In the PRC Data Challenge 2026, taxi-out time at ten European airports is the gap between two timestamps, and the off-block one is whatever the airport's database holds. Pick a real departure and watch where the stamp lands.
The aircraft moves off block when it moves. The airport's system stamps an off-block time, and the recorded taxi-out is the take-off time minus that stamp. Each case below is a real departure from 2025: choose one and watch the stamp land.
A recorded off-block within a minute of schedule is the signature of a copy, but punctual pushbacks land there too. The paper estimates that chance alone puts about 6 to 10% of records in that band. Nine airports sit at about that level. Fiumicino sits at twice it.
In this dataset, recorded taxi-out time is a record of what an airport's system wrote down, so the model has two jobs: measure the pushback where the record is real, and reconstruct the entry where it is not. The final model carries no value fitted to leaderboard scores and scores 233.7464 on January and July 2026. The competition model's 222.8507 includes such values and is not comparable.
MethodDepartures at ten European airports. Take-off times are given and only the off-block time is hidden, so the task is reconstruction rather than forecasting. Scored by RMSE in seconds. Validation is January and July 2025, scored out of fold; the 2026 figures are public leaderboard scores for January and July. All times on this page are UTC. ADS-B shows the first movement a receiver heard, which is not the same instant as brake release. Amsterdam routes are the receivers' own tracks; every other route is schematic; the times, stand and runway are from the record. Flight identifiers are shown as flown: the dataset's identifier field repeats letters of the airline designator on some rows (ARGG1141 for ARG1141, ITYTY680 for ITY680), which breaks the flight-plan match.
Our caveats: (1) the final model's 233.7464 comes from retraining once on four months, and February and June were scored without feedback; (2) two traces of the fitted values remain in it, the trajectory stage's search settings and the training target of one classifier, and the repository lists both; (3) the competition model was pushed with everything the leaderboard offered, so its 222.8507 is not a reference: removing most of its fitted constants costs 2.7501 RMSE-seconds, measured, and three more stay in both arms of that comparison; (4) 39.1% of scored departures have no surface track, all of Istanbul and nearly all of Paris Charles de Gaulle, a limit set by receiver coverage; (5) the mechanism behind the schedule-written and wrong-day records is not established, only the pattern in the data.
SourcePaper: Taxi-out Time as a Database Record. Code, build scripts and the value behind every constant: taxi-out-as-a-record, GPL v3.
DataDeparture records from the PRC Data Challenge 2026. Contains information from adsb.lol, made available under the ODbL 1.0, © adsb.lol contributors. Aircraft identities from OPDI.