Empirical confidence intervals for USDA commodity price forecasts

Abstract

Conventional procedures for calculating confidence limits of forecasts generated by statistical models provide little guidance for forecasts based on a combination or a consensus process rather than formal models, as is the case with US Department of Agriculture (USDA) forecasts. This study applied and compared several procedures for calculating empirical confidence intervals for USDA forecasts of corn, soybean and wheat prices over the 1980/81 through 2006/07 marketing years. Alternative procedures were compared based on out-of-sample performance over 1995/96 through 2006/07. The results of this study demonstrate that kernel density, quantile distribution and best fitting parametric distribution (logistic) methods provided confidence intervals calibrated at the 80% level prior to harvest and 90% level after harvest. The kernel density-based method appears most accurate both before and after harvest with the final value falling inside the forecast interval 77% of the time before harvest and 92% after harvest, followed by quantile regression (73% and 91% before and after harvest, respectively) logistic distribution (73% and 90% before and after harvest, respectively) and histogram (66% and 84% before and after harvest, respectively). Overall, this study demonstrates that empirical approaches may be used to construct more accurate confidence intervals for USDA corn, soybean and wheat price forecasts.