training data set, the predictand, 3-hour total "lightning report density", was matched with 112 potential predictors associated with moisture, convection, lift, and climatological controls for May to September 2000 and 2001. Predictor types are: temperature, dew point, and geopotential height at several levels; CAPE; convective stability indices; a severe storm index; helicity; elevation; convective cloud depth; Price and Rind predicted flash rate; tropopause height and temperature; precipitable water (total and above 700 mb); thickness of 4 layers; total and convective rainfall; wet bulb potential temperature; 700 hPa vertical motion; land/water and vegetation designations. Meteorological predictors were derived from archived 6-hr and 12-hr forecasts of the CMC’s GEM weather prediction model. For every three-hour period, the average, maximum (or minimum) value, and change of meteorological potential predictors is found. Tree-based models are derived with Classification and Regression Trees (CART) (Brieman et al., 1984). Verification of lightning probability forecasts for 2002 has been done and is underway for 2003 at the time of writing. The main method for verification is the "relative operating characteristic" (ROC) (Mason, 1982) applied to forecasts of lightning probability and observed lightning report density at all grid points at each projection time. Results show that overall, the forecasts are reasonably accurate for location of regions of elevated probability. Forecast probability values are less well correlated with observed lightning frequency when 1-grid-point verification squares are used, but are well correlated with observed lightning frequency when 5-grid-point verification squares are used.