Wildfires are a growing management concern in western US rangelands, where invasive annual grasses have altered fire regimes and contributed to an increased incidence of catastrophic large wildfires. Fire activity in arid, non-forested regions is thought to be largely controlled by interannual variation in fuel amount, which in turn is controlled by antecedent weather. Thus, long-range forecasting of fire activity in rangelands should be feasible given annual estimates of fuel quantity. Using a 32 yr time series of spatial data, we employ machine learning algorithms to predict the relative probability of large (>400 ha) wildfire in the Great Basin based on fine-scale annual and 16-day estimates of cover and production of vegetation functional groups, weather, and multitemporal scale drought indices. We evaluate the predictive utility of these models with a leave-one-year-out cross-validation, building spatial forecasts of fire probability for each year that we compare against actual maps of large wildfires. Herbaceous vegetation aboveground biomass production, bare ground cover, and long-term drought indices were the most important predictors of fire probability. Across 32 fire seasons, >80% of the area burned in large wildfires coincided with predicted fire probabilities ≥0.5. At the scale of the Great Basin, several metrics of fire season severity were moderately to strongly correlated with average fire probability, including total area burned in large wildfires, number of large wildfires, and average and maximum fire size. Our findings show that recent years of exceptional fire activity in the Great Basin were predictable based on antecedent weather and biomass of fine fuels, and reveal a significant increasing trend in fire probability over the last three decades driven by widespread changes in fine fuel characteristics.