Using machine learning in trading strategies for agricultural commodities has gotten a lot of hype lately. While there are plenty of crop yield models, the market has lacked a systematic, reproducible, quantitative approach to modeling which fully specifies its forecasts and has transparent explanatory power.
We, at Gro Intelligence, set out to change that. Earlier this year, we decided to openly share the structure and output of our yield model which marries agronomic process expertise with machine learning. Gro’s website featured the weekly forecasts for all to see and comment on from May to November 2017. Subscribers had access to daily updates of the yield forecasts at the county level and access to all the signals in the model.
Our data product “Gro” encompasses global agricultural information on all crops. We initially made a US corn yield model as a proof-of-concept for a more general method that can help analysis of a broad array of agricultural commodities globally.
Despite our forecast model being designed strictly to estimate the critical real-world variable of corn yield, it can also add significant value for traders. Using our public forecasts in a simple trading strategy makes this utility very clear. The strategy made money in 71% of the 2017 trades, ending the forecast season in November up $0.1775/bushel (or +$887.50/contract.) In a backtest, it profited in 61% of 76 trades from May 2006 to November of 2016 and earned a cumulative $2.79/bushel (or +$13,950/contract.) In contrast, the well known trade estimate numbers, if traded the same way through 2016, lost a cumulative $0.72/bushel (or -$3,600/contract.)
Every month, the USDA reports its estimate of corn yield in the US in the World Agricultural Supply and Demand Estimate, or WASDE. Grain traders all over the world monitor the WASDE very closely during the US corn season, as the USDA has historically been a good forecaster of crop yield, which determines the price of corn. Prior to each report from August to November, financial news organizations survey traders in order to arrive at a “trade estimate” of what the WASDE number will be.
OUR TRADING METHOD
Our trading method follows the simplest set of rules we could come up with for a systematic trading model. We generate a “buy” or “sell” signal based on the level of our corn yield model relative to the trade estimate when it’s available, or last month’s USDA yield when the trade estimate is not compiled (May, June, and July.) If our model has a lower yield than the trade estimate or old USDA number, we bought corn because a lower crop yield should lead to higher prices. If our yield estimate was higher than the trade or USDA number, we sold it. All trades were executed in liquid nearby futures at the closing price the day before the WASDE report was released. Trades were reversed and closed out at the daily closing price after the report.
For example: on the afternoon of 11 October this year, our website displayed a corn yield estimate of 172.7 bushels/acre. The published trade estimate average stood at 170 bushels/acre, and the latest USDA yield as of 12 September was 169.9 bushels/acre. Since our yield was above the trade estimate, we sold corn for delivery in December at $3.46/bushel. Then on 12 October at noon NY time, the USDA came out with a yield of 171.8. Corn prices rose three cents/bushel, despite the higher yield. We then closed out the trade on the afternoon of the 12th for a loss of $0.03/bushel.
The live and backtest details below clearly show that our model, which was tracked and explained for free to anyone who visited our website, generated significant value for corn traders who traded using our yield forecasts. As a comparison, we have also added a table which shows what the “trade estimate” P/L would have been through 2017 had the same trading rules been applied - this would have led to significant losses of $0.9075/bushel (or -$4537.50/contract) vs a gain of $2.9675/bushel (or +$14,837.50/contract) using Gro’s model. Restricting the Gro system to the months with a trade estimate (Aug-Nov) leads to a similar result: a gain of $1.7625/bushel (or +$8,812.50/contract.) Of course, profits would increase if the yield estimates were used as inputs to more sophisticated trading rules.
Gro’s strategy is up in 61% of 76 trades since May 2006 and earned a cumulative $2.9675/bushel (or +$14,837.50/contract.) Note that the WASDE wasn’t issued in October of 2013 due to the Federal Government shutdown.
“Trade Estimate” P/L following a similar rule to the simple Gro system:
The trade estimate model is up in only 17 out of 47 trades, or 36% of the time, for a total loss over the entire period of $0.9075/bushel (or -$4,357.50/contract).