How to Make Money With Gro Intelligence Yield Models

A strategy of trading futures contracts based on signals from Gro Intelligence’s US Corn Yield Model is off to a profitable start this crop season. The trading strategy registered profits in live paper trading in each of 2017 and 2018. For the current year, paper trades executed in May and June also have shown a net gain. The favorable results affirm the viability of using machine-learning-based yield models for trading.

Gro’s yield model covering the US corn market was the first such model we deployed that uses machine learning. It launched at the beginning of the 2017 Northern Hemisphere crop season. The model forecasts final yield using satellite- and ground-based variables at the county level, aggregated upwards to arrive at a national yield number. (Click here for a complete description of how the model works.) It aims to estimate yield as early and accurately as possible regardless of price movements. But we wondered if it could help futures traders, too.

Since we wanted to quantify the trading value of the model in simple terms, we measured its success in forecasting corn futures price movements on monthly USDA report days. To head off any concern about how we constructed the track record, we kept everything as plain as possible. If our model’s yield estimate was below either the USDA’s previous monthly forecast, or the trade estimate (when available), that signaled a buy. If our estimate was higher, it signaled a sale. If our model said “buy” we bought at the close on the day before the report, and sold at the close after the report was released. And vice versa if the model said “sell.”

In backtesting, for the decade ended 2016, that simple trading strategy made money consistently. That gave us some reassurance, but didn’t really impress anyone. After all, it makes sense that foreknowledge of corn yields would translate to trading profits retrospectively. But the fact that it has continued to make money since our model went live has attracted significant favorable attention. Now in its third year of live paper-trading performance, the strategy has so far profited by 11 cents a bushel, or $550 per contract, in just two 2019 report days of trading in May and June.

The table below details profits and losses from use of the trading strategy for each month of live paper trading (figures on red background) and backtesting (on white background), along with annual totals.

The past three years of paper trading profits have big implications for traditional discretionary traders and commodity trading advisors alike. If you’re trading agricultural futures without looking at a modern machine-learning-based model, why? We can help you. The USDA issues one yield estimate per month, exactly the same snail’s pace of 40 years ago, before the internet and the birthdays of many of the agency’s employees. Daily satellites pass over the corn crop 30 times between each of the USDA’s yield estimates. The agency runs a corn yield model similar to ours, but keeps its results a secret. Is it really that surprising that modern techniques can anticipate what the USDA will say, 30 daily model runs after their last statement?

The chart below highlights the benefit in cents per bushel earned of following a trading strategy for futures contracts based on Gro’s US Corn Yield Model (red line) versus a strategy based on trade estimates (blue line), which are usually published in financial media.

If you would like to learn more about Gro’s US Corn Yield Model or any of our other sophisticated machine-learning-based models of agricultural variables around the world, please contact us.