Yield series are an input, not an answer

The division holds yield series for eleven crops across six countries, some running back twenty-two years, assembled in the course of our own agri-science origination. They are the most requested inputs the research desk produces and the ones we attach the longest notes to. A yield series is a measurement of an agronomic outcome, and the thing being measured changes underneath the numbers. Three changes matter enough to break a model, and all three are documented rather than inferred.

The first is measurement method. A national statistics service that moves from farmer surveys to remote sensing does not create a discontinuity in the crop. It creates one in the record. In four of our eleven series the level shifts by between 3 and 7 per cent at the year of transition, in a direction that depends on which method over-reported. A regression run across that join attributes the shift to whatever regressor happens to move in the same year, and the coefficient it produces is a measurement artefact wearing the clothes of an agronomic finding.

The second is variety turnover. A yield series for a crop is really a series for a rolling population of varieties, and the population turns over every five to eight years. Gains that look like a response to rainfall or to input prices are frequently the arrival of a variety with different drought tolerance. The desk keeps a separate variety-share series for each crop and requires any yield model to include it, which typically halves the coefficient on the weather terms and removes the apparent trend that clients most want to extrapolate.

The third is the land itself. When prices are strong, marginal land enters production and average yield falls even though every field has improved. When prices fall, marginal land exits and average yield rises. The correlation between area and yield in our series is negative in eight of the eleven crops, and the effect is large enough to reverse the sign of a naive price model. Yield per hectare and total production answer different questions, and a model that uses the first to forecast the second without an area term is answering neither.

Trial data carries a further problem. Variety trials are run by parties with an interest in the result, and the plots that fail are less likely to be reported than the plots that succeed. We treat trial yields as an upper bound and never as a central estimate, and we down-weight any source that cannot supply the count of trials started as well as the count reported. Two commercial datasets offered to the division this year could not supply that count. Neither is in use.

The discipline this imposes is a rule rather than a preference. No series enters a model in this division without a written note stating where it came from, what it measures, where the joins are and what may lawfully be done with it. The rule costs the desk time and has removed several attractive datasets from consideration. It also means that when a position is sized on an agri-science forecast, the firm can say what the forecast rests on. The firm closes 2022 with US$312 million of client portfolio under stewardship and 142 people, and the research standard that supports those numbers is the one written down here.

Published 2022-12-13 by the Quantitative Strategies division. Research is prepared for eligible counterparties and does not constitute advice.