Why your flight is late

On time has a definition, and the airline writes half of it. A flight is on time when it arrives less than fifteen minutes after its scheduled arrival, the A14 convention the Bureau of Transportation Statistics has used since 1987, and the scheduled arrival is whatever the carrier chose to publish. From 2015 to 2025, across 72,583,211 flown flights, the share arriving on time went from 81.4% to 77.7%, a change of -3.7 pts (interval -4.9 pts to -2.6 pts).

Now the flying. On the same 3,631 routes, flown in every full year of the window and carrying 88.7% of its flights, each route weighted the same in every year, the time a flight actually spent between the gates moved +4.4 min (interval +4.0 min to +4.8 min).

The time the schedule allowed on those routes moved +4.6 min (interval +4.3 min to +5.0 min). The schedule's change less the flying's is +0.2 min (interval 0.0 min to +0.5 min). The schedule kept pace with the flying and no more, so the timetable did not move the on time rate either way: what moved it is delay the block times do not hold, the late departures that the chapter on inherited delay follows from one leg to the next. The rate is a comparison between what happened and what the airline said would happen, and only one of the two is outside the airline's control.

Padding is the scheduled block time minus the time the trip takes when nothing is in the way, measured as the gate to gate time that only 10% of flights beat on the same route, in the same three hour departure block and season, over 2015 to 2022. In 2025 the average flight carried 21.2 min of it, 14.4% of its scheduled block, and across the year that came to 2,362,990 h of scheduled time. On the same routes, carriers, hours and months flown in both 2015 and 2025, padding moved +3.98 min (interval +3.84 min to +4.13 min) while the actual block time on those cells moved +4.16 min (interval +3.77 min to +4.58 min). Padding's change less the flying's is -0.18 min (interval -0.60 min to +0.24 min), an interval that holds zero: the schedule grew by what the flying lost, no more and no less.

The growth did not arrive evenly. A changepoint search (PELT) over each of the 20 carriers' route adjusted, deseasonalized monthly series looked for moves of a minute or more that stayed, and found 204: 131 up and 73 down. The largest was JetBlue Airways in May 2020, a step of -9.3 min. Each is dated in the table under the chart, which is where a scheduler would look for the timetable change that caused it.

If carriers managed to the line, arrival delays would pile up just under fifteen minutes and thin out just over it. The chart counts every flown flight in the reporting years by whole minute of arrival delay, 24,231,039 flights between an hour early and three hours late.

At the line, the pooled distribution holds -3.84% more flights just under fifteen minutes than a smooth curve fitted around it predicts. That number means nothing on its own: every histogram has bumps.

So the same statistic was computed at 49 placebo thresholds between ten and sixty five minutes, where nobody has a reason to manage, to give its natural spread. The line's statistic sits -2.40 placebo standard deviations from their mean. Carrier by carrier, corrected together by Benjamini-Hochberg at q = 0.05, 0 of 15 carriers show bunching. That is a finding, not a failed test: in the reported arrivals, the line is not visibly gamed.

A delay reported on a flight is not a delay caused on that flight. To separate the two, every flown leg with a tail number was put back in order on its aircraft's day: 72,583,211 legs, 18,151,842 rotations, 10,019 aircraft. A leg is linked to the one before it only when it leaves from the airport that one landed at, after it landed, within 6 h. 369,126 pairs broke the chain, most of them aircraft swaps the tail number does not show, and were split rather than joined; 601,945 were impossible sequences, quarantined and counted. On the 18,381,035 linked legs of the reporting years, each minute the inbound aircraft arrived beyond the scheduled turn, less a minimum turn of 45 min, put 1.068 minutes onto the next departure (interval 1.067 to 1.070). That makes 34.9% of all arrival delay minutes inherited from an earlier flight (interval 34.8% to 34.9%), against 37.4% in the late aircraft field the airlines report.

The chart shows where a buffer would have stopped it. The longer the scheduled turn, the less of a late arrival reaches the next departure, and beyond the minimum turn 30 min of extra buffer halves the minutes passed on. The aircraft's day explorer on the rotations page shows the worst day of each month leg by leg.

The usual airline ranking sorts carriers by average arrival delay. That is a fair measure of what a passenger on each carrier sat through and an unfair measure of the carrier, because carriers do not fly the same routes, hours, months or aircraft. On the raw average in the reporting years, the three least delayed carriers were Endeavor Air, Republic Airways, Alaska Airlines, and the three most delayed were JetBlue Airways, American Airlines, Frontier Airlines.

