Published methodology

How the map is scored

Every color on the map comes from the formula on this page. Nothing is hand-tuned, and nothing is a prediction — the score describes the recent past, from open data anyone can check.

Source

Events come from the GDELT Project, which machine-codes worldwide news coverage in over 100 languages into structured events, updated every 15 minutes. Omnilus ingests that feed continuously. Map geometry is Natural Earth.

What counts as an event

Only conflict-relevant event types are scored, by CAMEO root code: 14 protest, 17 coercion, 18 assault, 19 armed clash, and 20 mass violence. An event must be reported by at least 3 independent sources before it counts — a single article never colors a region.

The score

For each country (and each US state), over the trailing 7 days:

raw score = Σ |Goldstein severity| × log₁₀(1 + article count)

Goldstein severity is a standard −10…+10 scale of how destabilizing an event type is. The logarithm means heavy coverage raises weight with diminishing returns, so one viral story can't dominate a region's score.

The two numbers you see

Risk index (the color) is the region's percentile rank of raw score within its cohort — countries rank against countries, US states against US states. 100 means the most severe 7-day window on the map right now.

Trend (the arrow) compares this week to the region's own trailing 90-day norm. A country in long-running conflict stays dark on the map, but its arrow only rises when things are worse than its normal. No trend is shown until a region has at least 14 days of history.

Zoom tiers

At global zoom you see countries. Zoomed into the United States, the map resolves into states, scored the same way against each other. Omnilus deliberately does not color counties or cities: news-derived data at that resolution measures news attention, not risk, and we won't paint confidence we don't have.

Known limits

GDELT measures the world through news coverage, so under-covered regions are under-scored and media-saturated events are amplified — the source-count floor and log damping reduce this but don't eliminate it. Machine coding also makes mistakes. Treat the map as a structured reading of open reporting, not ground truth, and click through to the underlying coverage — every region links to its sources.

Questions or corrections: the formula above is the whole method. If you can improve it, we'd like to hear.