How We Calculate Safety Scores

Overview

We aggregate data from 30+ trusted public sources to compute a composite safety score (1-10) for every country. Higher scores indicate safer destinations. Our scoring engine uses an uncertainty-weighted (Bayesian shrinkage) weighted geometric mean across five categories — Conflict (30%), Crime (25%), Health (20%), Governance (15%), and Natural hazards (10%) — so a single critical weakness (e.g. active war) cannot be masked by strong scores elsewhere. Countries with sparse or stale data are shrunk toward a conservative, region-anchored prior instead of being scored as if the missing evidence were average, and a smooth, down-only acute-risk term gives extra weight to severely low Conflict or Crime pillars. A calibrated, importance-weighted consensus of government travel advisories feeds into the Conflict pillar. There are no hard score caps or critical floors — every score responds continuously and predictably to the underlying evidence.

Formula v10.1: indicators, sources and editions

Each pillar combines the indicators below. We always use the latest edition of each source, and every country page shows each indicator with its own value, unit and data year.

Rules that decide what counts

• Latest edition only

  When a source publishes a new edition, the previous one weighs zero: editions are never blended, and old values are never faded in.

• Five-year maximum age

  An indicator whose data year is more than five years before the current year counts as missing, so it can neither raise nor lower a score. Government advisories are current by definition; population figures, used only to judge how reliable a homicide rate is, are exempt.

• Scale continuity

  Changing the indicators inside a pillar changes the spread of its values, and with it the pillar's real weight. The health, governance and natural-hazards pillars are therefore mapped back onto their previous scale with a fixed linear map, computed once on 14 October 2026 and never refitted.

• Partial data: no score below four sources or without crime data

  A place covered by fewer than four independent sources gets no official score. A place also gets no score when none of its sources measures crime (neither the Global Peace Index nor a World Bank homicide rate): without one, the crime pillar is a cautious estimate alone, and tests on the countries with complete data showed that estimate to run about half a point too low. Disaster alerts (GDACS, ReliefWeb) do not count toward the four, because they cover a place only while a disaster is under way. Nor does a source that supplies no value used in the score, such as a World Bank entry with only a population figure. Its page says “Partial data”, lists the sources we do have and asks travellers who have been there how safe they felt. Danger always overrides this: a place is scored when at least three governments advise against all travel, when do-not-travel advisories dominate, or when the Uppsala Conflict Data Program recorded conflict deaths there in the latest year.

• Community votes

  Travellers' votes count once a country has at least 30 votes from at least three countries of origin in the last 180 days, and no single country of origin may carry more than half of the weight. The adjustment is the gap between the community rating and the data score, scaled down when there are few effective votes, and capped at ±1.0 points.

• How often things change

  The score is recalculated every day. Government advisories and disaster alerts change daily; the international indices (Global Peace Index, INFORM, V-Dem, World Bank, UCDP) are published once or twice a year, so most pillar values stay the same between two editions.

Changelog

October 19, 2026: a crime measure is now required

• A place now needs at least one measure of crime (the Global Peace Index or a World Bank homicide rate) for a score, on top of the four-source minimum. Removing those two sources from the countries with complete data lowered their scores by half a point on average, as much as for the places with three sources that already showed partial data. The Comoros, Cabo Verde, the Maldives, the Solomon Islands, São Tomé and Príncipe and Vanuatu now show partial data; no other score changed.

October 17, 2026: formula v10.1

• Government advice to reconsider travel (level 3) now lowers a score too, at half the weight of advice against all travel (level 4). Before, only level 4 counted, so places where most governments advise reconsidering travel, such as the Gulf states and Cuba, kept a moderate or good score.
• Natural hazards: the exposure is now combined with the country’s capacity to cope (INFORM’s lack of coping capacity), the way INFORM itself rates risk. A well-prepared country such as Japan no longer ranks below a less exposed but less prepared one.
• Armed conflict: the conflict-death scale no longer stops at 24,000 deaths a year, so a full-scale war (Sudan, Ukraine) ranks below a smaller one.
• Missing data: when a pillar has little or no data, its cautious estimate can now rise as far as unanimous level-1 advisories justify, instead of stopping at 6 out of 10 (Liechtenstein, for example).
• A source counts toward the four-source minimum, and is listed on a page, only when it supplies a value used in the score. Nine places whose fourth source was a World Bank population figure (Brunei, Micronesia, Kiribati, the Marshall Islands, Nauru, Palau, Tonga, Tuvalu and Samoa) now show partial data.
• Each change was tested alone and together against verified consistency checks, travellers’ votes and an independent expert reference set; candidates that failed, such as more weight on homicide or on the crime pillar, were not adopted.

