Comparison with USDA

Pre-report trade estimates

Here are the trade estimates of the UDSA numbers for tomorrow. As you can see I'm way down at the edge of the corn yield estimates. This may be a year when I simply am much farther off from the USDA estimates. I'm right in the thick of it for soybeans, both yield and production. For cotton, I had some early sucess when I first built the model in predicting yields, but since this point, the model has been subject to error, much of it driven by the results in Texas.

However, at this point, the USDA still has historically made big adjustments between the October estimates and the final estimate (see the 2 posts before this one). The November estimates tend to be much closer to the final estimates.

Comparing USDA and Model Performance

In case you didn't look at the previous post, the dots are the mean and +/- 1 standard deviation for the USDA of the error of thier monthly estimate compared to the final. The lines are the same for my uncorrected model.

Here are some state by state comparisons of my error bounds each week relative to those of the USDA. As you can see I tend to be fairly competitive up until the October report. After the October report my error bounds don't decline, because I'm getting no new information. Now of course my errors are in-sample and the USDA errors are out of sample, but my error bounds are very similar to the USDA when looking at the October report. There is still room and precedent for the USDA to make pretty good changes to the yield estimate after October. For instance, the graph for Illinois corn shows a standard deviation of the change from October to Final of 5bushels. For my bounds I'm showing the uncorrected model. The corrected model simply takes my mean error each week and centers it back on zero, so the width doesn't change, I just recenter it. These graphs also show that the USDA has historically had a pattern to their estimates as they are not centered on zero.

Illinois Corn

Iowa Corn

Indiana Corn

Nebraska Corn

Missouri Corn

Minnesota Corn

Ohio Corn

Illinois Soybean

Iowa Soybean

Indiana Soybean

Nebraska Soybean

Missouri Soybean

Minnesota Soybean

Ohio Soybean

How does the USDA change its estimates through the year?

I've historically done a comparison piece where I compare my own errors in predicting final yields (in sample) to USDA's own errors (out of sample of course) to compare the relative accuracy and to see how errors decline throughout the growing season. I haven't done this for my own model yet (this is supposed to be a hobby) but I'll try to get some assistance in posting them but in the mean time, below you will find the USDA error distribution for corn and soybeans from 1986. The error I'm talking about here is each months guess relative to the final yield estimate. So for August, it is (August estimate-final). The blue dot is the average error while the red dot is plus/minus one standard deviation. From this you can really see that the movements from November are normally quite small, but there is still significant movement from October to the final estimate.

I'm farther away from the USDA estimates than 'normal' at this time of the year. It could be lots of things such as changes in technology, a oversized penalty for late planting based to much on history which no longer applies, some yet unknown error?

Not that I expect that I'll end up closer to USDA but I can hope. USDA could just as easily move further away from me on the upside for corn, or the downside for beans, but I'm given some hope by the recent corn and bean price movements that maybe on Friday I'll be a bit closer to USDA. I'll try to post those errors of my own work tomorrow and after the Friday report I'll give a state by state breakdown of how I compare. Email me if there is something else you want to see with respect to the errors. I've got some assistance in putting this all together so I can certainly promise somebody else's time to make it happen.

US Corn

US Soybean

Illinois Corn

Iowa Corn

Indiana Corn

Nebraska Corn

Minnesota Corn
Missouri Corn


Ohio Corn

Illinois SoybeanIowa Soybean

Indiana Soybean

Nebraska Soybean

Missouri Soybean

Minnesota Soybean

Ohio Soybean

USDA yield process

Presumably we should now get some harvested ear weights included in October estimates.
NASS objective yield process
So presumably the report on Friday should include some indication of test weight should it be an issue. Again there are readers out there much more knowledgeable on this issue than I am, you can either anonymously post comments or send me an email and I'll anonymously post any info you want to contribute if you indicate in the email.

After this report, the range of movement in the USDA estimate narrows significantly. I have calcualted the average error for the USDA in August through November by state and it is safe to say that after the October report, there is less chance of a significant movement. I'll try to post those historical error bands in the next day or two.

Week 27 results

Note the post after this one where I correct a question I had about how the 18 state totals are calculated. I know I've had that discussion in the past but I still made a incorrect statement about weighting. It is done on last years planted acres. Now my parameters are not fixed relative to each other that is to say Yield = a + b1(Poor) + b2(Fair) + b3(Good) + b4(Excellent) + b5(planting progress term) + b6(trend), etc. The parameters b1 through b4 are not restricted other than they must increase in value by some very nominal (essentially zero) amount. Please recall that I do this on a PLANTED AREA YIELD with a separate abandonment equation, not so important for corn and soybeans but more important for cotton. They are in bushels and the parameter values are basically the yield difference between the category in question and that in (Very Poor). So if b4=110, the difference in yield between an acre in Very Poor condition and that in Excellent is 110. But is is possible that b2 is only marginally greater than b3 or b4 only marginally greater than b3, etc. So not all condition movements have the same effect on yields and are not constrained across states, so when the 18 state total didn't move much last week, some states moved up while others moved down, this is enough for me to see a greater change that I might in an 18 state total model.

Week 27 spreadsheet (3mb)