Holding constant what each carrier flies (route, month, scheduled departure hour and aircraft type, fit by weighted least squares on 3,025,819 cells, with intervals clustered by day), the three least delayed become Endeavor Air, Republic Airways, Envoy Air, and the three most delayed Allegiant Air, Frontier Airlines, Hawaiian Airlines. 11 of 15 carriers change places.

PSA Airlines rises the most, from 11 to 7: its routes, hours and months are harder than the average carrier's, and its raw average charges it for them. Hawaiian Airlines falls the most, from 8 to 15: its flying is easier than most, and its raw average credits it for that.

Every flight that arrives fifteen or more minutes late carries a breakdown of its delay into five reported causes. Over the window, the carriers put 38.9% of the cause minutes on a late aircraft, 34.2% on the carrier, 21.0% on the national airspace system, 5.7% on weather and 0.18% on security. These fields record what was reported, by the carrier, under rules that keep the weather field for extreme weather; ordinary weather that slows the airspace is coded to the national airspace system.

The national airspace system's share is highest at EWR, SFO, LGA, airports where capacity, not the weather and not the airline, sets the pace much of the day.

Measured from the weather itself (precipitation, snow, gusts, low cloud, thunder, fog and freezing precipitation at the origin when the flight leaves and at the destination when it lands, on 8,320,779 flights between two of the FAA's Core 30 airports, with the inbound aircraft's lateness held constant), weather explains 18.8% of arrival delay minutes (interval 17.7% to 19.9%). The weather field on the same flights records 4.2%, and the national airspace system field 22.3%.

A meltdown shows in the daily numbers before it shows in the news. The detector scores every carrier's day, and every day at the 30 busiest airports, against that carrier's or airport's own previous four weeks, on the share of flights cancelled and the mean arrival delay. Its alert threshold was chosen on 2015 to 2022 by cost, with a missed disruption costed at 50.0 false alarms, and landed at 29.0 robust standard deviations, inside the grid. On the reporting years it flagged 8 of the 8 disruptions in the known events table, at a cost of 5.12 false alarm episodes per thousand carrier and airport days. The chart opens on Southwest Airlines in 2022-12-21: its cancellation rate against every other carrier at its busiest airports (DEN, LAS, MDW, DAL, BWI, PHX, HOU, BNA, OAK, SAN), with the disruption shaded. The storm hit everyone; the gap is the carrier's own.

Delta Air Lines in 2024-07-19 is the second panel: an outage of the software that runs its crew and aircraft scheduling, against the same kind of peer control. Over its window it cancelled 2,842 more flights than its peers' rate implies, and its passengers sat through 8,329 h of delay beyond theirs. Southwest Airlines's excess over its window was 5,614 flights and 2,531 h of delay.

Recovery is the number that separates them: Southwest Airlines took 9 days from onset to get back inside its own normal gap to its peers for three days running, and Delta Air Lines took 5. Across every alert episode in the window, carriers differ in how long their episodes last (Kruskal-Wallis H 23.11, permutation p 0.018, 15 carriers): how fast an airline recovers is a property of the airline, not only of the storm.

For the traveler, the safest flight is the aircraft's first of the day: in the reporting years the first leg of an aircraft's day arrived on time 85.8% of the time, against 76.0% for the legs after it, a difference of +9.8 pts (interval +9.5 pts to +10.1 pts). By the clock, departures scheduled at 05:00 were on time 91.2% of the time and those at 20:00 68.8%; November was the best month to fly and July the worst.

A connection is missed when the inbound flight is cancelled or diverted, or lands too late to leave 25 min to change planes. At ATL, the busiest airport, the chance of missing falls under 10% with a scheduled buffer of 65 min (interval 60 min to 65 min); at DEN it takes 70 min, and at DFW 90 min. Across the 20 busiest airports the buffer runs from 55 min to 100 min, and the planner page answers for any connection at them.

For the airline, a minute added to every turn takes delay off the rest of the aircraft's day only where a late inbound reaches into the turn, which was 25.3% of turns in the reporting years. Counting the minute it saves on the next flight and the share that would have carried further down the day, a buffer minute saves 0.46 arrival minutes on average. It does not pay for itself at any of the busiest airports on this count; the most it saves is 0.51 minutes per minute, at LAS, so a buffer is worth adding only where turns are reached far more often than average.