Methodology and data
How GeoForecast decides what to show, how serious it is, and what it cannot tell you. Last reviewed 15 November 2026.
Sources and authority
Every event names its source and how much weight that source deserves.
Every event carries its source and an authority tier, so an official reading and an automated news scrape never look equally reliable.
| Layer | Source | Authority | Coverage |
|---|---|---|---|
| Earthquakes | USGS Earthquake Hazards Program | Official government agency | Worldwide, magnitude 4.5+ |
| Disasters | GDACS (UN and European Commission) | Intergovernmental body | Worldwide; cyclones, floods, volcanoes, drought, wildfires; orange and red alerts |
| Internet disruptions | IODA, Georgia Institute of Technology | Research institution | Worldwide, by country |
| Weather warnings | WMO Severe Weather Information Centre, carrying the warnings of about 50 national weather services | Official national agencies, via the UN weather agency | Worldwide, where the national service publishes to WMO |
| Weather warnings, US | US National Weather Service | Official government agency | United States |
| Weather warnings, India | NDMA SACHET (India Meteorological Department and state disaster authorities) | Official government agency | India |
| Weather warnings, Malaysia | MET Malaysia, via data.gov.my | Official government agency | Malaysia |
| Weather warnings, Japan | Japan Meteorological Agency warning map | Official government agency | Japan, by prefecture or forecast region |
| Weather warnings, Philippines | PAGASA public alerts (CAP) | Official government agency | Philippines, by province |
| Health: outbreak news | WHO Disease Outbreak News | Intergovernmental body | Worldwide, by country; at least the last 60 days |
| Health: travel notices | US CDC Travel Health Notices | Official government agency | Current notices, by country, levels 1 to 4 |
| AI incidents | AI Incident Database, Responsible AI Collaborative | Curated research database | Worldwide; not geolocated |
| Volcanoes, storms, large wildfires | NASA EONET | Official government agency | Worldwide; activity inside the selected window; wildfires of 1,000 acres or more; no prescribed burns |
| Protests and violence in the news | GDELT Project event files | Automated media monitoring, unverified | Worldwide, located to a city, last 24 hours |
| Conflict trend by country | GDELT Project daily event files | Automated media monitoring, unverified | Worldwide, by country, last 14 complete days (UTC) |
| Published forecasts | Polymarket prediction markets (Gamma API) | Real-money prediction market. Not advice, not GeoForecast's forecast. | Open yes/no markets with at least $10,000 traded on war, military clashes, invasions, ceasefires, nuclear use, pandemics, outbreaks and disease; national elections and changes of leader; oil, gas and food commodity prices; shipping chokepoints; not mapped |
| Man-made accidents | Wikipedia Current events, "Disasters and accidents" section, each entry with the news report it cites | Automated media monitoring, unverified (volunteer-edited summary of news reports) | Worldwide; aviation, maritime, rail, road, mining, explosions, collapses, dam failures, spills, fires, crowd crushes, mass poisonings; the selected window, by day |
| Man-made accidents | Wikidata records of accidents and disasters (date, deaths, injuries, location) | Automated media monitoring, unverified (volunteer-edited database) | Worldwide; notable events, usually recorded within days |
| Flight disruptions | US FAA National Airspace System status (airport-events): ground stops, ground delay programmes, airport closures (NOTAMs), arrival and departure delays | Official government agency | United States airports; current state. See Flights and trains |
| Airspace warnings | EASA Conflict Zone Information Bulletins, each active bulletin's own page | Intergovernmental body | Conflict-zone airspace worldwide, as EASA lists it (16 active bulletins on 15 November 2026) |
| Rail disruptions | SNCB/NMBS (Belgian railways) unplanned disruption notices, read through iRail, a volunteer-run open API that republishes SNCB's own text; stations from iRail stations (CC0) | Company announcement, self-reported (the operator's own notices) | Belgium and its cross-border lines; current state |
| Flight and rail disruptions (reported) | Wikipedia Current events (all sections) and UN News, BBC, Al Jazeera, DW and France 24 news feeds, filtered by the published word rule in Flights and trains | Volunteer-edited reference, or named news outlet. Never above Moderate | Worldwide, as reported in English; the selected window |
| Politics | Wikidata: national elections, heads of state and government, coups | Automated media monitoring, unverified (volunteer-edited database) | Sovereign states: elections held in the last 90 days or scheduled in the next 180; new leaders in the last 180 days; coups and attempts in the last 180 days |
| Commodity prices | IMF Primary Commodity Price System (SDMX API) | Intergovernmental body | World monthly prices: Brent, WTI and Dubai crude; natural gas in Europe, the US and Asia (LNG); wheat, maize, Thai rice; urea, DAP; IMF energy, food and fertiliser indices; last 36 months; not mapped |
| Food price index | FAO Food Price Index | Intergovernmental body | World, monthly, with its cereals sub-index; last 36 months; not mapped |
| Shipping chokepoints | US EIA, World Oil Transit Chokepoints (oil flows); Wikipedia and Wikidata (coordinates) | Official government agency (flows) | Nine routes: Hormuz, Malacca, Suez and SUMED, Bab el-Mandeb, Danish Straits, Turkish Straits, Panama Canal, Gibraltar, Cape of Good Hope |
| Online services | The providers' own status pages (Cloudflare, GitHub, OpenAI, Anthropic, Vercel, npm, Atlassian, Discord, Zoom) | Company announcement, self-reported | Global; not mapped |
| Reported crime | police.uk (London, Birmingham), DataSF (San Francisco), Chicago Data Portal | Official government agency | Four cities; city totals only |
| AI opportunities | UK AI Security Institute, NIST, MIT News, Google DeepMind, Google AI, OpenAI, Hugging Face | Per item: official, research, or company announcement, self-reported | Announcements and research news; not geolocated |
| Headlines | UN News, BBC, Al Jazeera, DW, France 24 (RSS) | UN News: intergovernmental. The others: named news outlets | World news in English, at most 8 stories per outlet; not geolocated |
| Human needs: health capacity, vaccination, water and sanitation, food, electricity, internet, poverty | World Bank World Development Indicators (API v2), carrying figures from WHO, UNICEF, the WHO/UNICEF Joint Monitoring Programme, FAO, ITU and the World Bank's own poverty and electrification teams. Each indicator names its original producer. | Intergovernmental body | By country; annual; latest value per country, at most 12 years old. See Human needs |
| Human needs: universal health coverage index | WHO Global Health Observatory (OData API), indicator UHC_INDEX_REPORTED (SDG 3.8.1) | Intergovernmental body | By country; latest edition 2023 |
| Human needs: unemployment, GDP per person, inflation | IMF DataMapper API, World Economic Outlook | Intergovernmental body | By country; latest actual year and the IMF projection for next year, marked projected |
| AI preparedness | IMF DataMapper API, AI Preparedness Index | Intergovernmental body | By country; one edition (2023 data) |
| Country boundaries (needs map) | Natural Earth 1:110m Admin 0, as TopoJSON by world-atlas 2.0.2; country codes from the UN Statistics Division M49 table | Reference data | Natural Earth is public domain; world-atlas is ISC-licensed. Boundaries are shown for statistics only and imply no position on disputed territory |
| Where to help | ReliefWeb (UN OCHA) and WHO country pages | Intergovernmental body | Links only; each one answered when checked on 15 November 2026 |
Map tiles: OpenFreeMap, from OpenStreetMap contributors.