October 15, 2026: formula v10

• Removed indicators a traveller cannot act on: child mortality, PM2.5 air pollution, and a flood-risk column that was labelled as climate risk. The health pillar now reads INFORM’s access to health care instead of a column that actually measured conflict hazard.
• INFORM is read from its newest edition (INFORM Risk 2026) with exact column names; the governance indicator now reads INFORM’s Governance column instead of the Institutional one.
• Latest edition only and a five-year maximum age: stale values, such as homicide rates from before 2021, now count as missing instead of being shown as current.
• Places with fewer than four independent sources no longer get a score built from regional estimates: they are shown as partial data.
• Community votes can move a score by up to one point, under the rules above.
• Pillar weights and the conflict and crime pillars are unchanged; the environment pillar is now called natural hazards. Scores published before this date used formula v9.1, and trend charts mark the switch.

Category Weights

Categories are weighted to reflect their relative importance to traveler safety. Conflict carries the highest weight (30%) because active conflicts pose the most immediate and severe risk. Crime contributes 25%, as it directly affects personal safety day-to-day. Health (20%) remains important but is often manageable with precautions. Governance (15%) captures institutional stability as an indirect factor. Natural hazards (10%) reflects exposure to floods, storms, earthquakes and other natural hazards that, while relevant, tend to be more localized or seasonal.

Understanding Each Category

Click on each category below to learn what it measures and how to interpret scores.

Data Sources

Government Travel Advisories

We collect travel advisory levels from 25+ governments across all continents. Each advisory is normalized to a unified 1-4 scale (1=safe, 4=do not travel) and weighted by tier in the Conflict pillar.

Scoring Formula

Raw indicator values are normalized to a 0-1 scale (higher = safer) and combined into a raw pillar value p using each indicator's sub-weight. Each pillar is then shrunk toward a conservative, region-anchored prior μ using p̂ = (n·p + K·μ) / (n + K), where n is the pillar's precision — how much fresh, present data backs it — and K is a fixed shrinkage constant: a well-measured pillar sits close to its raw data, while a thin or stale pillar is pulled toward the cautious regional baseline instead of guessing. The five shrunk pillars are then combined with a weighted geometric mean — unlike an arithmetic mean, this ensures a single dangerously low pillar drags the overall score down more than a strong one can lift it. Two continuous refinements apply, with no thresholds or hard cutoffs: a down-only acute term gives Conflict and Crime extra weight when either is severely low, and a count-damped severe-advisory modifier gently discounts the score when government advisories broadly agree on "Do Not Travel" or "Reconsider travel" (a level-3 "reconsider" counts half as much as a level-4 "do not travel") — a single thin or contested record carries far less weight than a broad multi-government consensus.

Score Bands

Scores fall into five bands: below 5.0 is Danger, 5.0-6.0 High caution, 6.0-7.0 Moderate, 7.0-8.0 Good, and 8.0 and above Excellent. Because the formula shrinks uncertain estimates toward a cautious prior, scores rarely sit at either extreme: among the countries with an official score, the range runs roughly 3.4 to 8.9, with an average around 6.8. A country scoring 7.1 and one scoring 8.8 are both in a positive band, but they are not equally safe; always read the exact score, not just its band.

What Changed in Formula v9.1, and Why

Formula v9.1 builds on v9's Bayesian-shrinkage foundation with four additions. Crime now blends the Global Peace Index's Safety & Security domain with the World Bank's intentional-homicide rate, whose statistical precision is scaled by population so a handful of homicides in a very small territory is treated as noise rather than a worse-than-war crime signal. Conflict now also incorporates conflict-death counts from the Uppsala Conflict Data Program (UCDP Georeferenced Event Dataset), distributed via the Our World in Data mirror under a CC-BY license and log-normalized so a single catastrophic year does not saturate the signal. The Global Peace Index has been upgraded to its current 2026 edition, and — as is standard IEP practice — prior-year vintages are subject to retroactive revision. Finally, territories with very few tracked advisories now receive a smooth advisory-consensus nudge that ramps in gradually rather than switching on or off at a single threshold. Together these changes widen the practical score range to roughly 3.4-8.9, with a global mean around 6.60; the band thresholds themselves are unchanged.

