Economic forecasting is both an essential exercise and a way to discipline the construction of models. We show that pooling data from many countries in a cross-sectional dimension to macroeconomic data can statistically significantly improve the generalization ability of structural, reduced-from, and machine learning models in forecasting GDP. This procedure reduces root mean squared error (RMSE) by 12 percent across horizons for certain reduced form models and by 24 percent across horizons for structural DSGE models. A central result in this paper is that, in contrast with theory, DSGE or reduced from models forecast better, or at least as well, when estimated on more recent time series across a cross-section of countries rather than data that extends further back in time but only from the country of interest. Finally, given the comparative advantage of “nonparametric” Machine Learning forecasting models in a data-rich regime, we demonstrate that our Recurrent Neural Network (RNN) model and Automated Machine Learning (AutoML) approach outperform all baseline Economic models and even the Survey of Professional Forecasters at long term horizons when using pooled macroeconomic data. Robustness checks indicate that machine learning techniques are reproducible, numerically stable, and generalize across models.