How severity is assigned
One scale for every layer: Minor, Moderate, Severe, Critical. Each event states its rule.
One four-level scale is used across every layer: Minor, Moderate, Severe, Critical. Each event states the rule that produced its level under "Why this severity". On the map, severity is shown by size as well as colour.
| Layer | Rule |
|---|---|
| Earthquakes | The higher of two readings. Magnitude: 7.0+ Critical, 6.0–6.9 Severe, 5.0–5.9 Moderate, 4.5–4.9 Minor. USGS PAGER impact alert: red Critical, orange Severe, yellow Moderate, green Minor. A moderate quake under a dense city can rank Severe through PAGER. |
| Disasters | GDACS alert level: red Critical, orange Severe. Green alerts are not shown. |
| Internet disruptions | Judged on each signal's latest state, so recovered disruptions drop off. Critical: a nationwide drop confirmed by two or more independent measurement methods. Severe: a nationwide drop in one method. Moderate: drops in three or more regions. Minor: individual networks or regions. |
| Weather warnings | US: NWS Extreme is Critical, Severe is Severe; lower levels are not shown. India: IMD red is Severe, orange Moderate, yellow Minor. Malaysia: Dangerous is Severe, Bad weather Moderate, Alert Minor. Japan: a JMA advisory (注意報) is Minor, a warning (警報) Moderate, an urgent warning (危険警報, Level 4) or emergency warning (特別警報, Level 5) Severe; capped at Severe because each point is the issuing observatory, not the warned area; cancelled warnings and reports more than 48 hours old are not shown. Philippines: PAGASA's own CAP severity, Extreme Critical, Severe Severe, Moderate Moderate, Minor Minor; expired alerts, all-clear notices and alerts replaced by a later one are not shown. Other countries: the highest severity the national service stated, capped at Severe, because one district's extreme warning must not mark a whole country Critical. |
| Health: WHO outbreak news | WHO does not grade the severity of an outbreak notice, so every notice is Moderate. It is raised to Severe only when the notice's own risk assessment says WHO assesses the national or regional risk as high or very high; that sentence is quoted under "Why this severity". A notice is placed on the first country named in its title; other countries are listed in the summary. Notices for a region or the world ("WHO African Region", "Global") are listed, not mapped. |
| Health: CDC travel notices | The CDC level is the severity: Level 1 Practice Usual Precautions is Minor, Level 2 Practice Enhanced Precautions Moderate, Level 3 Reconsider Nonessential Travel Severe, Level 4 Avoid All Travel Critical. A notice naming several countries is placed on each of them. A notice for a province is placed at the country's centre. "Global" and continental notices are listed, not mapped. |
| Conflict trend | A count of news reports, not verified events, so never above Moderate. A country is listed when GDELT's daily files hold at least 20 reports of assault, fighting or mass violence (CAMEO 18, 19, 20) located in it over 14 days, each reported by at least two sources. Rising: the last 7 days have at least 25% more, and at least 10 more, reports than the 7 days before; falling is the reverse; otherwise steady. Moderate from 100 violent reports in the last 7 days or a rising trend; otherwise Minor. The escalation ladder shows 7-day counts for threats (13), military posturing (15), reduced relations (16), coercion (17), assault (18), fighting (19) and mass violence (20); a stage with no reports is shown as no evidence, never as a prediction. If a daily file is missing, the trend compares daily averages and the item says how many days were read; with fewer than 10 of 14 days the layer reports an error. |
| Published forecasts | No severity: every market says "Published forecast, not a severity rating". The probability shown is the market's current price for "Yes" (0 to 100%), with the traded volume, the market's end date and when the price was read. GeoForecast makes no forecast of its own. Which markets are includedA market is included only when it is open, has an end date that has not passed, has at least $10,000 traded, is a yes/no question, and its question matches the published keyword rules: health (pandemic, epidemic, outbreak, disease, virus, flu, influenza, H5N1, measles, Ebola, Marburg, mpox, cholera, plague, hantavirus, COVID, coronavirus, Nipah, polio, dengue, West Nile) or conflict (war, civil war, military action or clash, clash, invade, invasion, ceasefire, truce, airstrike, strike, attack, troops, and nuclear use: nuclear strike, attack, detonation or war, or using a nuclear weapon). A nuclear test, deal or acquisition is not nuclear use. Labour strikes, trade wars and courts "striking down" laws are excluded. Then, in order: shipping (the question names a chokepoint, such as Hormuz, Bab el-Mandeb, the Red Sea, Suez, Malacca, the Panama Canal, the Bosporus or Gibraltar, or a blockade), prices (crude oil, Brent, WTI, natural gas, LNG, gasoline, petrol, diesel, wheat, maize, rice, soybeans, fertiliser or urea, together with a price: "$", "price" or "all-time high"; questions about production, output or exports are not prices) and politics (elections, presidents, prime ministers, chancellors, referendums, coups, impeachment, resignations, nominations, a leader "out"; mayoral, local, regional and by-elections are excluded, as in the politics layer). Election, price and shipping events can hold dozens of markets, so each keeps its five most-traded. Read from the tags geopolitics, world, middle-east, pandemics, disease, global-elections, world-elections, oil, commodities and natural-gas, most-traded first. Where no market exists, the page says so. |
| Man-made accidents | Both sources are volunteer-edited summaries of news reports, so never above Moderate. Moderate from 10 stated deaths or 50 stated injuries; otherwise Minor. A count is used only where the text states it next to "killed", "dead", "injured", "wounded" or "hospitalised" (or "death toll rises to"), in digits or words; "dozens" or "many" stay unknown, and nothing is estimated. Where Wikidata holds two figures, both are shown and the higher decides the level; no midpoint is made up. See Man-made accidents. |
