This paper tests the out of sample predictive ability of reduced form models found in the literature on forecasting CO2 emissions. We show that for a newly available panel data set covering the fifty U.S. States and Washington D.C. during the years 1960 - 2000, the benchmark models found in the literature are outperformed by over 3000 of the considered models. A search over a large universe of models selects a ”best” performing model using the mean square forecast error from a simulated out of sample forecast experiment. Forecasts of aggregate CO2 emissions for the United States are provided.
