Methodology
How We Score Aircraft Safety
How SafetyScorer rates 70 airliner families from 0 to 100, using fatal passenger accidents from ASN and flight counts from OpenSky and the World Bank.
Updated
Our aircraft score answers one question: if you board a random passenger flight on this type, how likely is it that you will not survive? Modern jets such as the Airbus A350, the A320 family, the Boeing 777 and the 737 NG score in the 90s, which means roughly one death-equivalent in 20 to 50 million flights. Old Soviet types and small bush turboprops score in the 50s and 60s. The airline and the place you fly matter more than the type, so read the score as one input, not a verdict.
What counts as a fatal accident?
We start from the Aviation Safety Network (ASN) database: 27,463 records up to 7 October 2026. We keep every event on one of 70 airliner families in which at least one person on board died. That gives 1,891 events. We drop 12 of them as unrelated to the aircraft (see below), and 1,070 of the rest count toward the score. An event counts when:
- the flight carried passengers (scheduled, charter or air taxi), and
- it was an accident, or its cause was never established. MH370, China Eastern 5735 and the still-open Air India 171 investigation count.
We leave out 656 events on cargo, ferry, training, test, military and other non-passenger flights, because our flight counts cover passenger flights only. We also leave out 148 acts of unlawful interference, such as bombs, hijackings and shoot-downs (Pan Am 103, MH17, Ukraine International 752), and 5 crashes that an official investigation found a pilot caused on purpose (Germanwings 9525, LAM 470, EgyptAir 990, Royal Air Maroc 630 and Japan Air Lines 350). These say little about the aircraft. Excluded events still appear in each aircraft's history.
Deaths not related to the aircraft or its operation are neither counted nor listed: falls while boarding or leaving a parked aircraft, ground staff hurt at the stand, stowaways and natural deaths. This is in line with the accident definition in ICAO Annex 13. The 12 we dropped are eight falls at the stand, one stowaway found in a wheel well, a passenger who opened an exit in flight, and two deaths on parked or stolen aircraft caused by people with no link to the flight (self-inflicted or by others, also outside the Annex 13 definition). Turbulence, system faults and injuries during an evacuation do count, because they are part of the flight.
We checked every passenger event of the types still flying by hand and corrected ASN's labels where needed: Pan Am 103, for example, was filed as an accident. We also read the ASN narrative of every counted event in which the aircraft was undamaged or only slightly damaged.
How do we measure severity?
A crash that kills everyone is worse than an overrun that kills one passenger, so we use the Full-loss equivalentA way to weigh a fatal accident by the share of people on board who died. A crash that kills everyone counts as 1.More in the glossary (FLE) developed by MIT statistician Arnold Barnett: each accident adds the share of people on board who died. Air France 447 adds 1.0. Asiana 214 at San Francisco, with 3 deaths among 307 people, adds 0.01. Total FLE divided by the number of flights is the chance that a random passenger on a random flight is killed.
How many flights has each type flown?
No public source counts flights by aircraft type over 60 years, so we estimate them:
- World Bank indicator IS.AIR.DPRT (from ICAO): departures by registered carriers for the world, its regions and 12 large countries, 1970 to 2023.
- OpenSky Network flight lists for 2019 to 2022 (117 million flights, on Zenodo). We count flights by passenger airlines for each aircraft type and re-weight them country by country to the World Bank totals. This corrects most of the bias of OpenSky's volunteer receivers toward Europe and North America.
- A fleet model for other years: deliveries, a retirement curve and flights per aircraft per year, calibrated on the OpenSky years. Types OpenSky barely sees use averages of similar aircraft, except Concorde, which flew far less than other airliners and gets its own estimate of about 200 flights per aircraft per year from its published flying hours; Soviet types after 1991 are capped by the departures the former Soviet republics reported.
In total that is about 1.2 billion flights, from 236 million for the A320 family to 2.8 million for the A350.
The formula
rate = (Full-loss equivalentA way to weigh a fatal accident by the share of people on board who died. A crash that kills everyone counts as 1.More in the glossary + k × r) / (D + k)
score = 100 − 25 × log10(1 + rate / 0.05)
D is flights in millions, r is the FLE rate per million flights of the aircraft's technology generation, and k is 3 million flights.
The score is on a log scale, so each tenfold rise in risk costs about 25 points: one death-equivalent in 20 million flights scores 92, one in 10 million 88, one in a million 67 and one in 100,000 42. Grades: A from 85, B from 70, C from 55, D from 40, F below 40.
Why do new aircraft start near their generation's average?
