Skip to content

About

How accurate are adsbdb and hexdb.io callsign-to-route lookups? 82 routes verified across 9 regions: US 27%, elsewhere 79% - and it is not database staleness.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

adsbdb / hexdb.io route accuracy — measured

DOI

Why is adsbdb returning the wrong origin and destination? Why does hexdb.io have no route for my callsign? Is the route data stale?

Short answer: the error is overwhelmingly regional, not stale. adsbdb is 27% correct on US domestic flights and 79% correct everywhere else. Record age barely matters — Australian routes verify at 100% on database rows with a median age of 11.5 years.

Update 2026-09-05: the US figure has been replicated at thirteen times the sample size — 28.6% on 395 flights, by an independent method. See Replication below.


How accurate are free callsign→route lookups, and where do they fail?

ADS-B broadcasts a callsign but no origin or destination. Integrations and apps that want to show "this aircraft is flying Denver to Chicago" resolve the callsign against a free database — usually adsbdb or hexdb.io. Both are widely used and both are frequently wrong.

This is a measurement of how wrong, and of where the error actually lives.

Result

It is not database staleness. It is which market the flight is in.

82 routes verified against an independent third source, across nine regions:

Region n hexdb correct adsbdb correct median hexdb record age
Oceania 6 100% 100% 11.5 y
Nordics 15 93% 87% 1.3 y
Europe 9 78% 67% 1.3 y
Canada 11 73% 73% 2.4 y
SE Asia 5 60% 60% 7.8 y
United States 30 13% 27% 7.5 y
Latin America 5 0% 80% 11.4 y

US 27%. Everywhere else 79%.

With 95% intervals: US 14–44% (n=30), everywhere else 66–88% (n=52). They do not overlap, so the split is not sampling noise. ⚠ The individual regional rows are a different matter — four of the seven have n≤6, and Oceania's 100% carries an interval of 61–100%. Those rows show where the sample went, not what a region's true rate is.

Why it is not age

The obvious hypothesis is that old records go stale. hexdb exposes an updatetime per route, so it is testable. Holding record age constant and varying region:

Records ≥ 4 years old n hexdb adsbdb
non-US 22 55% 82%
US 29 14% 28%

Same age band, threefold gap. And adsbdb exposes no record age at all, yet splits the same way — so the effect cannot be about database freshness.

Australia is the clearest single case: median record age 11.5 years, older than the US sample, and 100% correct on both sources. A 14.5-year-old record for QFA776 Perth–Melbourne is still right. A 14.9-year-old record for CPA135 Hong Kong–Melbourne is still right. Those city pairs have not changed.

The mechanism

Inferred, not measured. These databases key on flight number. Where a number means a fixed city pair for a decade — Australia, East Asia, the Nordics — every database is right at any age. Where numbers churn seasonally and are reused across regional operators, every database is wrong at any age. The US failures are heavy with regional carriers flying under mainline codes (SkyWest, Endeavor, Envoy, Republic) plus mainline domestic.

Replication: US 28.6% on 395 flights

The original US cell was n=30, sampled on a single day. It has since been re-measured against a larger sample by a different method, and the result holds.

Original Replication
date 2026-08-27 2026-09-05
n (US) 30 395
ground truth aviationstack filed FAA flight plans
adsbdb correct 27% 28.6%

Full breakdown of the replication:

n share
adsbdb returned a route 395 98.8%
exact match 113 28.6%
both ends wrong 241 61.0%
one end right 35 8.9%
reversed (origin/destination swapped) 6 1.5%

The pattern is high coverage and low accuracy. adsbdb answers almost every callsign and is wrong about two thirds of them, and the failures are not near misses. SWA938 returns KMDW-KBOS where the flight plan says KBWI-KDEN. DAL2768 returns KOAK-KSLC against KSEA-PANC. Those are different flights, not stale variants of the same one.

By operation type:

Category n correct
Commercial 364 29.1%
On-demand / air taxi 25 8.0%
Cargo 6 83.3% ⚠ n=6

On-demand charter is the worst-served category by a wide margin — those callsigns are not published schedules and no curated database has them.

