Accurate prediction of forest fire spread is a critical management and scientific challenge as the world adapts to rapidly changing fire regimes. We reconstructed 5,400 daily burned area progression maps for 196 U.S. Northern Rocky Mountain wildfires (2012-2021) and used machine learning to estimate daily fire growth given local weather, hydroclimate, fuels and topography. Optimized models explained 36% of the variation in daily fire growth, increasing to 56% when an index of fire activity the previous day was included. Soil moisture and plant hydraulic stress were the dominant predictors of fire spread, increasing accuracy by 8%-9% over models with only fuel and weather. Wildfire danger forecasts and fire spread models in the U.S. use short‐term weather indices and don't consider longer‐term drought. Our findings suggest that soil moisture and vegetation stress are critical indicators of fire spread potential in this region, with implications for fire modeling and prescribed burn planning.