Yield Forecasting

Blog/

How regressing an index curve against historical yield actually works

Close-up of a dense, lush green wheat canopy in a field.

If you've ever heard someone say they're "regressing the index curve against yield" and nodded along without quite knowing what that means, here's the mechanics behind it. No black box, just a statistical relationship between two things that move together season after season.

What the index curve is

A vegetation index, usually something like NDVI or a close cousin, is a number derived from how a crop canopy reflects light in different bands. Healthy, dense green tissue reflects near-infrared strongly and absorbs red light. Stressed or sparse canopy does the opposite. Track that number for a field or a region across a growing season and you get a curve: it climbs through emergence and vegetative growth, peaks somewhere around flowering or grain fill depending on the crop, then tapers as the canopy senesces.

That curve by itself tells you something about crop condition relative to prior weeks. It doesn't tell you bushels. To get from curve to production number, you need history.

Building the regression

Take a region with ten, fifteen, twenty years of index curves, each one paired with the harvested yield for that crop year. Now you've got a dataset: curve shape and amplitude on one side, realized yield on the other. A regression model fits a relationship between features of the curve (peak value, cumulative index over the season, rate of green-up, timing of peak) and the yield that region produced.

Once that relationship is fit, you can run this year's partial curve, the one that's still accumulating week by week, through the same model and get a projected yield. Early in the season the projection carries wide uncertainty because you're extrapolating from a short piece of curve. By the time you're weeks out from harvest, the curve has filled in enough that the estimate tightens up considerably, because the model has seen most of what it needs to see.

This is the same logic a basis trader applies to a thin market: more data points narrow the range, and you weight recent prints more heavily than the open. The index curve regression does that mechanically, season-to-date data pulled against a weekly print, region by region.

Why region matters more than field

A single field's index curve is noisy. Planting date variance, a hail event, a drainage issue in one corner of the farm, any of these will throw a field-level reading around in ways that have nothing to do with regional production. Run the regression at the regional level instead, averaged across enough acreage to smooth out the one-off events, and the curve tracks what a production call needs: how the crop in that region is trending relative to the years it's been trained on.

That's also why this approach handles planted area and harvested area separately from the yield-per-acre question. The regression gives you a yield trend. You still need a planted-area number from elsewhere to turn that into total production, and harvested area if abandonment is a factor in a given crop year.

Where this beats waiting on a harvest report

A combine doesn't start running until the crop is ready, which means a harvest report by definition arrives after the window where a trading desk needs a number. A single end-of-season satellite look is better than nothing but gives you one data point instead of a trend. Regressing the index curve weekly means the estimate updates as the season develops, and you can watch the trend line move instead of waiting for one static read right before combines roll.

That's the version of this we run as a weekly regional production call, refreshed region by region as the season's index curve fills in against the historical yield relationship for that area.

If you're pricing carryout off a harvest report that's always two weeks behind the crop, ask about getting on the weekly print instead.

Start a project