| Flights and trains | Severe: an airport closed or under a ground stop, airspace closed, or a national strike or network-wide suspension, stated by the authority or the operator itself. Moderate: a ground delay programme, an airport closed during stated hours or with conditions, delays with a stated maximum of 60 minutes or more, a line closed or services suspended, cancellations, a local strike, or an EASA conflict-zone airspace warning. Minor: other delays and disruptions, and anything reported as resolved. Reports from news outlets and Wikipedia never rise above Moderate. Never Critical: the layer measures disruption, not harm to people. See Flights and trains. |
| Politics | Not a hazard: elections and changes of leader say "Political event, not a severity rating" and sit at Minor so they never outrank a hazard. A coup or coup attempt recorded in Wikidata is Moderate, as an unconstitutional change of power; it is not a measure of violence. See Politics. |
| AI incidents, headlines | Not assessed. These sources carry no severity, and none is invented. |
Location precision
Each point is only as exact as its source. Country-level signals sit at the country centre.
- Earthquakes and disasters: the source's coordinates.
- Internet disruptions are plotted at the centre of the country and drawn as hollow rings, because the disruption is national or regional, not at a point.
- Weather warnings are plotted at the centre of the warned area for the US, India and the Philippines (each warned province), at the centre of each named state for Malaysia, at the issuing JMA observatory for Japan (one point per prefecture or forecast region, listing the areas warned), and at the centre of the country elsewhere, because WMO's feed lists warnings without their area. One country point stands for all of that country's active warnings.
- Headlines and AI incidents are never placed on the map, because their sources record where a story was published, not where it happened.
- Accidents are placed at the accident's own Wikipedia article or Wikidata point where one exists; otherwise at a place linked in the report only when the sentence puts the event there ("in Golaghat", "off the coast of Kébémer", not "a bus travelling towards Zagreb"), skipping rivers and seas (Wikipedia places them at their mouth) and any linked place more than 3,000 km from the country named; otherwise at the centre of that country, marked as such. A report naming no country and no usable place is listed, not mapped.
- Politics: the centre of the country.
- Flights and trains: FAA items at the airport's own coordinates (from the FAA feed), SNCB items at the first station named in the notice, EASA bulletins and news-derived reports at the centre of the country; a bulletin or report covering several countries, or none, is listed, not mapped.
AI incident categories
Each AI incident gets one category from published keyword rules. First match wins.
The AI Incident Database feed does not categorise reports, so GeoForecast assigns one category per report using published keyword rules, applied in this order. The first match wins.
- Agents acting outside their authority. AI agents doing things nobody authorised: bypassing access controls, using credentials or funds without permission, reaching internal systems, escaping sandboxes. The clearest signal of AI acting beyond its brief.
- Autonomous and cyber attacks. AI used in or for attacks, malware, intrusions, security incidents.
- Deepfakes and disinformation. Synthetic media, voice cloning, influence operations.
- Fraud and scams. Impersonation and financial fraud.
- Safety-critical failures. Autonomous vehicles, medical systems, drones.
- Rights, bias and surveillance. Discrimination, facial recognition, wrongful identification, privacy.
- Harmful outputs and model misbehaviour. Chatbot harms, hallucinated content in official work, deception, jailbreaks.
- Legal and regulatory action. Court rulings, lawsuits, fines, enforcement.
Rules are English-first, with terms in Spanish, French, Dutch, Arabic, Chinese and Hungarian that the feed regularly carries. On the November 2026 feed, 6% of reports fell outside every rule and are listed as "Other". New phrasing can be missed until the rules are reviewed. The rules are in the open-source code, file api/_sources.js.
Protests and violence in the news
News reports read automatically, not confirmed events. Faded on the map, never above Moderate.
There is no free, verified, worldwide record of protests and political violence that is current to the day. ACLED, the standard source, requires registration and restricts republishing. GeoForecast therefore reads the GDELT Project's event files, which code world news every 15 minutes, and keeps protests (CAMEO code 14), assaults (18), fighting (19) and mass violence (20) located to a named city. Events are grouped by place, and a place appears only when at least two separate articles report it. Government landmarks such as the White House or the Kremlin are excluded, because news is datelined there rather than happening there.
These are news reports, not confirmed events. Several articles may describe one incident, a place may be mentioned in passing, and a busy news day in one city can outweigh a quiet war elsewhere. So they are drawn faded on the map, kept in their own section, never counted in the verified headline or priority list, never announced as alerts, and rated no higher than Moderate: Moderate from 5 articles for violence and 10 for protests, otherwise Minor.
Man-made accidents
From Wikipedia and Wikidata, both volunteer-edited. Never above Moderate.
Two volunteer-edited sources, both built from news reports. Wikipedia's Current events portal lists accidents the same day, each with the news report it cites; GeoForecast links that report, not the encyclopedia. Wikidata records notable accidents with dates, deaths, injuries and locations as fields, usually within days. When a Wikipedia report concerns an event Wikidata also records, the two are merged into one item that keeps the news report as its link. Only dates recorded to the day are used. Wikidata records that are also crimes, attacks, bombings, military operations or weapons tests are excluded, as are natural hazards and attacks in the Wikipedia section: other layers cover those.
