SpatioNiti
How accurate are we, and how early

Accuracy, graded — and how far in advance

Two questions a buyer should ask any forecaster: how wrong are you versus just assuming a normal year (the market's default), and how early do you see trouble. Both answered below from the archived walk-forward record — every number reproducible, losses included.

Skill vs the naive baseline (assume-normal-year), leave-one-year-out 2006–2024

Advance warning — when our +180d signal first crossed, vs when the news flagged it

For every crop year with a realized anomaly beyond ±6%: the first date our published 180-day forecast leaned the right way (beyond ±0.2σ), and how many days that preceded the matching verified public event. Positive lead = we signalled first. Years where we never crossed are absent — that is the honest miss list in the tables below.

Season-by-season record

Naive baseline = predict the 5-year trailing mean (0% anomaly) — what a procurement plan assumes by default. Our prediction: season weather signal × LOYO-fitted slope. Advance-warning uses the blended +180d series exactly as archived. Skill is honest, not uniform: Côte d’Ivoire carries validated skill; Ghana partial; Ecuador, Indonesia+PNG, Nigeria+Cameroon do not clear the bar and get no volume calls. © 2026 SpatioNiti.