When one geography grows most of the world's supply of a commodity, the state of that land — moisture, canopy, stress — moves the global price weeks before it prints. Cocoa is ~70% West Africa. Palm is ~85% Indonesia and Malaysia. Sugar and coffee cluster in a handful of origins. For these, a satellite that watches the producing regions is watching the supply side of the order book.
For wheat, corn and soybeans — grown across dozens of substitutable origins — the same satellite data is noise an over-parameterised model will happily overfit. Most vendors sell satellite data as a universal overlay. It is not, and we can prove it, because we tested ourselves against ourselves.
We ran an eight-year, walk-forward backtest — retrain every cycle on a rolling five-year window, every forecast stored against the eventually-realised outcome, zero look-ahead — across 25 liquid agricultural instruments and five horizons. Then we ran it again without the satellite arm, and compared head-to-head on identical cells.
| Configuration | 30d directional hit-rate | vs baseline |
|---|---|---|
| No satellite (baseline) | 50.5% | — |
| Satellite everywhere (blanket) | 49.6% | −0.9pp |
| Satellite compressed (PLS 256→16) | 48.8% | −1.7pp |
| Satellite gated by crop geography ★ | 50.8% | +0.2pp |
The portfolio headline understates the product. Inside the gate, the lifts are large and stable across every variant we ran: sugar +3.6pp, palm +2.7pp, cocoa +2.5pp at 30 days, with strategy Sharpe above baseline at three horizons. Outside the gate, forcing satellite on costs up to 4pp on wheat — which is why we don't do it. The gate was fixed from agronomy before scoring, not fitted after.
We let the archive answer the questions traders actually ask. Ninety days ahead of Brazil's 2021 black frost, the model was long arabica on 49 of 51 signals into a +21% realised move. Through the 2023-24 West African dry spell it was long cocoa on 76% of signals while NY cocoa tripled. Through Vietnam's 2024 drought, 80% directional on robusta. It missed the first rain-driven leg of the cocoa crisis, called soymeal but not beans through the Amazon drought, and scored 2/26 through Suez — a logistics event a crop model has no business claiming. Every one of these is recomputed live on the stress-test pages; nothing is curated away.
Two things changed. Satellite foundation models (Clay v1.5 today; Google AlphaEarth being integrated as a second arm) compress the entire scene over a producing region into embeddings a forecaster can use — beyond any hand-built index. And the softs complex entered a structurally tighter, climate-stressed decade: the same concentration of origin that makes these crops fragile makes them forecastable from orbit.
A model suite — four base learners and single-modality experts fused by a seasonally-gated mixture-of-experts — is implemented and queued for full walk-forward validation; results ship only after they beat the current production model out-of-sample. AlphaEarth 2017–2025 embeddings are extracted and will enter the next training cycle the same way. The discipline does not change: nothing reaches this terminal that the archive has not scored.
We are raising to scale the validated engine across the model suite, the second GeoFM arm, and live deployment with design partners — a hedge fund, a trading house, and a consumer-staples procurement desk. The demo behind this note is the live system. Click any number; it will defend itself.