Volume ledger — forecast vs actual, origin by origin

Crop-year production anomaly predicted from the season's physical signal (leave-one-year-out, zero look-ahead), against FAOSTAT actual tonnes. Volume calls in kilotonnes are issued only where the calibration has validated out-of-sample skill (Côte d'Ivoire, Ghana). Elsewhere the signal is shown and explicitly marked uncalibrated — no invented confidence.

Prediction lineage — trace any number to its sources

Pick a belt, horizon and date: the full chain from raw data sources through feature values, model vintage and blending to the published forecast, the realized outcome, the error, and every later revision of the same target date.

What-if — conditional scenarios on the fitted model

Finite-difference sensitivities of the trained models, evaluated at the latest observed state. These are conditional deltas (“if X shifts by one standard deviation”), not data.

Demand and buyer side

Who takes the beans: destination imports (UN Comtrade, annual), world grindings and balance (ICCO), and the demand context a buyer sees. Grind-hub imports are the physical demand pull; the ICCO balance is the market-wide scoreboard.

Data explorer - every input, visible

Each dataset the system uses, charted raw. Nothing hidden, nothing synthetic: gaps render as gaps.