This article develops new model selection methods for forecasting panel data using a set of least squares (LS) vector autoregressions. Model selection is based on minimizing the estimated quadratic forecast risk among candidate models. We provide conditions under which the selection criterion is asymptotically efficient in the sense of Shibata (1980) as n (cross sections) and T (time series) approach infinity. Relative to extant selection criteria, this criterion places a heavier penalty on model dimensionality in order to account for the effects of parameterized forms of cross sectional heterogeneity (such as fixed effects) on forecast loss. We also extend the analysis to bias-corrected least squares, showing that significant reductions in forecast risk can be achieved.