As coastal populations increase every year, greater numbers of people and buildings to support them are left vulnerable to severe hazards associated with hurricanes, which have shown signs of increasing strength and frequency related to climate change. Community-level decision making is essential to adequately prepare populations for the risks associated with imminent hurricanes and to adapt buildings to be more resilient. This creates a need for state-of-the-art methods such as data-driven machine learning to predict the damage that buildings will experience during hurricanes and support decisions for community stakeholders. Previous research has attempted to proactively forecast hurricane damage using numerical frameworks for individual building archetypes or by incorporating a narrow spectrum of input features. The focus of this study is a novel machine learning framework trained on building, hazard, and geospatial data to hindcast damage from Hurricanes Harvey, Irma, Michael, and Laura, with the objective of forecasting expected damage from future hurricanes. Performance of different algorithms were investigated including k-nearest neighbors, decision tree, random forest, and gradient boosting trees algorithms. In predicting qualitative damage states, random forest outperforms other algorithms with 76% accuracy in the hindcast. Parametric studies identify which features contribute the most to accurate predictions and demonstrate that prediction accuracy increases linearly for this case study with additional reconnaissance data to train the model. Finally, a comparison is drawn between this model and the ability of Federal Emergency Management Agency’s Hazus Multi-Hazard Hurricane Model to estimate building-specific damage on the same hindcast set of buildings.