The development of a high-quality wildfire occurrence model is an essential component in mapping present wildfire risk, and in projecting future wildfire dynamics with climate and land-use change. Here, we develop a new model for predicting the daily probability of wildfire occurrence at 0.1° (∼10 km) spatial resolution by adapting a generalised linear modelling (GLM) approach to include improvements to the variable selection procedure, identification of the range over which specific predictors are influential, and the minimisation of compression, applied in an ensemble of model runs. We develop and test the model using data from the contiguous United States. The ensemble performed well in predicting the mean geospatial patterns of fire occurrence, the interannual variability in the number of fires, and the regional variation in the seasonal cycle of wildfire. Model runs gave an area under the receiver operating characteristic curve (AUC) of 0.85–0.88, indicating good predictive power. The ensemble of runs provides insight into the key predictors for wildfire occurrence in the contiguous United States. The methodology, though developed for the United States, is globally implementable.