Lightning causes one third of the 9000 wildfires that occur in Canada. Annually, these lightning-caused fires account for 90% of the area burned and cost Canadians at least 150 million dollars in suppression costs and values destroyed. Unlike the fires caused by human negligence, lightning-caused fires often occur in multiple numbers in remote locations. A modern fire control organization can suppress all of these fires while they are still small only if it has time to position sufficient suppression forces before the fires occur. Therefore, predicting the occurrence of lightning fires hours in advance is an essential component of a successful suppression strategy. This paper describes the method currently used to predict the daily number and location of lightning-caused fires. A network of automated lightning sensors provides the locations and numbers of cloud-to-ground lightning flashes. For each flash the appropriate weather, fuel type, and moisture data are combined with models of the ignition, smouldering, and detectability processes. The ignition model predicts the chance of a flash causing ignition. The detectability model forecasts the probability of a fire being visually detectable during the burning period. The smouldering model tells us the chances of a fire surviving overnight (usually in a smouldering state). Because fires can remain in a dormant state for long periods, each flash that occurred during the previous 10 days is considered a potential ignition point for the current day. Fires predicted to have been ignited up to 10 days earlier are given the opportunity to smoulder; they are removed from consideration after detection. Remaining fires combined with likely new fires and the expected number of detectable fires during the next burning period gives the number of fires predicted for that day. Evaluation results are presented and discussed. In general, the prediction program produces fair to good results for small to medium morning storms and medium to large overnight storms. As well, for the previous day, the smouldering/survival model seems to work well. Poor predictions are generated, however, from afternoon storms, from occasions when rainfall data is not available, and from the smouldering model for periods longer then two days. The prediction program is perhaps best thought of as being an expert system where specific knowledge of lightning physics, rainfall patterns, and fire behavior are combined with expert opinions of the various lightning fire occurrence processes. There is still much to learn about lightning physics, how fires are ignited, the conditions necessary for ignition, the smouldering process, and the conditions needed for smoke production.