This study analyzes forecasts of U.S. ending stocks for corn, soybeans, and wheat issued by the USDA. The proposed efficiency tests focus on forecast revisions. Forecast errors are decomposed into monthly unforecastable shocks and idiosyncratic residuals. The error covariance matrix allows for heteroscedasticity and auto-correlations. Results suggest that the USDA forecasts are inefficient, providing strong evidence that the USDA is conservative in forecasting the ending stocks. Unforecastable shocks are heteroscedastic, and idiosyncratic residuals are small. Results are consistent across the three decades analyzed, but soybean forecasts are found to be considerably worse from 2005 to 2015.