TellusLabs Trend Estimates
TellusLabs is pleased to share our long-term trend yield estimates for corn, soy and wheat crops in the northern Hemisphere.
In a matter of weeks, we will be launching satellite and weather-based models for wheat, which will be soon followed by corn and soy models (in June). A crucial initial stage in that modelling process is determining the long-term trend for yield for each country-crop pair.
Around the world, evidence points to a secular improvement in yields, driven in large part by innovations in seed technology and farm management practices. Incorporating this structural improvement into our models by analysing recent historical yields is the starting point for our modeling process.
Today, we are sharing these “starting point” numbers.
New models vs. 2017
In Summer of 2017, we produced live yield models for US corn and soy. Over the past several months, we’ve deployed live (satellite and weather-informed) models for Argentine and Brazilian corn and soy. For this Northern Hemisphere growing season, our models will cover corn, soyand wheat crops in 8 countries (in this case, we are calling the EU a single country - the Kernel product shares forecasts by individual nation as well).
Our trend results
*US Wheat estimate is average of Spring and Winter Wheat weighted based on 2017 production figures of 27.1% of spring wheat, and 72.9% of winter wheat and forecasts of 48.9 bu/ac (Winter) and 47.1 bu/ac (Spring).
**Predominantly Spring Wheat
It is important to bear in mind that these trend estimates have no information from the current season. They are a reflection of long-term trends only (see a brief discussion of methodology below). While it is convenient to compare these trend estimates to current government forecasts, they are different in kind.
To model the national end of season yields across the combinations of countries and crops reported here we employed unique linear fixed effects model for each country & crop combination. The models are built from county-level data sourced from various public data sources. Linear effects were calculated by state and then enforced for individual counties. The intercept for each county was allowed to vary in the models. We conducted a grid-search to determine the optimal number of years to employ in training and the threshold for outlier exclusion. Our objective in that search was to minimize root mean squared error versus historical end of season yields for the last ten years at the national scale. Further methodology details available on request; we welcome feedback and ideas on this important first step in the modelling process.