The U.S. Department of Agriculture (USDA) publishes monthly Ending Stocks projections, providing an estimate of the end-of-marketing-year inventory of a particular commodity, which effectively summarizes supply and demand outlook. By comparing USDA's projections of balance sheet variables against their realized values from marketing years 1992/3 to 2019/20, we decompose ending stocks forecast errors into errors of the other supply and demand components. We apply a decision-tree-based ensemble Machine Learning (ML) algorithm, Extreme Gradient Boost Tree (EGBT), that uses a gradient boosting framework and is robust to multicollinearity. Our results indicate export and production misses to be the major contributors to ending stocks projection errors. Because foreign imports are likely tied to foreign production deficits, we likewise investigate how U.S. export errors are linked to USDA's foreign production and export forecast misses, country-by-country, and show that misses on production and export levels in China, Mexico, Brazil, and Europe cost USDA the most.