← Back to Millet.coop

METHODOLOGY · VERSION 1

How this works

Millet.coop estimates 2026 yield for 829 mapped millet regions. It gives a weather-led estimate with an uncertainty label; it does not observe individual fields or calculate a complete market-price forecast.

What the estimate means

Model v1 combines growing-season weather with official national millet-yield history. It anchors each regional estimate to modeled 2020 regional yield and adjusts it using the national yield trend and weather response. The snapshot is dated September 30, 2026. Every result is a modeled estimate, not a measured harvest. “Moderate” and “wide” describe the model’s relative uncertainty; they are not guaranteed probability ranges.

Satellite vegetation is not an input to model v1. NOAA vegetation data are available for map context and were tested in backtesting, but were not retained in the locked model. Current crop health is not directly observed by this yield estimate.

Sources

What the backtest says

In a leave-one-year-out national-yield backtest, model v1’s production-weighted mean absolute error was 115.6 kg/ha, compared with 119.5 kg/ha for a simple trend. With each country-year weighted equally, model v1 was 311.0 kg/ha versus 307.1 kg/ha for trend. This is a national backtest; it does not prove that an individual regional estimate is accurate.

What it does not tell you

Yield alone does not establish available supply, carryover, exportable stocks, demand, or whether prices are more likely to rise or fall. Regional harvested area and current crop damage are not confirmed. Use the estimates as decision context, with the uncertainty and source dates in view.

Locked model artifact: assets/global-yield-estimates-2026-phase4.json · SHA-256 b28dc72afce61f96c64f579ea4b84e07e7a6fe818fa049c45951f9c406b2bcf0.