GeoAI · agricultural commodities

Satellites see supply before markets price it.

SpatioNiti reads crop state from orbit — satellite foundation-model embeddings over the regions that grow the world's softs — and turns it into directional return forecasts. Every number below comes from a walk-forward backtest with zero look-ahead, scored against what actually happened.

Book a call

spationiti.ai — live forecast terminal

cane-belt AOI · monitored daily

GeoFM embedding read here — a 256-dim crop-state vector that prices SB1 30–90 days out

🛰 MODIS true color · Center-South Brazil · satellite state in

SB1 raw sugar · price signal out →

The model reads crop state from orbit and turns it into a directional price path — this line is the real SB1 signal from our archive.

How it works

Observe

Satellites read crop state daily over every producing region we cover — moisture, canopy, stress.

Infer

GeoFM embeddings + weather feed walk-forward models, gated by crop geography. Zero look-ahead, ever.

Act

Directional 15–180 day signals for 25 instruments — each with its verified track record attached.

From raw pixels to price · the SpatioNiti pipeline

1 · RAW SIGNAL · MODIS DAILY

Côte d’Ivoire cocoa belt, exactly as the satellite sees it — every day.

→

e[37]=0.83dense canopy

e[112]=−0.41moisture stress

e[58]=0.12young replant

e[201]=0.67flowering flush

2 · GeoFM EMBEDDING · 256-DIM / FIELD

Our GeoFM reads each field into a 256-dim crop-state vector — the land, as numbers (illustrative channels shown).

→

3 · PRICE · EXPLAINED & FORECAST

Those vectors flagged the 2024 cocoa supercycle — and price the next 90 days, live.

181,777scored forecasts per run · ~8 years walk-forward

25live securities · futures & ETFs across 11 crops

5horizons · 15 / 30 / 60 / 90 / 180 days

liveforecasts as of · refreshed daily from market close + satellite

Live MODIS imagery · what our models watch, today

Center-South Brazil · sugar & coffee belts

Côte d’Ivoire & Ghana · ~70% of world cocoa

Sumatra & Kalimantan · palm estates

01 · The signal

A foundation model looks at farmland so you don't have to.

Hand-built indices like NDVI compress a satellite scene to one number. We keep the whole picture: our GeoFM embeddings — 256-dimensional monthly vectors from Sentinel-2 imagery — summarise everything the satellite sees over each crop's producing regions, on top of a physically-interpretable remote-sensing core. Region by region, weighted by production share, mapped to each tradeable security.

GeoFMour GeoFM · Sentinel-2

256-dim monthly scene embeddings over crop AOIs — the learned representation of "what the land looks like".

VegetationMODIS NDVI

Daily vegetation vigour, lags and rolling anomalies per producing region.

WaterCHIRPS rainfall · ERA5 soil moisture

Precipitation and root-zone moisture — the supply shocks that move softs.

HeatERA5-Land temperature · GDD

Growing-degree-days, heat stress and frost windows per phenology calendar.

02 · The discovery

Satellite data is a selective edge for commodities and commodity-linked securities.

We tested five configurations head-to-head on identical (security, horizon) cells — price + weather (rainfall, temperature, soil moisture, NDVI) is the baseline; the four arms add GeoFM satellite embeddings in different ways. Blanket use hurts. Compression doesn't save it. The only configuration that beats the baseline portfolio-wide is the one that applies satellite only where geography says it should work.

Price + weather baseline · rain, temp, soil, NDVI

50.5%reference

+ GeoFM satellite · blanket, all 256 dims

−0.9pp49.6%

+ GeoFM satellite · PLS-compressed 256→16

−1.7pp48.8%

+ GeoFM satellite · thin-ETF gate

−0.1pp50.4%

+ GeoFM satellite · concentrated-crop gate ★

+0.3pp50.8%

30-day directional hit-rate vs the price + weather baseline · n≈28 securities · identical cells across arms · walk-forward, zero look-ahead. Bars show the change against baseline. The concentrated-crop gate is the only arm that adds portfolio-wide — +0.3pp — and lifts strategy Sharpe at every horizon: 15d 0.265→0.287 (+8%), 30d 0.213→0.237 (+11%), 90d 0.504→0.512 (+2%). On the satellite-gated softs themselves the lift is +2.5 to +3.6pp.

03 · The gate

One rule, fixed before the test: follow the crop's geography.

When one region grows most of the world's supply, the satellite state of that region genuinely drives global price. When supply is spread across dozens of substitutable origins, 256 extra dimensions are noise to overfit. So satellite goes ON for concentrated-origin softs and OFF for distributed staples — decided a-priori from crop geography, not fit to results. The losers prove the rule as much as the winners.

Satellite ON · concentrated origins

The same three names lead every run

SB1 · Sugar #11 Brazil Center-South crush
+3.6pp
IJ1 · Palm oil ~85% Indonesia + Malaysia
+2.7pp
CC1 · Cocoa ~70% West Africa
+2.5pp
KC1 · Coffee arabica Brazil-dominated
+0.7pp

Satellite OFF · distributed origins

Where satellite hurts — so we don't use it

W 1 · Wheat CBOT dozens of substitutable origins
−4.1pp
MW1 · Wheat spring same story
−3.8pp
Corn · soy complex diluted signal, gated off
0.0pp
30d hit-rate deltas if satellite were forced on

04 · The money

What +2.5pp buys, in dollars.

A couple of points of directional accuracy sounds small. At institutional scale it is not — being right 2.5% more often, on moves that average ~6% a month, compounds across every position and every cycle. Assumptions stated in full — the value page computes these live.

≈ $3.6M / yrSystematic desk · $100M book, 10 × $10M positions, monthly cycle · 2 × 2.5pp × 6% avg |move| ≈ $30k per position-cycle

≈ $2.0M / yrAgro-procurement · 50kt cocoa per quarter (≈$400M/yr spend) · hedging right 2.5pp more often on ~10% quarterly swings

≈ $1.9M / yrPhysical trader · 2,000 lots raw sugar (≈$54M notional) · the same arithmetic at trade-desk scale

Indicative arithmetic — edge × notional × average move, before costs and sizing. Not a P&L promise: the gated softs carry +2.5 to +3.6pp measured lift, and the value page recomputes these numbers from today’s live forecasts.

05 · Seeing is the edge

The market sees a price. We see the physical geography through AI behind it.

Drag the divider. Left: sugar as every terminal shows it. Right: the same series with what our system was watching — the producing regions, the supply shock as it built, and the satellite drivers behind each call. Same instrument, same dates, real archive.

MODIS true color · Center-South Brazil · real

WITH SPATIONITI · drivers + regions + supply shocks

MARKET VIEW · price only

⇔

SB1 raw sugar, cumulative realised path from the archived forecast record · purple band = 2023 India monsoon deficit + export curbs · chips = the model's actual top drivers in that window.

06 · The product

Every forecast is archived, scored, and explained.

The demo is the real system — live exchange prices, live satellite features, and the full forecast archive. Nothing in it is illustrative.

Today grid · 25 × 5

Current directional forecasts for every security and horizon, colour-coded by what drives them — earth observation, 🛰 GeoFM embeddings, or both.

Any-day time travel

Pick any trading day since 2018. See exactly the forecast the model made with data available then — and what happened next.

Supply-shock replay

The 2024 cocoa crisis, Indonesia's palm export ban, Brazil's 2021 frost — see the model's calls through each event, hit by hit.

Driver attribution

Every cell exposes its top features — and how much of the signal came from the satellite foundation model vs physical indices.

Equity curves vs buy & hold

Non-compounded signal-following curves per (security, horizon), with supply shocks shaded on the timeline.

Producing-region maps

The exact AOI polygons the satellites watch, with phenology and climate-risk notes per region.

Stress-test it yourself.

Jump to the COVID crash. The 2024 cocoa supercycle. Any day of your choice from May 2014 through today. Every number you'll see is a real archived forecast scored against the realised return — the same engine now producing live forecasts from today's closing price.

Book a call →

07 · Book a demo

See your commodity through our satellites.

Leave your details and we will set up a 30-minute walkthrough.