How we score travel safety

Every IsMyTripSafe report answers one question — is this place safe to visit? — with a 0–100 score (100 = safest) built the same way for every destination. Nothing on a report page is hand-written per city; everything is computed from the sources below, so scores are comparable across destinations.

The composite index

20+ indicators are grouped into six hazard families: crime, conflict & terrorism, official guidance, institutions & rule of law, everyday hazards, and health & environment. Within a family the indicators are averaged — they are several noisy readings of one underlying thing, so averaging cancels noise. Across families they are not averaged.

That distinction is the whole design. Averaging across families is "compensatory": a good indicator cancels a bad one. For travel risk that is simply wrong — clean air does not offset an armed conflict. So families are combined on the risk scale through a generalised mean that weights severity, which means high risk in any one family dominates the result instead of being diluted by the calmer ones. Family weights: crime 45%, conflict 15%, official guidance 15%, institutions 15%, everyday hazards 6%, health 4% — renormalised over whichever families actually have data, and never raised above the hard caps below.

Raw values are normalised onto a 0–100 safety scale with piecewise bands calibrated per indicator (a homicide rate of 1/100k ≈ 94, 20/100k ≈ 26). Governance percentiles are used as-is. Counts are always per-capita, never absolute. When a primary source has no data, the pipeline falls back through independent secondary databases (UNODC via Our World in Data, WHO, V-Dem), then regional aggregates — each fallback labeled as such on the report. Every report also carries a confidence figure: the share of the weighting actually backed by measured data.

Hard caps

Some facts are categorical, not quantitative. "Do not travel" is not a data point to be weighed against air quality — it is a statement that bounds the answer. So a handful of conditions set a ceilingon the score rather than contributing a term to it: a pair of Level 4 ("Do Not Travel") advisories caps the score at 20, a Level 3 advisory average at 44, extreme homicide (50+/100k) at 28, sustained terrorism (25+ deaths per million) at 32, and bottom-5% political stability (active conflict) at 22. When a cap binds, the report says so and names the reason. No other indicator — and no traveller-type re-weighting — can lift a score above its cap.

Reporting reliability

Recorded-crime statistics are only as trustworthy as the institutions producing them. A country in the 40th percentile for rule of law reporting 1.2 homicides per 100k is not making the same claim as one in the 87th percentile reporting 0.5. So where institutions are weak, flattering crime figures are shrunk toward what the institutional data supports. The adjustment only ever lowers a crime score, never raises one — a country with strong institutions honestly reporting high crime keeps its high crime.

Missing data

Dropping a missing hazard family and renormalising the rest quietly assumes it matches the families we can see — which systematically flatters closed and unsurveyed countries, exactly the places least likely to publish crime statistics. Instead, a missing core family is filled with a conservative prior derived from the families that did resolve, marked as estimated on the report, and reflected in the confidence figure.

AI field research — and how it's bounded

Databases lag reality, so each report also includes AI-assisted field research: an agent searches current news and traveller reporting for recent incidents, unrest, district-level safety, scams and sentiment, with the database profile and official advisories supplied as fixed ground truth it must not contradict. Its street-crime ratings and traveller-sentiment read contribute 18%of the final published score; the databases keep 82%. Robbery and pickpocketing are scored on separate scales, because a threat to your person and a threat to your wallet are not the same risk. The field research is bound by the same caps as the composite — a positive sentiment read cannot argue away a "Do Not Travel" advisory. Advisory levels are never AI-generated — they are shown verbatim from the issuing governments' own feeds. We disclose this on every report because you should know which parts are measured and which are researched.

That research is published as a score of its own — the traveller-sentiment score, shown beside the crime score on every report. It combines how safe travellers report actually feeling day to day (35%) with current robbery (45%) and pickpocketing (20%) risk, on the same 0–100 scale as everything else. It is not a second opinion sitting next to the headline number: it is exactly the 18% term inside it, the part the databases cannot see.

The score pyramid

Each report leads with the composite, then the five scores behind it, arranged by how directly a traveller meets them. Crime and traveller sentiment sit on the upper tier — the two you feel walking down a street. Advisories, stability and health & air sit below as context. Every tile reads straight off the same engine as the headline, so a tile can never disagree with the number above it, and each one links to the evidence it came from.

Traveller-type adjustments

Per-traveller scores (solo, family, nightlife, 65+) are deterministic re-weightings of the already-computed category scores — extra weight on crime for a solo nightlife trip, on health & care for families and older travellers, capped so the general score stays the backbone and re-bounded by the same hard caps. Same inputs, same number, every time.

Who is in the party counts too. A women-only party puts a further 12% of the weight on the crime family, which is where the two indicators that track this most directly live — sexual-violence victimisation and the share of people who say they feel safe walking alone after dark; a mixed party adds 4%. The answers also change the written report: the findings pulled to the top are the ones that bear on that party, quoted from the same data shown further down the page. Nothing about the personalisation is AI-generated, and it can never lift a score above a hard cap.

Update cadence & dates

Live feeds (advisories, air quality, disaster alerts, weather) are fetched at report build time. A report's "last updated" date is the moment its data was actually rebuilt — we never bump dates without rebuilding the data behind them. Reports are rebuilt on demand once they're older than seven days.

Limitations

• Most statistical indicators are national; city-level nuance comes from Numbeo (where cities have enough contributors), OpenStreetMap, live environmental feeds and the field research.
• Survey-based indicators (victimisation, feeling safe at night) can be years old — every value shows its year on the report.
• Crowdsourced indices reflect perception, which can diverge from recorded crime.
• A score is a summary, not a guarantee. Conditions change fast; check your government's advisory before you travel.

Every source we use