This week's results are just below in the following post, but let us look at a bit of historical performance in this post. I've pulled these results out of the models for the 2007 and 2008 crop year. The lines represent my model results while the red dots represent the USDA reports in the approximate week they were released. Again, I do not calibrate in any way to the USDA reports
and planted area (not harvested area) is taken as given. The models are re-estimated at the end of each year to make use of the
additional year of historical data. The first year where condition data in five
categories was available nation-wide was 1986 so each additional observation remains quite valuable. In 2007 you can see that the the crop size grew through September in the
USDA's estimates and the model tracked this but was at a lower level and kept growing through October, where conditions usually end as harvest is well underway. The November and January USDA numbers showed a decline. The model estimates were competitive through October, showing a good estimate. I've examined the model and will share some
additional look at model errors, but the conclusion is the model is very competitive against USDA estimates up to the October estimate, at which point no new information from the model is available but the USDA continues to collect information and the November USDA estimate usually shows only modest revisions in January. We can see that in 2008 the corrected model showed a good estimate of the final production estimate while the uncorrected model tracked USDA estimates quite well. I wouldn't draw a broad conclusion from this as I expect that the USDA estimates would be by
their nature a corrected model but perhaps somebody with more insight into
their procedure can comment. you can also see the nature of convergence of the two models where estimates in September, where the equations are estimated are identical.
Again the model predicts the final
production, does not seek to estimate the USDA estimate nor does the USDA estimate enter into the prediction of the final yield. That is to say, I don't take the USDA
estimate in August and try to use this information to predict final yields. I may put up some historical error analysis between the model and USDA performance if there is interest. Also if there are other graphs such as this that you would like to see, leave a comment and I'll see what I can do to put up the desired graphics.
and planted area (not harvested area) is taken as given. The models are re-estimated at the end of each year to make use of the additional year of historical data. The first year where condition data in five categories was available nation-wide was 1986 so each additional observation remains quite valuable. In 2007 you can see that the the crop size grew through September in the USDA's estimates and the model tracked this but was at a lower level and kept growing through October, where conditions usually end as harvest is well underway. The November and January USDA numbers showed a decline. The model estimates were competitive through October, showing a good estimate. I've examined the model and will share some additional look at model errors, but the conclusion is the model is very competitive against USDA estimates up to the October estimate, at which point no new information from the model is available but the USDA continues to collect information and the November USDA estimate usually shows only modest revisions in January. We can see that in 2008 the corrected model showed a good estimate of the final production estimate while the uncorrected model tracked USDA estimates quite well. I wouldn't draw a broad conclusion from this as I expect that the USDA estimates would be by their nature a corrected model but perhaps somebody with more insight into their procedure can comment. you can also see the nature of convergence of the two models where estimates in September, where the equations are estimated are identical.
No comments:
Post a Comment