The type of accident comes from published word rules, first match wins: aviation (an aircraft, helicopter or flight with a crash, collision or emergency landing), maritime, rail, road, mining, dam failure, explosion, structural collapse, spill or leak, fire, crowd crush, mass poisoning. Limits: coverage follows English-language news and Wikipedia's editors, so small accidents, and accidents in countries with little English coverage, are under-represented; this is not a complete record. Deaths and injuries are what the text states on the day and are often revised. Sources checked and not used: Aviation Safety Network (its security service blocks automated requests), NASA EONET's man-made category (nothing since 2018), GDELT's knowledge-graph themes (a theme is tagged to an article, not to a place) and ReliefWeb (needs an approved app name).
Politics: elections, leaders, coups
National elections, new leaders and coups, from Wikidata, for sovereign states only.
- Elections: national elections and referendums held in the last 90 days or scheduled in the next 180, typed presidential, parliamentary (legislative or general), referendum, or other national election. Local, regional, mayoral, party and by-elections are out of scope. A scheduled election is a date someone has published, not a prediction; it is shown with how exact that date is ("29 Nov 2026", or "December 2026 (day not set)"). Dates known only to the year are not shown, because Wikidata stores them as 1 January. A record whose name gives a year two or more away from its date is treated as mislabelled and left out. Two-round elections list both dates.
- New leaders: people whose term in their country's head-of-state or head-of-government office (as Wikidata records those offices) started in the last 180 days.
- Coups: events recorded as a coup d'état or an attempt in the last 180 days.
Limits: Wikidata is edited by volunteers and can be late, incomplete or wrong; every item links to its article so it can be checked. Acting and interim office-holders are not distinguished. Results are not shown. Published odds on elections and leaders come from Polymarket in the forecasts list, labelled as a real-money prediction market, never as GeoForecast's forecast.
Flights and trains
One item per disruption, with only what its source states. First-hand for the US, conflict-zone airspace and Belgium.
One item per disruption, each with what its source says and nothing more: the cause, the impact (numbers only as stated), the recovery (an end or reopening time, only when one is stated), the measures the authority or operator says it is taking, and whether it is ongoing, scheduled or resolved. Where the source does not say, the field says "not stated"; GeoForecast never fills it in.
Covered:
- Flights in the United States, from the FAA's own status feed: ground stops (with the FAA's stated end time and its rating of the chance of extension), ground delay programmes (with the stated average and maximum delay and the flights included), airport closures from NOTAMs (a NOTAM says what is closed, not why, so closures carry no cause), and general arrival and departure delays with the FAA's reason. NOTAMs that close an airport only to unscheduled general aviation, and de-icing on its own, do not affect scheduled flights and are left out.
- Conflict-zone airspace, from EASA's bulletins: the cause is the first sentence of EASA's own description after the sentence saying who issued it, the impact is the affected airspace, the measures are EASA's recommendations to operators, and the recovery is the bulletin's validity date, which is a review date, not a forecast of when the risk ends. A bulletin covering several countries is listed, not mapped, so no single point is invented.
- Trains in Belgium, from SNCB/NMBS: unplanned disruptions, with SNCB's stated cause and advice, placed at the first station named; planned engineering works are left out (they are routine and dated in free text), except announced strikes, shown as scheduled.
Elsewhere, only as reported. Wikipedia Current events and five news outlets are read with a published word rule: an item counts only when a transport word (flights, airport, airspace, airline, air traffic; trains, railway, rail services, metro) and a disruption word (cancelled, suspended, closed, shut, grounded, ground stop, halted, diverted, delayed, disrupted, stranded, reopened, resumed, or a labour strike or walkout) stand within 80 characters of each other in the same sentence. "Strike" counts only as a labour strike, never an air, drone or missile strike, and a crash or derailment alone is an accident (see Man-made accidents). Cause, impact, recovery and measures are short phrases copied from the text, never rewritten. A cause is taken only when the text asserts it: not after "not", "no evidence", "ruled out", "whether", "if" or "unclear" in the same clause (on 15 November 2026 a first version reported "cyberattack" from "there was no evidence this was a cyberattack"; that is now a tested case). These items are marked news-derived, link the report, are placed at the centre of the first country named (or listed when none is named) and never rise above Moderate.
Not covered, and why (checked 15 November 2026): UK National Rail incidents (the open data feeds need a registered token; HTTP 401 without one), Dutch NS (subscription key required; HTTP 401), Deutsche Bahn (its API marketplace answers HTTP 401 without a registered client id and secret; the bahn.de traffic page is a web application, not a documented feed), SNCF (its open-data portal has punctuality and timetables but no live disruption dataset; the SNCF API answers HTTP 401 "no token"), Amtrak (alerts are a web page, not a feed), and EASA's JSON export (its address has a query string, which EASA's robots.txt disallows; the bulletin pages themselves are allowed). No free, keyless source of airline schedules, cancellations by airline, or rail disruptions in Asia, Africa or the Americas outside these was found. Coverage outside the US and Belgium therefore follows English-language news.
Prices that move household costs
Monthly world prices and nine shipping chokepoints. Observed, never predicted.
Monthly world prices from the IMF Primary Commodity Price System (crude oil, natural gas, grains, fertiliser, and IMF price indices) and the FAO Food Price Index, for the last 36 months, at /api/feeds?layer=markets. Names and units are as the IMF publishes them. For each series GeoForecast shows the latest month, and the change over 1, 3 and 12 months, compared only across exact month gaps.
Spike rule, a statistical flag on published prices, not a forecast: a spike is a latest 1-month change more than two standard deviations (sample) of the series' earlier monthly changes in the 36-month window, upward; a drop is the same downward. The latest change is not part of its own yardstick. With fewer than 12 earlier monthly changes, or a missing month, nothing is judged. A series whose latest month is more than three months old is marked late.
