Author
Hubbs, Todd
We compare the accuracy of crop condition models from existing literature in forecasting U.S. corn and soybean yields. The data for the study consists of weekly state and national crop condition ratings from the USDA over 1986 through 2022. A battery of statistical tests is applied to perform out-of-sample forecasts over 2000 through 2022. While there are differences in the accuracy of the models, test results are uniform in suggesting that no model has statistically significant superior forecast accuracy. A key finding is that relatively simple models perform just as well as more complex and computationally demanding models.
The USDA produces four forecasts of net farm income for each year; these forecasts are closely monitored by decision makers across the agricultural sector. However, little is known about the performance of these forecasts. Traditional forecast evaluation tests suggest that between 1975 and 2016, the long-horizon forecasts systematically under-predicted realized values. In addition, the shorter-horizon forecast revisions overreact to new information. The findings suggest that forecast users should adjust their expectations and that the USDA may want to consider other forecast approaches to supplement current procedures.