If the world gets scary, where’s safe?

05How the model thinks

A haven is only safe from something. So the model does not rate places once. It rates them on twelve things that keep people alive when systems fail, each built from published data, and lets your scenario decide which of those things matter. The structure is the one the humanitarian risk indices use (hazard, exposure, capacity to cope), turned round to ask where people are best placed rather than where they are worst off.

How your worries become a scenario

Each card is a specific threat, and each one declares which of six underlying risks it pushes and how hard, from the literature it leans on. Tapping sets how much weight it gets: a little worried counts for 40% of its push, worried for 70%, very worried for all of it. The time band sets when it arrives: within five years, five to twenty, more than twenty, or not in your lifetime. Within each risk the strongest card sets the level and every other card adds a tenth of its own push, so worries compound but their order never matters. The weights and the table below are our judgement, stated so you can dispute them.

Twelve channels, from published data

Each channel is the weighted mean of its indicators. Every indicator is scaled across all countries that report it (trimmed at the 5th and 95th percentiles so one outlier cannot stretch the scale) and turned so that 100 is always the safer end. A gap is filled with the median of the country's sub-region, and the page tells you when a place rests on partial data.

How each risk uses the channels

Each risk draws on the channels in two profiles. At moderate severity systems still work, trade still flows, and institutions, money and hospitals matter. At extreme severity trade and states fail, and what a place can grow, burn and drink, and who can reach it, matter more. The model slides from the first profile to the second as you push a risk from 40 to 80. The direction is well supported (Xia et al. 2022; Moersdorf et al. 2024; Jehn et al. 2025; Boyd & Wilson 2023); where exactly the switch happens is our assumption. The table shows each loading as moderate → extreme.

The arithmetic, in plain words

• How much each risk counts. Severity counts more than linearly (a catastrophe is worse than twice a crisis), and a risk that peaks further away counts less: half as much for every fifteen years of distance.
• No trade-offs at the edge. Within each risk, channels combine as a weighted geometric mean, not an average, so a place with no water is not rescued by good government. The UN's Human Development Index and the INFORM risk index made the same choice for the same reason.
• Risks feed each other. A risk that arrives first raises the severity of those that follow: climate stress raises the odds of conflict, war breaks trade, rapid automation shocks the economy. Eleven links, each small, each shown on the page when it fires.
• Beyond geography. Some share of each catastrophe is extinction-level, and no place shelters anyone from that. We take it from the ratio of extinction to catastrophe in the largest forecasting tournament on these risks (about 21% for AI, 9% for pandemics, 4% for nuclear war). Those forecasts were made in 2022; surveys up to mid-2026 ask for one figure or the other, never both from the same people, so nothing newer can replace them yet. The war figure covers nuclear war only, so it slightly overstates the share for war as a whole; no climate or economic collapse has been judged extinction-level. We phase it in above severity 60, take it out of the ranking and show it on the page as the no-refuge share.
• The composite. The worst risk sets the floor; the others raise it in proportion to the room left. It answers "how bad does it get"; the timeline answers "what comes first".
• Regions. 149 regions of 29 large countries start from the national score and move up or down on each channel, by at most 25 points, where there is cited evidence of a difference: missile fields, water stress, heat, cyclone coasts, fault lines, distance from the capital. The adjustments are centred, so a country's regions average back, by population, to its national score: the national indices set the level, the regional evidence only the differences.
• Wealth is not counted twice. The climate channel uses ND-GAIN's vulnerability score, not its headline index, because the headline adds a readiness measure that is largely wealth and governance, which already have their own channels. Critics of earlier haven rankings made exactly this point.

The presets

Two presets are our reading of the Existential Risk Persuasion Tournament (Forecasting Research Institute, 2023), the largest structured forecasting exercise on these risks: 89 superforecasters and 80 domain experts, four months of argument, and a tenfold disagreement at the end. We chose the cards that match where each group put its weight and when; the cards are our translation, not their words. The other presets are illustrations. None of them is the truth; they are there so you can see how much the answer depends on whose fears you start from.

06What it cannot tell you

• It is a rehearsal, not a forecast. Nobody knows the order in which these things arrive, or whether they arrive at all. The model shows what follows if, and which if matters most.
• Some risks have no geography. At the runaway end of the AI scale, no postcode helps: if networked systems act beyond human control, distance buys little. The AI channel scores exposure to the disruption on the way there, and distance from the chip and compute hubs a conflict over AI would target. It does not claim a refuge from the end state.
• Reaching a haven is not modelled. Distance, visas, closed borders and the cost of moving are not in this version. A haven you cannot reach is not one.
• The best havens attract people. A place that is safe on paper may be overwhelmed by those arriving. This version does not model inflow; the next one will.
• Some data is old. The Global Health Security Index is from 2021, the Global Food Security Index from 2022, tourism arrivals from 2019. Nuclear-winter food figures assume no trade at all, which later studies show overstates famine where trade survives.
• Averages hide places. A national score is an average over everyone in the country. Regions help for the largest countries; everywhere else, the coast and the hills get the same number.
• The weights are judgement. The loadings are anchored in the literature cited below, but they are choices, and they are all on this page so you can disagree with them specifically.
• Two cards borrow for now. A bot swarm that takes the internet down, and an Atlantic current collapse that chills north-west Europe, have their own mechanisms (networks going dark; regional cooling) that this version does not yet model. They borrow the nearest existing risks, so their effect on the map is approximate. Both get their own channels in the next version.
• Borders are the recognised ones. The map draws borders as recognised under international law, not as currently controlled: Crimea is Ukraine, the Golan Heights are Syria, Gaza and the West Bank are Palestine.
• Not advice. This is a way to think through trade-offs, not a relocation plan.

07Tell us what we got wrong

08Sources

• Xia, L., Robock, A. et al. (2022). Global food insecurity and famine from reduced crop, marine fishery and livestock production due to climate disruption from nuclear war soot injection. Nature Food 3, 586–596.
• Boyd, M. & Wilson, N. (2023). Island refuges for surviving nuclear winter and other abrupt sunlight-reducing catastrophes. Risk Analysis.
• Boyd, M. & Wilson, N. (2020). The prioritization of island nations as refuges from extreme pandemics. Risk Analysis 40(2).
• King, N. & Jones, A. (2021). An analysis of the potential for the formation of 'nodes of persisting complexity'. Sustainability 13(15), 8161.
• Forecasting Research Institute (2023). Existential Risk Persuasion Tournament.
• European Commission JRC. INFORM Risk Index. University of Notre Dame. ND-GAIN Country Index. Institute for Economics & Peace. Global Peace Index. Fund for Peace. Fragile States Index. NTI & Johns Hopkins. Global Health Security Index. WRI. Aqueduct. IFHV. WorldRiskIndex. World Bank. World Development Indicators and Worldwide Governance Indicators. IMF. AI Preparedness Index. Harvard Growth Lab. Economic Complexity. FAO. FAOSTAT food security indicators.
• Every input, its source, year and weight is in the data file behind this page.
• Search names for states, provinces and regions: GeoNames, first-level administrative divisions (CC BY 4.0).

The Havens Index is an illustrative model built from your own weighting of six broad risks and their timing, applied to published country data. It is not a forecast, not expert consensus and not advice. Channel scores are built from the indices cited above; regional adjustments and the loadings between risks and channels are documented judgements. Treat the ranking and the scenario as a way to think through trade-offs, not a prediction or a relocation plan.

Contexxt Editorial · first published September 2026, updated October 2026