The accuracy of any model is determined through out-of-sample forecasts over a period of time. In our case, we analyzed out-of-sample model performance over the last 17 years (primarily because USDA forecasts did not begin in May until 1993).
Forecast results shown below were generated using a two-step process. First, our models were estimated using up-to-date data (excluding data for the year being forecasted). Second, up-to-date observations of weather and crop development were used to make a forecast for the year being forecasted. This forecast simulation did not incorporate future weather forecasts and therefore should result in conservative estimates of YieldCast™ forecast accuracy. The YieldCast™ report that is released each Tuesday will incorporate future weather forecasts from T-storm Weather to help us reach our final yield forecast more quickly.
As shown in the top chart, over the last 17 years, U.S. corn yield forecasts from YieldCast™ outperformed USDA and trend forecasts from June through September — especially so in August and September*. Although USDA forecasts outperformed YieldCast™ in some years, corn forecasts from YieldCast™ have been more accurate on average.
Forecasts for each year over 1993 through 2009 are also provided below (beginning in 2009 and ending in 1993).
† Average forecast error is calculated as if all U.S. corn yields from 1993-2009 were the same as in 2009 (164.9 bushels per acre)
* Based on the Mean Absolute Percentage Error forecast accuracy statistic for U.S. corn and soybean yields over 1993-2009.