Shipping chokepoints. Nine routes with oil flows from the US EIA (first half of 2025, crude oil and petroleum liquids; EIA gives none for Gibraltar, so none is shown) and coordinates from Wikipedia or Wikidata. Each carries a signal, not a prediction: news-coded violence reports in its bordering countries (listed on the item) over the last 7 days against the 7 before, from the conflict layer, using that layer's rising, falling and steady rule. Countries below the conflict layer's listing threshold are named and not counted. Related Polymarket questions are shown with their odds, volume and end date where they exist. The precedent is quoted from the EIA: after Houthi attacks on commercial ships in the Red Sea began in November 2023, about 18% less crude oil passed the Bab el-Mandeb in December 2023 than the 2023 average to November, a Persian Gulf to Rotterdam voyage takes 19 days via Suez and nearly 35 around the Cape of Good Hope, and longer routes put upward pressure on freight rates (EIA, 1 February 2024).
Limits: these are world benchmark prices, monthly averages published weeks after the month ends; they are not the price in any shop or pump. Petrol and diesel pump prices are not covered yet: the IMF lists gasoline and heating oil but publishes no values for them, and the US EIA's data service needs a key. The EU Weekly Oil Bulletin is free but published only as spreadsheets, which GeoForecast does not read yet. GeoForecast does not predict prices; published odds on oil prices appear in the forecasts list.
Regions and search
Regions are approximate boxes. The filter applies to everything on the page.
The region filter applies to the map, headline, tiles, lists and charts. Regions are approximate boxes: Middle East and North Africa, Europe, Sub-Saharan Africa, North America, Latin America and Caribbean, Asia, and Oceania and Pacific. AI items and online services have no location, so they stay worldwide and say so. Search matches an event's title, place, type and country.
Reported crime and case status
City totals from police open data, with where each case stands. Never mapped one by one.
There is no worldwide crime feed. GeoForecast includes the cities whose police publish recent, open data that records what happened to each case. It shows city totals only: individual crimes are never mapped, which protects the people involved and keeps the map calm. Crime is drawn in grey, because the number of crimes is not a severity.
Each case is placed in one of four groups from the outcome the police published:
- Under investigation: "Under investigation", "Awaiting court outcome", "Open or Active".
- Action taken: an arrest, charge, caution, penalty notice, court result or local resolution.
- Closed without charge: "Investigation complete; no suspect identified", "Unable to prosecute suspect", "Unfounded", or closed for exceptional reasons.
- Status not published: no outcome recorded. This is kept separate and is never counted as closed. Anti-social behaviour in the UK is recorded without outcomes, and Chicago publishes only whether an arrest was made, so Chicago cases without an arrest appear here.
UK figures cover a one-mile radius of each city centre for the latest month police.uk has published, about two months behind. San Francisco and Chicago cover the whole city for the last 30 days. Excluded, and why: Los Angeles (its open dataset stopped in December 2024) and Manchester (police.uk returned no records for the latest month when checked on 15 November 2026). Outcomes change as cases progress, so recent months show more cases under investigation than older ones will.
AI opportunities
Releases, research and policy, each with its own authority tier.
Beside AI harms, the report lists new AI capabilities, research, safety work and policy, taken directly from the organisations doing the work. Government and research sources keep their own tier. Posts from AI companies are labelled Company announcement, self-reported: they are the primary source for what a company released, but they are promotional, and listing one is not an endorsement. Each item gets one category by keyword rules, in this order: Safety and evaluation, Policy and governance, Science and health, Open models and tools, Model and product releases, then Other developments. NIST news is filtered to items about AI. The window is at least seven days, because these sources post a few times a week.
New-item alerts
A short notice for each new severe event or AI item, at most three at a time.
While the page is open it refreshes every five minutes. A short notice appears for each new severe or critical event, and for each new AI incident or opportunity, at most three at a time. Nothing is announced on first load, except a single summary of what is new since your last visit, which your browser remembers locally.
What you could do
Options to weigh, not instructions, each naming the public guidance it comes from.
The Overview and every event offer three groups of options: getting ready if you are in or near it, helping from anywhere, and keeping it in proportion. Preparedness options come from Ready.gov hazard pages and, for air quality and mental health, from WHO; giving options link to ReliefWeb (UN OCHA) appeals for the country and to the Red Cross and Red Crescent; the family-tracing option links to the ICRC. Where an authority gives an instruction, such as moving to high ground after a strong coastal earthquake, it is quoted as theirs. A few practical lines, such as checking with an airline before travelling, have no named source and show none. With Near me on, a faraway event leads with how to help and states its distance; GeoForecast does not estimate anyone's personal risk. Your local authority's advice always comes first; the WMO register lists the official warning service for each country.
Live news and webcams
Broadcasters' own news streams and established webcams, one at a time, muted.
The Live panel plays broadcasters' own 24-hour news streams (Al Jazeera English, DW News, France 24, Sky News, Euronews, CNA, NHK World-Japan, ABC News Australia, Bloomberg) and city webcams from established operators (EarthCam, Fuji News Network). Each was checked on 15 November 2026 to be live, owned by the named broadcaster, and open to embedding. GeoForecast does not edit, verify or endorse what they show. Webcams from anonymous channels restreaming conflict-zone cameras were found and deliberately left out, because their origin cannot be verified.
How the headline numbers are counted
Events in your time range and things in effect now are counted apart, each situation once.
Events and things in effect are counted separately. Earthquakes, disasters, natural events and AI incidents are events: they are counted only if they fall inside the time range you chose (24 hours, 7 days or 30 days). Weather warnings, outbreak notices and internet disruptions are in effect now: they are counted whatever day they were issued, and the headline says so ("Plus N warnings and notices in effect now"). A travel health notice issued three months ago is current; it did not happen this week.
Each situation counts once. Successive WHO updates and the CDC notice for the same outbreak in the same country are one item, shown with "+N related notices". A US weather warning issued county by county is one item per warning type and state. Countries are counted by country, not by place name.
Crime: the share still under investigation is out of cases whose status the police publish. Chicago publishes only whether an arrest was made, so its other cases are excluded rather than counted as open. Cities publish different periods; the city view shows each period.
AI incident reports are dated when the database publishes them, often in batches, so no daily trend is drawn for them.
Crisis simulator
A what-if model of 13 linked crises and AI capability. A scenario model, not a forecast.
