When is an error not a mistake?
If we’re talking about forecasting, then “error” is just the term for how much difference is ultimately noted between the formerly predicted value and the actual value.
In 2018, U.S. producers ultimately planted 89.129 million acres (ma) of corn compared to a March Prospective Plantings prediction of 88.026 ma. That’s an “error” of 1.103 ma or 1.25%.
But that’s not to say there was anything “wrong” about that original prediction. “Error” is merely the mathematical term to describe the quantity. It was the best information available at the time (surveyed responses of real intentions).
Things just change; intentions change from one month to the next.
This year, the March Prospective Plantings survey suggested U.S. farmers intend to plant 92.792 ma of corn, 84.617 ma of soybeans and 45.754 ma of all varieties of wheat. It’s safe to assume the final planted acreage of each of these crops won’t come out to exactly those numbers. They’re likely to be off by a few hundred thousand acres or so. That doesn’t necessarily make them bad forecasts. They may be an exact correct measurement of what would have been planted by U.S. farmers on March 1 if they’d been physically capable of planting all their spring crops on that date. All the “error” is showing is that there is some unavoidable difference between expectation and reality.
Errors in the Forecast
Sometimes, however, we can be judgmental about forecasts with a lot of error — if those forecasts are biased in one direction or another.
If, every year, U.S. farmers consistently indicated in the March Prospective Plantings survey that they intended to plant 5% more corn acres than they actually ended up planting, well, then we could reliably adjust our ultimate acreage expectations down exactly 5% from the surveyed number.
On the other hand, if the number is 2% high one year, 2% low the next year, 1.6% high the following year and 1.6% low the year after that, then the “errors” would all average out. The expected value of the forecast error would be zero and we would believe that the forecast isn’t biased. It might be wrong one way or another, but we can’t predict from history which way it will be.
WADSE Reliability Tables
The last few pages of any World Agricultural Supply and Demand Estimates (the monthly WASDE report released by USDA economists) always contain some reliability tables that start with the term “root mean square error,” which causes everyone’s eyes immediately gloss over and ignore them. However, what those tables are trying to show is whether or not the mathematical models that USDA are using to calculate its production, export, usage and ending stocks estimates are unbiased. Some of the results are rather interesting.
For instance, over the past 36 years, USDA’s April estimate for U.S. corn ending stocks overshot the ultimate September 1 measured number 18 of those years, and also undershot the ultimate number 18 of those years; that’s pretty balanced.
AgFax Weed Solutions
The average error of their April forecast is 151 million bushels (mb) in one direction or another and the USDA’s economists can say with 90% confidence that their April estimate of U.S. corn ending stocks won’t be off by more than 22.4%. So, having estimated on April 9 that 2018-19 corn ending stocks will ultimately be 2.035 billion bushels (bb), they are 90% confident that on September 1, the actual number won’t be less than 1.579 bb or more than 2.491 bb (seems reasonable).
Working from 36 data points (the past 36 April rounds of projections compared to final marketing year numbers), not every crop gets such confidence from USDA’s economists, however.
There’s an expected 33.3% incidence that their U.S. cotton ending stocks projection could be above or below the final estimate by more than 13.7% and that their U.S. soybean ending stocks projection could be above or below the final estimate by more than 35.7%. Still, that’s in either direction — i.e. unbiased. Their April estimates of world wheat and rice production and exports have been too low 29 of the past 36 years and too high in only seven of their past 36 attempts at forecasting. Therefore, these reliability tables do give us some clues about which way we should be skeptical of the government’s models, if we’re inclined to be skeptical of such things.
Farmer Acreage Projections
Meanwhile, the planted acreage projections released at the end of March weren’t the product of any mathematical model built by the government’s economists — they were direct responses from farmers themselves answering a survey about what they intend to plant. There would certainly seem to be some potential for bias in those answers.
However, when I looked at past data (and admittedly, I didn’t look at 36 years’ worth of data), the Prospective Plantings numbers didn’t seem to lean in any significantly biased direction or another. Actual planted corn acreage over the past five years has turned out to be 98.8 percent of intentions, then 98.7%, then 100.4%, then 100.2%, then 101.3% — an expected value of 99.9%, or virtually right on the nose.
Practically, we might expect the corn crop, which is relied on for feed in livestock producers’ regular crop rotation, to be more stable from one year to the next than other crops such as soybeans or wheat — and that’s exactly what we see.
Soybeans’ actual acres compared to March intentions have ranged from 97.7% to 102.2% (with an expected value of 100.5%) and wheat’s actual acres compared to March intentions have ranged from 99.3% to 101.8% (expected value of 100.7%).
Basically, what farmers say they’re going to plant in March is indeed what ultimately gets planted on a nationwide aggregate basis.
All of that assumes that spring weather allows intentions to become reality. Nevertheless, to calculate confidence intervals for this spring’s weather, we’d have to start talking about the mathematical models behind the current weather forecast … and that’s a whole other level of uncertainty.
Elaine Kub is the author of “Mastering the Grain Markets: How Profits Are Really Made” and can be reached at firstname.lastname@example.org or on Twitter @elainekub.