Risk by region.

Aggregation: Route A (cheap). For each region we pull the existing per-country /v1/forecast/{cc}/7day output for every constituent country and take an evidence-volume-weighted mean of the per-country max risk over the next 7 days. Weights are the number of evidence rows the country has produced in the last 30 days (floored at 1.0 so quiet countries still contribute). The headline number is weighted_avg_probability_7d.

Why evidence weighting, not population? A country's probability of a shutdown isn't a per-capita quantity — it's a yes/no event for that country's network. We don't ship a population table on the forecasting server, and using population as the weight would let India and China dominate every Asian average. Weighting by evidence volume puts more confidence on countries we actually have measurements for, which is the right epistemic move when aggregating model output.

Single high-risk countries pull averages up. If MENA looks 12% but Sudan is 80%, the region number obscures that. Always check the per-country drill-down above before citing a region figure — the “# ≥50%” column tells you how many countries are doing the lifting.

Region definitions. Continents follow UN M.49 geographic groupings (Africa / Americas / Asia / Europe / Oceania). MENA is a custom grouping — UN doesn't define MENA — composed of Northern Africa (DZ, EG, LY, MA, SD, TN) plus Western Asia, minus IL, CY, AM, GE, AZ which journalists typically don't count as MENA. World is every country in the country_geography table.

Route B (per-region model) — not shipped yet. A dedicated XGBoost on region-level features (sum of evidence, count of countries in elevated risk, etc.) would be more accurate than aggregating per-country probabilities, but it needs its own label assembly and a temporal holdout before it can replace Route A. We expect to ship it after the first month of journalist feedback on which regions get most-cited.

Calibration limits. The underlying per-country forecast model is isotonic-recalibrated only for the 30 watched censorship-heavy countries (see /atlas/forecast). Countries outside that set get raw uncalibrated probabilities, so the world and oceania averages should be read as relative rankings, not absolute calibrated figures.