A scenario model, not a forecast. It shows what a set of assumptions implies; it cannot say when, or whether, the world collapses.
It is a system-dynamics model of thirteen stresses: armed conflict, great-power confrontation, outbreaks, climate and hazards, infrastructure and cyber, AI misuse and loss of control over AI (where a shock can start); energy prices, shipping and supply chains, food prices and hunger (knock-on systems); and economic downturn, displacement, and unrest (where it lands).
Each year every stress grows on its own (logistic growth, scaled by the "amplify" setting), is pushed up by the stresses linked into it, and is pulled down by resilience times the strength of the response. About 37 links connect them. Where a source documents a pathway (UNHCR, WFP, the World Bank, the US EIA, the IPCC, IDMC, the IMF, CISA) the link names it; the rest are labelled model assumptions. Every strength is an assumption.
Resilience is rebuilt by investment and cooperation and worn down by stress. Wellbeing combines them: 100 × resilience0.45 × (1 − weighted stress)1.3. Its bands name five states: Utopian (72+), Stable (52+), Strained (35+), Crisis (15+) and Collapse (below 15, defined as global coordination failing so that food, health and power systems stop working across regions).
AI capability runs alongside on a 0–1 scale with five milestones, following the steps the AI 2027 scenario describes without its dates: routine tasks, coding automated, AI doing AI research better than the best human teams, AI surpassing people at most cognitive work, and superintelligence. It grows logistically; at speed ×1 the "AI researcher" milestone arrives about 3.5 years from today, an assumption. Capability pushes AI misuse up, and capability cubed times (1 − oversight) pushes loss of control up. Past "coding automated" it displaces jobs, which investment can cushion. When oversight is strong, capability strengthens the response to health, climate, economic and food stress.
Decision points. Four turning points fire when a stress crosses a line: AI race or coordinated slowdown (at "AI researcher", after AI 2027's two endings), escalate or negotiate (great-power confrontation at 45%), share vaccines or hoard (outbreaks at 45%) and export bans or open trade (food at 45%, see the IFPRI export-restrictions tracker). Each option changes growth, responses, oversight or cooperation from then on. Today's course assumes the option closest to recent behaviour and says why. Playing the simulator pauses at any decision the person has not chosen. Endings compare the person's scenario, the same scenario with the other AI choice, and the two presets.
Starting events and the unchanged baseline. A scenario can start from up to four shocks: live events (from the person's selection or the most severe events now), each sized by its severity alone (0.12 for Minor to 0.30 for Critical, an assumption), and preset what-ifs. With any shock set, every chart and ending also shows the same settings without the shocks. With two or more, each shock is also run on its own, and the page compares "all together" with "one at a time, added up" from one run of the person's settings. The model is not additive (shocks interact through its links, its limits and the resilience they share), so the page reports the difference instead of assuming that effects add up. A short list of the model's limits sits above the controls.
Play. The simulator's Play mode loads a separate game module only when opened. It receives the same model and the live events; if the module is not available, the page says so and the model is unchanged.
Scenario numbers. People facing acute hunger (WFP, 266 million in 2026), people forcibly displaced (UNHCR, 117.8 million at the end of 2025), deaths in organised violence (UCDP, at the rate of its latest 30 days), Brent crude (IMF) and world growth (IMF projection, 3.1% for 2026) start from those real figures and move with the matching stress: today × ek × change in stress, or linearly for growth. The factors k are assumptions, listed in the simulator. Every such number is labelled "scenario value".
Starting point from live data. Each starting stress is a simple formula of dashboard numbers, listed with its inputs in the simulator: for example conflict uses the share of countries with rising violence reports and the average published odds of the five most-traded escalation questions (one market per question, and never a ceasefire or peace market). Energy uses Brent's 12-month change, food the FAO index, supply chains the chokepoints with rising violence nearby. Where no live measure exists (displacement, loss of control over AI) the starting value is a stated assumption.
Ranges, not dates. Every setting is run 120 times, each time with every growth rate, link strength and the AI speed varied at random (log-normal, spread 0.3, 0.3 and 0.35). The page shows the middle 80% of runs and the share of runs in crisis or collapse by a given year. These are shares of model runs, not the chance that anything happens. The coefficients were chosen so that today's settings drift slowly, the utopian preset recovers over a decade and the dystopian preset collapses within years; they are not fitted to history, because no dataset of past global collapses exists.
Chain reactions
Pick one or more shocks, preset (for example a Gulf war closing the Strait of Hormuz) or a live event sized by its severity, and the model plays each forward 120 times over five years, using the simulator's settings and the person's links. In every run it plays the same world twice with the same random rates, with and without the shock. A crisis counts as hit when it rises 3 points above the no-shock run; its cause is the hit crisis pushing hardest into it at that moment, or weaker resilience when no linked crisis explains it. A chain is the path from the shock to each crisis it reaches. In the list of every chain, impact is that crisis's weight in wellbeing times its largest extra stress. Where a prediction market prices a related question, its odds are shown beside the chain, separately, and never mixed into the ranking.
Outcomes: what the shock leads to. The view first groups the same 120 runs by where each ends up and how it gets there, and shows 2 to 4 outcomes, most common first. No extra runs are made: each run's record (its random draws, its wellbeing with and without the shock, its chains, its end state) is kept from the chain search.
- Grouping. Runs are grouped by the state the shocked world is in after five years (utopian, stable, strained, crisis, collapse). If one group holds 60% or more of the runs, it is split once, by how it got there: by timing, if the times it entered that state have a clear break (a gap of a year or more, with at least 10% of all runs on each side); otherwise by the same run without the shock, which ends in the same state ("shock or not"), a better one ("the shock tips it") or a worse one. Groups of fewer than 3 runs are listed by name, not shown as cards. With more than four groups, the smallest joins the group closest to it in end wellbeing. Names come from that rule: the end state, plus how it got there.
- Share. "In 74 of 120 runs" is a share of model runs, not a real-world probability; the page says so beside every outcome. Shares of the cards and the rarely seen groups add up to the 120 runs. When one outcome holds 95% or more, the page says that nearly every run ends the same way and still shows what tips the rest. Most common first; where two outcomes have the same share, share × impact (the wellbeing lost against the same runs without the shock) breaks the tie, as it ranked chains before.
