Descartes Labs Can Make Corn Yield Predictions Better Than USDA! How Do They Do It?

Descartes Labs released its 2016 U.S. Corn Production Forecast in August not only to the US government but to the public as well.

However, in a feat that will marvel even forecasting experts, the company seems to have consistently updated its corn forecast every Tuesday. 

Interested readers can find the numbers on their website, mobile app, and newsletter. But the need to forecast corn numbers nationwide appears to have a more practical and pressing meaning at hand.

According to Medium, this is best explained by elaborating that corn is the largest and most important crop in the United States. Majority of it, however, does not end up being eaten.

The U.S. is the number one producer of corn in the world, with harvests reaching 13.6 billion bushels last year in 80.7 million acres of land. It is used for animal feed, ethanol, and corn syrup, among other things. This means corn is integral in a number of industries as well. 

The U.S. Department of Agriculture already does an incredible job of forecasting the crop, even better than any other country in the world. It releases monthly forecasts for corn production at the state and country level beginning August, and the final report comes out in January. This is the first time country-level yields were reported.

The USDA does these reports by surveying farmers and touring farms to sample data and uses all these information to make projections.

Descartes Labs does a similar forecast, but in a much larger and more efficient scale. Its forecasts are generated weekly and even include in-season county-level yields. This gives better local visibility to farmers and grain buyers.

For instance, it is important to know that grain elevators price corn locally based on the local supply of corn from farmers. Understanding this supply can help elevator managers price corn for better market efficiency. The company's reliance on satellites and stations can survey every field in every farm every day as well. 

The company combines initial data with detailed weather data and how weather variables deferentially affect corn to build its models. Its system can analyze a petabyte of data in just 16 hours with the help of 30,000 computer cores.

This means its model is more frequent as it can produce forecasts every two days, and it has made vast improvements in its accuracy. Descartes Lab has thrice the data that its needs, and its report has 2.37 percent statistical error versus the August 4 USDA report. 

Not only does Descartes Labs and its usage of huge datasets and machine learning help forecast the future on a local and global level, but it appears they also have plans to expand to soy and other major commodity crops.