Method

Filed flight plans carry the departure and arrival airport and, on most, the filed lateral trajectory. They are available free through the SWIM Cloud Distribution Service, which publishes the TFMS R14 Flight Data feed.

A warning before you subscribe. Free is not the same as unencumbered. Taking an FAA feed means signing a data access agreement, and that agreement governs what you may display and redistribute.

The FAA's Limiting Aircraft Data Displayed programme lets an aircraft owner ask to be excluded from displayed data. The list is republished weekly and subscribers are expected to apply the current one, including to data already collected — an aircraft that enrols on Tuesday has to disappear from what you show, not merely from what you gather next. A public artifact built from this feed therefore has to be rebuilt on that cadence rather than when it suits you.

How far that reaches is not something this repository can tell you. The obligation plainly covers the FAA's own data. Whether it also reaches aircraft you receive independently over ADS-B is a question about the terms you sign, and one worth putting to the programme office rather than inferring — the wording and the programme's stated scope do not obviously agree. Assume the stricter reading until someone with authority tells you otherwise.

None of this applies to ADS-B reception or to the other databases here. An aircraft's own broadcast is yours to receive, and adsbdb and hexdb impose no such terms. If you only want route data, use them. Subscribe only if you need what the filed plans give you, and read the agreement first — the obligations attach to you, not to your code.

  1. We collected flight plans continuously and aggregated by callsign.
  2. We kept only routes whose filed trajectory corroborates the filed endpoints — first and last trajectory point each within 25 km of the stated airports.
  3. We queried adsbdb for the same callsign and compared.

Step 2 is the arbiter. Measured across 4,000 routes, the filed trajectory agrees with the filed endpoints 99.88% of the time, which is what makes it usable as a check.

⚠ What this does not establish

The trajectory is not independent ground truth. It establishes that the flight plan is internally consistent — the filed path agrees with the filed endpoints. It cannot detect a plan filed correctly and then flown somewhere else, and where no flight plan exists adsbdb cannot be judged at all. That is weaker evidence than the aviationstack verification used in the original study, and it is why this is presented as a replication of one cell rather than a replacement for the whole table.

It is a US number. The feed covers US airspace, so this sample is US- heavy by construction. 28.6% is comparable to the 27% US figure above, and not to the 79% measured everywhere else. Quoting it as a global accuracy rate would misrepresent both measurements.

Four routes were excluded. All four Cathay Pacific flights through Hong Kong showed an identical 29 km offset between the trajectory endpoint and the airport — a systematic error in one airport record rather than four bad routes, but it would distort the arbiter.

Three distinct failure modes

  1. Records exist and are wrong — the US. No source-switching helps; both sources fail on the same flights.
  2. Records exist and are right — Australia, East Asia, the Nordics, regardless of record age.
  3. Records are absent — Timor-Leste (0 of 12 in hexdb), South Asia (16 of 22 callsigns with no hexdb record). Not wrong, missing. Nothing to fall back to.

Latin America is its own oddity: hexdb 0/5, adsbdb 4/5, Brazilian domestic.

What actually helps

A geometric plausibility check — reject a route the aircraft's own position contradicts. No API key, no account, no regional assumption. Measured against verified truth:

adsbdb answers n actually correct
passed the check 25 72%
rejected by the check 18 11%

Roughly seven-to-one separation. It discards a correct route about 11% of the time it fires — the price of turning a confident wrong answer into a blank.

Corrected 2026-09-07. This table first read 8 / 22 at 75% and 9%. Those figures came from a flag stored during collection, which was only computed for one 60-record cohort rather than for every verified row. Recomputing the check over all 43 verified rows that carry coordinates gives the numbers above. The separation and the conclusion are unchanged; the sample is larger and the population is now stated.

Restricted to US flights alone, where the underlying route data is worst, the check separates 7 kept at 71% from 16 rejected at 12%.

