Yield Prediction Models
The corn yield prediction is based on Jiang et al. (2018), in which authors developed a Long Short-Term Memory (LSTM) model to
redict corn yields in ten corn belt states—Illinois, Indiana, Iowa, Kansas, Michigan, Missouri, Minnesota, Nebraska, Ohio,
and South Dakota—that achieved promising results with the sample data. Overall, the model prediction is only 0.83 bushel per acre (bpa)
lower than actual corn yields, a smaller difference than the corresponding prediction from USDA.
About 80% of the LSTM county-level corn yield predictions fall within +/-20 bpa of actual yields.
The model uses ten variables that affect corn yields, which were constructed from data sets provided by the USDA, USDA-NRCS,
USDA-NASS, NOAA, and IBM weather underground.
The original model was re-calibrated in 2021 with the two changes. First, we replaced IBM weather undergroud data set with public assessible
GRIDMET data set by the Climatology Lab. Second, Wisconsin corn yield model were added. Thus,
the new US new yield model includes 11 midwest states which accounted for about 85% US corn production in 2020.
Current Year Yield Prediction
Current yield prediction relies on three sets of information. First, our yield model will predict county corn yields by combining the realized
weather conditions (from GRIDMET along with historical weather conditions between April 1st - October 31st
(The growing season chosed in our yield prediction work). Second, the county harvest infromation, such as total harvested grain acreage and county share over total state harvested
grain acreage, queried from USDA quickstats web portal. Third, forcasted state and national corn production information in
USDA crop production report .
For states in our model, we first calculate the forcasted county harvested acreage by multipling forcasted acreage in crop production reports with previous year's county harvest share. Then, the forcasted
state yield is calculated based on forcasted county yields with these acreages as the weights. In the calculation of forcasted national yield, we treat state yield in non-modeled states as in the crop
production reports and calculate national yield by weighting our state yield forcasts for 11 midwest states and state yields for other states according to forcasted state harveste acreage in the reports.
Paper Link:
Jiang et al. (2018)