MIT Tech Review: Crystal Ball for Corn Crop Yields Will Revolutionize Commodity Trading.

May 08, 2017

When talking about TellusLabs, it’s easy (and fun!) to spend a lot of time discussing satellites and machine learning - those topics certainly garner a lot of attention! That said, we think it’s helpful to understand what we’re doing in the context of the problem we’re solving in the agricultural markets.

Agricultural marketplaces date back to Sumer as early as 4,500 BCE, so you could probably call ag commodities TheOriginal Asset Class. The market for ags was a big deal then and it’s a big deal now: Food and agribusiness together are a $5 trillion global industry. Here in the United States, last year’s corn and soy harvest alone was worth nearly $100bn in direct farm cash value, and as much as 15x that was traded in the futures markets. Although the supply side for ags plays out in plain view on farms around the world, market intelligence is still driven largely by gut instinct, phone calls with farmers, and weather-only forecasts. The result is one of the more volatile asset classes in the global economy.

We can do better. Last Spring, two remote sensing scientists, David and Mark, decided to prove it by applying decades of experience working with NASA and the Fortune 500 to agricultural intelligence.

Commodities traders and researchers take a lot into account when they make investment decisions around grains, but by far the most important source of variation is often crop yield: How many bushels is a single acre of farmland going to produce? What matters in this game isWhen do you know?andHow sure are you?In January 2017, the USDA announced that on average each of the 82.7 million acres of US farmland devoted to soy yielded 52.1 bushels of beans. The market spent the year trying to predict this “right answer” (which was also a historical figure - a record harvest for Soy in the US).

Throughout the season, the USDA issues monthly reports, with forecasts for this end-of-season number. These monthly production reports get noticed; the futures market moves within minutes of each announcement, driving the most volatile trading days of the entire year. Depending on the crop and timing, a reset in expectations of just 1 bushel per acre can translate into dramatic fluctuation.

Every day from mid-August to mid-October of last year, TellusLabs refreshed our own forecast of the end of season number and sent it in pdf form to a community of beta participants that grew from a couple dozen friends and family to a community of over 600 growers, traders and researchers from around the world. These beta participants were witnesses to a pretty astounding result. Not only did our final October forecast of 52.1 bushels per acre prove to be the exact end of season outcome, but we closed within 1% of that final number in September, two full months before the USDA.

If you were part of that beta community receiving a daily forecast from TellusLabs, you would’ve had a 2 month head start on the market. (Also: We also closed within 1% of the USDA’s January estimate for Corn.)

How do we do it?

We start with with data from satellites. Spaceborne intelligence is not new to the marketplace -- in recent years companies have used high-res imagery to count cars in parking lots, or monitor oil reserves.

What we do is different.  

We are not investigating individual sites or counting things. We are building planetary-scale, daily-cadence models of (in this case) all 170 million acres of US corn and soy.  We work with Earth Observation data coming from satellites that’ve been orbiting the planet for over 15 years.

We tame this massive collection of satellite, weather and ground data with our Earth database. We then get down to work building features: practical, science-grade, physically meaningful “vital signs” for the Earth. Data scientists appreciate the importance of feature engineering -- strong features make for strong models. We then apply machine learning to build the sort of predictive models that deliver the results we saw in 2016.

That’s what going on “under the hood”.... but our customers care about where the rubber meets the road. In 2017, that’s the Kernel product.

Last year, 600 growers and traders signed up for our Kernel beta and received a daily pdf (created with care every morning by David himself).

The fantastic team we’ve assembled over the past year allow us to take a more ambitious approach for 2017.

For this year’s corn and soy season, we’re providing an immersive web app alongside an intuitive API. We know that just giving an answer (even the right answer) isn’t enough. To help our users make better decisions, we need to arm them with context: the “why” matters as much as the “what” - market participants won’t be satisfied with another black box. InKernel, users can see confidence intervals and historical comparisons for our forecasts, they can drill down to individual states and even counties, and they can see how much each forecast is influenced by satellite data, weather data or ground observations. Via our new API, customers can access all of the same insights on their own terms -- bringingKernel into the heart of their own models and systems.

Getting the answer right for ags means takingKernel far afield of Iowa and Indiana.  This very moment, half of the global harvest for soybeans is in the ground in Argentina and Brazil. This region matters for ags, but it’s a tough place for the markets. Opaque reporting, extreme weather (like the flooding event we tracked in La Pampas), and more complex cropping cycles fuel an order of magnitude more uncertainty than we have here in the US. This April, we launched a first of its kind, daily cadence satellite and weather-based forecast for Argentina’s soy crop (which we’ve finalized in the months since).

We’re pushing to make this product as intuitive and insightful as the technology behind it because we think the market has depended on haphazard methods for too long. Kernel aims to empower ag market participants with real insights to allow them to make better decisions.

We enjoy letting our imaginations run wild about the other applications for satellite and predictive modelling. We’ve already started R&D work on applications in energy, forestry and water resources - and over time our company will have things to say on each of these themes. We’re looking for problems that are large in scale, feature fast-changing data sets, and are meaningful both to markets and society.

Today, we are laser-focussed on agriculture. We’re delivering a solution that brings objectivity and insight to a market that has gone without it for too long; and we’re excited to find partners for the journey.

The above is taken from a presentation we gave at FinTech Sandbox’ Demo Day 4.0 on April 20, 2017, at Fidelity’s offices in Boston. It was fantastic event and we are very appreciative to have been included!

Also in News

TellusLabs' Summer Adventures
TellusLabs' Summer Adventures

September 22, 2017

Recap of TellusLabs' summer adventures, from San Francisco to Nova Scotia to Grand Island, Nebraska.
USDA August crop production report commentary
USDA August crop production report commentary

August 10, 2017

TellusLabs new Kernel models show a 2017 US growing season back on track
TellusLabs new Kernel models show a 2017 US growing season back on track

June 28, 2017