An aircraft with 2 million flights and no fatal accident has not proven that it is safer than one with 200 million flights and eight accidents. So we borrow strength from similar aircraft (Empirical BayesA statistical method that pulls a small or new record toward the average of similar cases, so a short lucky streak does not look like proof of safety.More in the glossary), grouped by Airbus' classification of airliner technology:
| Generation | Examples | Full-loss equivalentA way to weigh a fatal accident by the share of people on board who died. A crash that kills everyone counts as 1.More in the glossary per million flights |
|---|---|---|
| Gen 4 jets: Fly-by-wireA flight control system in which the pilot's inputs go to computers, which then move the control surfaces. Many such systems also stop unsafe manoeuvres.More in the glossary with envelope protection | A320 family, A350, 777, 787, E-Jet | 0.041 |
| Gen 3 jets: glass cockpit and flight management system | 737 NG and MAX, 757, 767, CRJ | 0.17 |
| Gen 2 jets: integrated autopilot, analogue cockpit | 727, DC-9, 737-200, Tu-154 | 0.88 |
| Gen 1 jets | 707, DC-8 | 0.71 |
| Modern turboprops | ATR, Dash 8, Saab 340 | 0.92 |
| Classic turboprops | Twin Otter, An-24, Let 410, F27, Islander, Metro | 2.9 |
A type's own record and its generation's average weigh the same at 3 million flights. We took k from the data: a fit of how much families in one generation really differ, beyond chance, gives 1.5 to 4.4 million. Halving or doubling k moves most scores by 0 or 1 point and none by more than 4.
Worked example: Boeing 737 MAX
Two counted accidents, Lion Air 610 (189 of 189 died) and Ethiopian 302 (157 of 157), give Full-loss equivalentA way to weigh a fatal accident by the share of people on board who died. A crash that kills everyone counts as 1.More in the glossary = 2. The MAX has flown about 10.5 million passenger flights; the Gen 3 average is 0.169.
- Rate = (2 + 3 × 0.169) / (10.5 + 3) = 0.19 per million flights, about 1 in 5.4 million.
- Score = 100 − 25 × log10(1 + 0.19 / 0.05) = 83, grade B.
The 737 NG, with 7.2 FLE over 165 million flights, scores 93.
How sure are we?
Confidence is high from 20 million flights, medium from 3 million and low below that. Low-confidence scores, such as those of the A350, A380, A220, C919, Superjet and An-148, mostly reflect their generation.
Why do some aircraft share a rank?
Aircraft types whose scores are within 1 point of each other share a rank, shown as "1=". A gap that small is not a real difference: our flight estimates can be off by about 25%, and the choice of k alone moves most scores by 0 or 1 point. We sort the types by unrounded score and form groups: a group starts with the highest remaining score, and every type less than 1 point below that first score joins it. All types in a group share a rank, numbered as in sports tables: the rank is one plus the number of types in the groups above, so five types sharing rank 1 are followed by rank 6. With scores of 95.0, 94.5 and 94.0, the first two share rank 1 and the third ranks 3. Lists still run from lowest to highest risk, and retired types are not ranked.
Where do the event descriptions come from?
Every event in an aircraft's history has a short description that we wrote in our own words from the ASN narrative, and an event type based on the ICAO occurrence categories (CICTT), such as controlled flight into terrain, loss of control in flight or Runway excursionWhen an aircraft leaves the runway surface during take-off or landing, either off the end (an overrun) or off the side (a veer-off).More in the glossary. Where the cause is not known, the type says so. Each event links to its Aviation Safety Network record, which has the full details. ASN is the source of all the accident data on our aircraft pages.
What are the limits of the data?
- OpenSky coverage. Receivers cluster near big airports, and many older turboprops had no ADS-B transponder before the 2020 mandates. Turboprops in remote regions are under-counted, so they may look slightly worse than they are.
- The fleet model. Years before 2019 rest on approximate delivery and retirement figures: about ±25% for types OpenSky covers, ±50% for the others. A 25% error moves a score by about 2 points.
- Soviet types before 1992. With no Soviet traffic data, we assume Western utilisation, which probably flatters their score.
- Air taxi. ASN counts air taxi flights as passenger flights, but our flight counts come mostly from airlines, so the Caravan, Twin Otter and Islander rates are likely too high.
- Missing occupant counts. For 36 events ASN gives no occupant count; we assume 75% of the typical seats.
- Cut-off. The data end on 7 October 2026. Later accidents are not yet included.
Before you book, check who operates the aircraft and where it flies. A type scoring in the 80s or 90s with a well-run airline is a very small risk; for a type in the 50s or 60s, the operator's record matters most.