Tested on data it was never tuned against

The figures above have an obvious weakness: the check was designed and scored on the same 82 flights. A rule evaluated on the data that shaped it flatters itself.

So it was re-scored against the 395 replication flights, which it had never seen, using filed flight plans as the arbiter:

passed the check rejected separation
designed-against, n=43 72% 11% 6.5 : 1
unseen, n=395 59% 7% 8.4 : 1

95% intervals on the unseen data are 51–66% and 4–11%, which do not approach each other. The separation is not an artefact of tuning.

The absolute rates differ for two reasons worth stating. The arbiter is weaker — agreement with a filed plan rather than with a verified flight. And the aircraft position used is the midpoint of the filed trajectory, which is by construction far from both airports, so the rule that accepts anything within 150 km of either end almost never fires. That biases the unseen test toward rejection: 58% of flights were rejected there against 42% in the original.

See route_audit.py, which is standalone and runnable against any location.

Limitations, stated plainly

  • ⚠ The sample is stratified, not random. Flights were drawn in cohorts chosen to test specific ideas — a 30-flight random pull, then deliberate cohorts for Canada, Sweden, Norway, Denmark, Oceania, Latin America, East and Southeast Asia, plus old_eu and young cohorts selected on record age. So "79% everywhere else" is an average over regions that were picked, not an estimate of global accuracy. The US–non-US contrast is the finding; the aggregate is not a population figure.
  • One day for the regional table — all sampling on 2026-08-27. The US replication was sampled separately on 2026-09-05.
  • One ground-truth source for the regional table (aviationstack). Its own coverage may vary by region; that confound cannot be ruled out. A check confirmed mismatches were against every leg a flight number flew that day, not just the chosen one. The replication used a different source — filed FAA flight plans — which is why it is evidence rather than repetition.
  • Small cells. n=5–6 for several regions. ⚠ The US cell was n=30 and is now separately supported at n=395; the non-US cells have not been replicated and remain small. The 79% figure rests on 52 flights.
  • ⚠ The two measurements are not the same test. aviationstack verifies against what was flown; the replication verifies against what was filed. They agree here, which is informative, but they are not interchangeable.
  • A rate is not a quality score. Discard rates below are counts, not verdicts on whether discarding was right.
  • Per-flight records are not published — see below. Nobody can check the arithmetic; they can only re-run the method. That is a real weakness and the honest cost of the licensing position.

Why the raw data is not here

Ground truth came from aviationstack under a personal-tier licence, and adsbdb's route data carries a restriction on copying or republication. A table of per-flight answers would redistribute two restricted datasets at once.

Published here: the method, the scripts, and aggregate statistics — which are measurements over that data rather than the data itself. Withheld: raw API responses and per-flight rows.

Reproducing

The regional table:

python study_collect.py  <lat> <lon> <radius_nm> <max>   # free
python study_verify.py   --spend N                        # uses your quota
python study_crosstab.py                                  # free

The US replication, which spends no quota but needs your own SWIM/TFMS subscription and a route table built from it — see the warning above before you go and get one:

python compare_against_filed_plans.py --db routes.db --airports airports.csv

study_verify.py needs an aviationstack access key. It refuses to run without an explicit --spend N so quota cannot be burned by accident, and it caches every response, because a verification spent is gone for the month.

⚠ Point the collector at a nearby airport, not your own coordinates. Your sampling centre ends up in your results.

Citing this

⚠ Cite the concept DOI, which always resolves to the most recent version:

Strausberg, Lee. adsbdb / hexdb.io route accuracy: measured.
Zenodo. https://doi.org/10.5281/zenodo.22652002

To cite this specific release rather than the latest, use 10.5281/zenodo.22652003.

A CITATION.cff is included, so GitHub's "Cite this repository" panel and most reference managers will read the metadata directly.

Licence

MIT for the code. The findings are measurements, freely usable. The underlying route and flight-status data belongs to its respective providers.

About

How accurate are adsbdb and hexdb.io callsign-to-route lookups? 82 routes verified across 9 regions: US 27%, elsewhere 79% - and it is not database staleness.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages