SpatioNiti · GeoAI commodity forecasting

satellite-driven commodity forecasts
as of about
explore: 📊 The 10 stress-tests 🌍 Globe view 📄 Investment note 💼 What today is worth 🤖 Ask Niti
cell color = what drives the forecast: earth observation (NDVI / rain / temp / soil) 🛰 satellite embeddings (Clay) both price/other only hover = details · click = drill: equity curve, drivers, events, map
replay a moment: — or pick any date above; you see exactly what the model knew then
Method. Walk-forward XGBoost per (security, horizon) — retrained on a rolling 5-year window, every prediction scored against the realised outcome, zero look-ahead · 181,777 backtest forecasts across ~8 years · Remote-sensing core (MODIS NDVI, CHIRPS rainfall, ERA5-Land temp/GDD/soil-moisture) + Clay v1.5 satellite foundation-model embeddings (256-dim monthly, Sentinel-2) · Satellite is gated ON only for concentrated-origin softs (cocoa ≈70% West Africa, palm ≈85% Indonesia+Malaysia, sugar, coffee) — a rule fixed a-priori from crop geography, not fit to results · Second GeoFM arm (Google AlphaEarth, 64-band annual) in build for cross-model explainability.

SpatioNiti Research · model: walk-forward XGBoost, rolling 5-year window, zero look-ahead · satellite arm: Clay v1.5 (Sentinel-2, 256-dim monthly) gated to concentrated-origin softs · imagery: NASA GIBS MODIS true-color, captured at the forecast date · all statistics computed from the archived forecast record. Not investment advice.