New England forests provide numerous benefits to the region’s residents, but are undergoing rapid development. We used boosted regression tree analysis (BRT) to assess geographic predictors of forest loss to development between 2001 and 2011. BRT combines classification and regression trees with machine learning to generate non-parametric statistical models that can capture non-linear relationships. Based on National Land Cover Database (NLCD) maps of land cover change, we assessed the importance of the biophysical and social variables selected for full region coverage and minimal collinearity in predicting forest loss to development, specifically: elevation, slope, distance to roads, density of highways, distance to built land, distance to cities, population density, change in population density, relative change in population density, population per housing unit, median income, state, land ownership categories and county classification as recreation or retirement counties. The resulting models explained 6.9% of the variation for 2001–2011, 4.5% for 2001–2006 and 1.8% for 2006–2011, fairly high values given the complexity of factors predicting land development and the high resolution of the spatial datasets (30-m pixels). The two most important variables in the BRT were “population density” and “distance to road”, which together made up 55.5% of the variation for 2001–2011, 49.4% for 2001–2006 and 42.9% for 2006–2011. The lower predictive power for 2006–2011 may reflect reduced development due to the “Great Recession”. From our models, we generated high-resolution probability surfaces, which can provide a key input for simulation models of forest and land cover change.