- When. The time each run first entered its end state: the median, with the 10th to 90th percentile as the range. A shock that lands hard puts some runs there at once ("within weeks").
- Main chain. The most common chain among that outcome's runs. When several are about as common (within 20%), the one most typical of this outcome compared with all runs is chosen. Selecting an outcome lights its main chain in the network, and the network's boxes then show the share of that outcome's runs reaching each crisis.
- What tips it. Every rate in a run is today's value times espread × z, with z drawn at random; the model keeps each run's z for every crisis's own growth, every link and AI progress. For each of these assumptions the page compares the mean z of this outcome's runs with the runs of the next worse outcome (or the next better one, for the worst), divided by what chance gives for groups of those sizes. The largest difference is named, in plain words ("Stronger link: energy prices → food and hunger"), with how many runs on that side of the middle ended here. Below 3 (about what the largest of 50 chance differences reaches) the page says no single assumption stands out. It shows which assumption the model's outcome is most sensitive to, not a cause in the world.
- Key numbers. The three scenario values that move most in that outcome at year five, from today's real figures moved by the median stress of its runs (see the simulator's scenario numbers).
- Every chain. Behind "Show every chain": chains found in every run are folded into one line, since they do not tell runs apart; the rest are drawn by step, with one box per crisis per step, saying how many paths reach it. Directly shocked crises are labelled "Starting point".
- When the outcomes say little. The five states are wide (strained covers wellbeing 35 to 52), so many shocks end most runs in the same state; the split by the run without the shock then carries the difference. The draws vary rates only; the starting point, decisions and links are the same in every run, so outcomes do not cover what the model leaves out.
Connecting crises yourself. Any built-in link can be switched off or made weaker or stronger, and any two crises can be connected with a link of your own. Your links are labelled as yours, change the simulator too, and are kept in the page link.
One home for chains. Chains are drawn only in the Chain reactions view (since 2026-11-19; before, a map panel and the simulator drew them too). With no shock picked it shows two things kept apart: the documented pathways, each citing a source, with their two ends coloured by live data (an arrow is lit when both ends are active, which is not evidence that one is driving the other); and the model's whole network, where links with a source are solid and model assumptions are dotted. Overview, the simulator and every event link to it, carrying their starting points.
Your selection and the connection check
Keep events in a selection, then simulate them or check what links them.
Any event can be added to a selection that stays with the person across views and in the page link (event ids only; no personal data). The selection can start a simulation, start chain reactions, or be checked for a connection.
The connection check never says that one event caused another. For each pair it reports four kinds of evidence, labelled separately: observed (great-circle distance, same country, days between the reports); documented links between their kinds of crisis (a direct link in the model that cites a source, such as conflict to displacement, UNHCR); model assumptions (a direct link with no source, or a path of up to three steps through other crises); and nothing. A link that runs against the order in which the two were reported is flagged. Two reports within 500 km (or in the same country) and 14 days with no link are called close in place and time, which can be coincidence. When none of this applies, the page says "No supported connection identified": a valid answer, not missing data. Events outside the model (reported crime, for example) are compared by place and time only. Report times are when sources published, not always when things happened.
Game theory lens
Off by default: a two-sided escalation game per conflict hotspot. A model, not a forecast.
Off by default. When on, each conflict hotspot gets a two-sided escalation game: each side chooses to hold or escalate. Payoffs to one side: both hold 0; it escalates while the other holds g (gain from striking first); it holds while the other escalates −x (cost of being caught unprepared); both escalate g/2 − c (cost of open war). If c < g/2 + x, escalating is each side's best move whatever the other does (an escalation spiral). Otherwise it is brinkmanship: in the mixed equilibrium each side escalates with probability q = g / (g/2 + c − x), and a direct clash happens with probability q².
Starting values: g = 1 + week-on-week momentum in violence reports (0 to +100%); x = 1; c = 1 + 2 × share of reports that are armed fighting + 2 × the published odds that a ceasefire for that country holds. Every value can be changed on the card. When the lens is on, the simulator multiplies conflict growth by 1 + 0.5 × the average clash chance across hotspots. This is a model of incentives under stated assumptions; real decisions involve far more than two sides and three numbers.
Outlook: what GeoForecast does and does not predict
GeoForecast makes no predictions of its own. It shows observed trends and other people's market odds.
GeoForecast makes no predictions of its own. The Conflict view shows two things. The first is observed trends: 14 days of news-coded events per country, and which steps of an escalation ladder (relations cut, threats, military posturing, coercion, assaults, armed fighting, mass violence) appear in the reports. The second is published forecasts: prices from Polymarket, a real-money prediction market, with the end date and amount traded. A market price of 12% means traders treat the event as about a one-in-eight chance by that date. Markets can be thin, wrong or manipulated, and nothing here is advice. Disease outlooks work the same way. Where no published forecast exists, the page says so.
Escalation cards are ranked by reports of armed fighting and mass violence. News coding files ordinary crime as "assault", and coverage leans towards countries with large English-language media. GDELT also often places an event at the country carrying it out, such as a strike abroad recorded at the striking country.
Human needs: annual official statistics
Yearly official statistics per country, with their year. Not live events.
The needs view shows, per country, the state of basic needs: health capacity (hospital beds, physicians, the universal health coverage index, out-of-pocket health spending), vaccination (DTP3, measles), water and sanitation, food (undernourishment, food insecurity), electricity and internet, jobs and income (unemployment, GDP per person, inflation, poverty at $3.00 a day), and the IMF AI Preparedness Index.
- These are yearly official statistics, not live events. Most are one to three years old when published, and survey-based ones (poverty, some health figures) can be older. Every value shows its own year. The latest value for each country is used; a country whose latest value is more than 12 years old is left out, and the indicator says how many were left out.
- Projections are marked. The IMF publishes estimates and projections in its World Economic Outlook. The API does not flag them value by value, so GeoForecast follows the Outlook's convention: every year from the release year onward is a projection (the October 2026 release projects 2026 and later). Two figures are shown: the latest year before the release, and the IMF projection for next year, labelled projected. For some countries the latest year is itself an IMF estimate; the API does not say which.
