Today, we’re expanding our model coverage of global food supply by announcing the launch of three new yield forecast models: one for Ukraine corn and a pair for Canada corn and Canada soy. These new models are now live in the Kernel product.
We are excited to announce our current end-of-season model forecasts.
Canada Soy: As of September 5th, our model forecasts a yield of 2.36 MT/ha.This is below 2017’s end of season yield number of 2.97 MT/ha and below our long-term trend-implied yield of 2.87 MT/ha.
Canada Corn: As of September 5th, our model forecasts a yield of 9.24 MT/ha.This is below 2017’s end of season yield number of 9.96 MT/ha and below our long-term trend-implied yield of 10.15 MT/ha.
Ukraine Corn: As of September 5th, our model forecasts a yield of 6.18 MT/ha.This is below 2017’s end of season yield number of 6.59 MT/ha and below our long-term trend-implied yield of 6.43 MT/ha.
OurKernel model are updated daily as they take more weather and satellite data into account.
How do we do it?
Our data science team trains machine-learning models on historical data to forecast what we expect the yield to be at end of year. We take into account a range of factors, including weather, field conditions, and satellite-derived crop health, to make our predictions as informed and accurate as possible.
How do we ensure accuracy?
We backtest our models against historical results to evaluate their accuracy. By testing our models against known results, we can see how well they would have predicted past outcomes. We continue to refine our models to reduce our margin of error through bias correction and using known techniques to avoid overfitting. We will only publish a model if backtesting suggests more accurate estimates than those of high-quality reference sources such as the Foreign Agricultural Service. (Learn more about backtesting from our previous blog post.)
At TellusLabs, we know that when it comes to actionable information, time is valuable. We provide daily data updates - a much higher frequency than monthly government reports. We use proprietary predictive models to process large amounts of raw weather and satellite data to make high-quality insights and predictions readily available for our customers.
Want live access to our newest models? Check out a free trial of our product, Kernel.