New residential construction is a primary user of wood products in the United States; therefore, wood products projections require understanding the determinants of housing starts. We model quarterly US total, single-family, and multifamily housing starts with several model specifications, using data from 1979 to 2008, and evaluate their fit out of sample, 2009-14. Goodness-of-fit statistics show that parsimonious models outperform general models in out-of-sample predictions. Monte Carlo simulations of total housing starts to 2070 project median starts ranging from 0.86 million/year at 0% real gross domestic product (GDP) growth to 1.91 million/year at 5% real growth, with 90% uncertainty bounds ranging from 0.52 to 2.13 million/year. Assuming that future GDP growth equals the average rate observed over 1990-2015, there is less than 9% probability that housing starts will exceed 2.0 million in any given year, 2016-35. Results show no evidence of structural change in the determinants of total or single-family housing starts coincident with the recession of 2007-09. Using these housing projections in a softwood lumber consumption model shows that GDP growth slower than 2% is consistent with stagnant or declining median softwood lumber consumption.