- Regions and income groups are never shown as countries. The World Bank and IMF return them alongside countries; GeoForecast keeps only codes on the World Bank's country list (or, if that list is unavailable, the UN M49 list). The world figure, where the source publishes one, is shown separately.
- Values are rounded to three significant figures. Use the source link on each indicator for the exact figure.
- The poverty line is $3.00 a day (2021 prices). The World Bank moved its international poverty line from $2.15 (2017 prices); the series code is unchanged.
- A missing indicator is shown as missing. If one source does not answer, the others are still shown and the page names what is missing.
- Higher is not always worse or better. Each indicator says which direction is worse. Inflation is marked "higher is worse", but falling prices are not good news either.
What no open global data covers. These are stated plainly rather than estimated:
- Availability of essential medicines. WHO's SDG 3.b.3 series covers only 25 countries (17 with a value from 2016 or later); older generic-medicine series stop in 2013. WHO indicator page.
- Medical oxygen. No WHO Global Health Observatory indicator covers it, and no other free official per-country series was found. WHO on medical oxygen.
- Hospital beds available right now. No open global source publishes current bed availability. The hospital beds figure is a yearly count of beds that exist, not beds free today.
Checked against the WHO Global Health Observatory's list of 3,092 indicators on 15 November 2026. The World Bank's own copy of the universal health coverage index is archived, so that index is read from WHO, which produces it.
Offline use and search
Data is saved in your browser for 30 days, so search and the page work with no connection.
Everything the page loads is saved in your browser for 30 days (at most 5,000 records). The page, the latest data, the fonts and the map areas you have viewed are cached, so GeoForecast opens with no connection.
Search is in the header on every view (Ctrl+K, Command+K or / to jump to it). It searches live records, records saved on this device, countries (with their two- and three-letter codes and common names such as DRC, UAE or USA), published odds, views and tools, sources and these methodology sections. Ranking runs in your browser with no download: BM25 word matching (title, place and country count more than the summary), three-letter fragments so misspellings still match, a one-typo allowance per word, a short list of hazard synonyms ("flood" also finds rain warnings), and a recency and severity order. The three rankings are combined with Reciprocal Rank Fusion. Each result says whether it is live (and when it was fetched) or saved (and when it was saved). When you are online, a search about a topic also refreshes that topic's feeds if they are more than five minutes old; when you are offline, the panel says so and shows saved results. Nothing you type is sent anywhere except those feed refreshes, which carry no query text.
Neural embeddings were considered and not used: a useful on-device model is a download of 20 MB or more, the opposite of what a weak connection needs. The full results page also answers with counts and quotes of the matching saved records, each cited, and links them by country, region, type and source. No text is generated.
Limitations
What the data cannot tell you.
- Verified conflict data runs behind the news. ACLED monthly country totals (from HDX), UCDP candidate events (released monthly, about a month behind) and ACAPS crisis severity (scored monthly) are refreshed once a day. ACAPS's list of active risks is shown with the date ACAPS last reviewed it, and is never mapped, because those events have not happened. The faster conflict trend counts news reports coded by GDELT: busy news markets produce many reports and a quiet war few, so compare a country with itself over time.
- Weather coverage depends on each country publishing. About 50 national services send warnings to WMO's feed and many do not, so a country with no point may be quiet or may not publish there. India and Malaysia are read direct because they do not publish to it; Japan and the Philippines are read direct because their own feeds are more detailed, and are not repeated from WMO. Countries read direct: United States, India, Malaysia, Japan, Philippines. Warning names appear in the issuer's language.
- Disaster and outage times are update times, not the moment the event began. That is why only earthquakes appear in the per-day trend chart.
- "Most affected places" is approximate for earthquakes, taken from the nearest named place in the USGS description.
- A source can be late or wrong. GeoForecast does not correct source data; it validates it and drops records that fail (see below).
- Accidents and politics come from volunteer-edited sources. Wikipedia and Wikidata follow the news and their editors' attention; they miss events and can be wrong. Treat them as leads to check, linked on every item.
- Flights and trains are covered first-hand only in the United States (FAA), for conflict-zone airspace (EASA) and in Belgium (SNCB). Elsewhere a disruption appears only if Wikipedia's editors or five English-language news outlets report it, so absence on the map is not evidence that services are running normally.
- Prices are world benchmarks, monthly. They lag by weeks and are not local prices. Pump prices for petrol and diesel are not covered (see Prices).
- The game theory lens is not live. They will not be switched on until they run on real inputs and have been tested against past events.
Reliability and failures
A failing source is always shown as failing, never as a quiet world.
- A missing source is always shown as missing. If a source fails, a banner says so and the source panel marks it unreachable. "Nothing on the map" never silently means "the feed is down".
- Last good data, clearly marked. If a source fails after a successful fetch, its last good data is shown for up to six hours, labelled stale with its age.
- Validation. Every record is checked before display: coordinates in range, a plausible timestamp, a severity on the scale, and a link that is an ordinary web address. Records that fail are dropped and counted in the API response.
- Freshness. Most sources are re-read every 10 to 30 minutes (flights and trains every 10); accidents hourly, the conflict trend and markets every 12 hours, politics daily; and the page refreshes every 5 minutes while open. Live status for every source is at
/api/status. - Tested. Each source's processing is tested against recorded real responses, including the failure cases above.
Citing GeoForecast
Cite the original source first, with the time you read it.
Cite the original source shown on the event, with the time you accessed it. If you cite GeoForecast as the aggregator, use: GeoForecast, data from [source], accessed [date and time, UTC], https://[this site]/?[view link]. The "Copy link to this view" option produces a link that reopens the same layers, filters and selected event.
Data access
No public API yet. Export any view as CSV or GeoJSON.
GeoForecast's own data service is not public: it answers only the GeoForecast site. Every event links to its original source, which you can use under that source's own terms, and any view can be exported from the page as CSV or GeoJSON for your own work. API access is planned for GeoForecast Pro; register interest from the site.