What Changed in Formula v9, and Why

Earlier versions of our formula had three specific weaknesses. The Crime pillar was actually measuring rule of law rather than criminal safety. A rigid score cap and a critical floor could make the score jump discontinuously around a single threshold. And countries with very little data were blended toward an advisory-derived estimate that could look more confident than the underlying evidence justified. Formula v9 fixes all three: Crime now uses the Global Peace Index's Safety & Security domain directly, with rule of law moved to Governance where it belongs; every score now responds smoothly and continuously to new evidence, using a calibrated advisory consensus and a down-only acute-risk term instead of hard thresholds; and countries with sparse or stale data are shrunk toward a conservative, region-anchored statistical prior (Bayesian shrinkage) rather than blended toward an ad hoc estimate. The result is a score that moves gradually and predictably as new information arrives, alongside a confidence figure that shows how much evidence actually backs it.

Confidence

Alongside each score, our data carries a confidence value from 0 to 1 — a precision-weighted measure of how much fresh, present data backs it across all five pillars. High confidence means a country has rich, recent coverage from most or all pillars; low confidence means the country has thin data and its score sits closer to our conservative regional prior than to hard local evidence. Confidence is a transparency signal for every country, not just small or data-sparse ones.

Known Limitations of Formula v9.1

A few edge cases are worth knowing about. Palestine (around 4.6) is currently rated slightly above Iran (around 4.2) — this is evidence-based, not an oversight: Iran carries a near-universal multi-government "Do Not Travel" advisory load and weaker governance and environment scores, while Palestine's Gaza conflict-death toll already fully sinks its Conflict pillar. Homicide rates for very small-population territories are statistically reliability-damped: a handful of homicides on a population of only tens of thousands is treated as noise rather than a worse-than-war crime signal, since a per-capita rate on so few people is not reliable at full statistical precision. UCDP conflict-death data is published quarterly and lags real-world events by roughly three months, and the most recent 2025-2026 figures should be treated as preliminary. And Japan's MOFA travel advisories are now displayed correctly on country pages (a fetcher bug fixed in this release), but remain excluded from the score itself pending a dedicated recalibration of their historical severity mapping.

Baseline + Signal Architecture

Baseline sources (World Bank, V-Dem, INFORM, GPI) are updated annually and anchor each pillar with high-precision evidence. Signal sources — real-time government travel advisories from 25+ governments, ReliefWeb, and GDACS — update daily or weekly: advisories combine into a single calibrated, importance-weighted consensus that feeds the Conflict pillar, while ReliefWeb and GDACS active-disaster alerts inform the Natural hazards pillar. Every indicator's statistical precision is scaled by a freshness-decay multiplier, so stale data — however recent it once was — never carries as much weight in the final score as current evidence.

Data Freshness Decay

Each data source has a configured half-life for freshness decay. If a source's data is one half-life old, its statistical precision is roughly halved; at two half-lives, roughly a quarter; and so on. Beyond a source's configured maximum age, its precision drops to zero, and the affected pillar shrinks further toward the conservative regional baseline instead of relying on stale evidence. This ensures scores always reflect the most current available information while gracefully degrading — never abruptly — when sources are temporarily unavailable.

How INFORM Enriches the Scoring

The INFORM Risk Index, published twice a year by the European Commission’s Joint Research Centre, feeds four indicators: exposure to natural hazards, epidemic risk, access to health care and governance. A fifth column, lack of coping capacity, is not scored on its own: it weighs the natural-hazard exposure, because a well-prepared country runs a smaller risk from the same hazard. We read them by their exact column names from the newest edition (INFORM Risk 2026) and use only that edition.

Global Peace Index

The Global Peace Index (GPI) from the Institute for Economics and Peace is integrated into the Conflict pillar. GPI indicators (overall score, safety & security, militarization) provide a comprehensive picture of peacefulness based on 23 qualitative and quantitative indicators across 163 countries.

Community Sentiment

Below each country's data-driven pillars, travelers can tell us whether the safety score feels right, on a five-level scale from “way too high” to “way too low”. Where independent data is too thin for an official score, they rate how safe they felt instead, from “very unsafe” to “very safe”. Every vote becomes a perceived score: the score shown when the vote was cast, moved half a point per level (a direct rating counts as 2, 4, 6, 8 or 10). We average the votes of the last 180 days, giving newer votes more weight (it halves every 30 days), and no single country of origin can carry more than half of the total weight. The community only counts once a country has at least 30 votes from at least 3 countries of origin. From then on it can move the safety score by up to ±1.0 point, and the fewer the votes, the smaller the effect: it refines the data but never overrides it. Until a country gets there, its page only says that there are not enough votes yet; it never shows how many votes a country has. Of each voter we keep only the country of origin (country level, never a raw IP address).