Making Up Time in the Air

Summary

Subtract a flight's departure delay from its arrival delay and the schedule falls out of the arithmetic exactly: what is left is the time the aircraft actually took from gate to gate, measured against the time its schedule allowed. Across 17,073,131 completed departures from the 64 airports FlyerIntel covers, departures-weighted throughout, the average flight finished the journey 5.74 min inside its own schedule, and 73.3% of flights beat their block time. That spare time is the reason a departure delay usually shrinks by the time you land. What it is not is a response to being late. Comparing flights on the same route flown by the same airline, one that pushed back an hour or more late handed back 3.85 min against 6.13 min for the same route leaving on time — 2.28 min less, not more. Effectively none of that difference is in the air: airborne time between the two groups is 1 second apart, in the direction of the late flight being airborne longer. The pooled figure is itself two small opposite effects cancelling — short sectors airborne slightly longer when late, long ones slightly less — and the larger of them is 2.07 min on a sector where the aircraft is airborne about 288 min. The minutes the late flight loses are spent taxiing, at both ends.

There is a sentence every frequent flyer has heard from the flight deck after a long wait at the gate: we have got a good tailwind today and we will try to make up some of the time in the air. It is reassuring, it is plausible, and it is one of the very few claims an airline makes about its own operation that can be checked exactly, for free, against a public record. The US Department of Transportation publishes, for every flight, the delay at pushback and the delay at the gate on arrival, and separately the minutes spent taxiing at each end. That is enough to say precisely how long the aeroplane spent in the air, how long its schedule expected it to, and whether the two differ when the flight is running late. This report does that for every flight in the database.

Departure delay is the actual gate departure minus the scheduled one. Arrival delay is the actual gate arrival minus the scheduled one. Subtract the first from the second and the two scheduled times collapse into the scheduled block time, the two actual times into the actual block time, and what is left is the difference between them. No timezone enters it, because both delays are differences taken at a single airport. A flight with an arrival delay smaller than its departure delay took less time gate-to-gate than its schedule allowed; the number of minutes is exact and needs no assumption at all. Everything in this report is measured on one basis and one population: one record is one completed flight that departed one of the 64 airports FlyerIntel covers, and every network figure is departures-weighted from summed counts rather than averaged across airport averages. Cancelled flights never left and diverted flights did not arrive where they were scheduled, so neither has a block time and neither is counted. Nothing here is an arrivals measure and no figure mixes the two.

Over 17,073,131 completed departures from covered airports. A negative number means the flight took less time gate-to-gate than its schedule allowed. This is not an accusation of padding: a schedule has to cover a bad day as well as a good one, and a block time set at the average would make half of all flights late by construction.

The distribution above is the reason a late pushback is not a sentence. Departures-weighted across covered airports, the mean departure delay is 13.07 min and the mean arrival delay 7.33 min: the system gives back 5.74 min of every flight, on average, between the two gates. 73.3% of flights finish inside their scheduled block and 23.0% finish a quarter of an hour or more inside it. So far this is the folklore confirmed, and it is worth saying clearly before the rest of this report complicates it: schedules do contain slack, the slack is large, and it does absorb delay. The question this report exists to answer is a different one. Is the slack spent *because* the flight is late?

All nine columns are departures measures over one population. The taxi-in column is an attribute of these same flights at the far end, not an arrivals average for any airport.

Read the fifth column of that table downwards. A flight that pushed back on time or early finished 6.00 min inside its schedule. One that pushed back between 15 and 59 minutes late finished 5.31 min inside it. One that pushed back an hour or more late finished 3.88 min inside it. The direction is the opposite of the announcement: the later the flight left, the less of its schedule it managed to give back. That raw comparison is not yet worth much, because the three groups are not the same flights. Late departures are concentrated at congested airports, in bad weather, at the worst hours, and on whatever route mix happens to go wrong most — and route length alone changes how many spare minutes a schedule contains. Everything that follows holds those things constant.

Every row is computed inside the same 2,103 (origin, destination, operating carrier) cells — each with at least 200 flights in every one of the three groups — and reweighted onto one common mix, so the rows describe an identical set of routes and airlines rather than three different ones. The airborne and scheduled block columns carry each route's own constant timezone offset, which is why they are only ever compared down a column within this fixed set of routes and never quoted as absolute flying times. The five columns satisfy an exact identity — minutes against schedule = airborne + taxi-out + taxi-in − scheduled block — and the generator refuses to publish this report if they do not reconcile.

Holding the route and the operating carrier constant, the effect gets stronger rather than weaker. A flight leaving on time gave back 6.13 min; the same route flown by the same airline leaving an hour or more late gave back 3.85 min. The gap is 2.28 min, and the table above says exactly where every second of it went. Extra taxi-out at the origin: 2.14 min. Extra taxi-in at the destination: 0.36 min. A difference in the scheduled block times themselves, pulling the other way by 0.23 min, because a delayed departure is not drawn evenly from a route's own timetable. And airborne time: 0.01 min, which is 1 second, in the direction of the late flight being airborne longer. That is the whole of the tailwind. Two robustness checks say the same thing. Additionally holding the scheduled departure hour constant, so the two groups are drawn from the same slot in the same day, the gap is 1.09 min with airborne time moving -0.44 min; additionally holding the calendar month constant, 2.98 min with airborne time moving 0.72 min.

The same cells as the previous table, split by the length of the route. Airborne minutes carry each route's timezone offset and are comparable only within a band, down the column, never across bands.

Pooling the whole country hides two effects that happen to be pointing in opposite directions, and reporting only the pooled figure would be the exact mistake this report is trying to avoid. Split by route length, the airborne column does move, and it moves sensibly: on flights under 750 miles the late aircraft is airborne 0.38 min longer than the on-time one, while on flights of 2,000 miles and over it is airborne 2.07 min shorter. A long sector has enough cruise in it to be flown faster and a short one does not, which is what an operations department would predict. But the size is the point. The largest airborne saving in the table is 2.07 min on flights of 2,000 miles and over, on a sector where the aircraft is airborne roughly 288 min — under one per cent of the flight. And it does not survive contact with the ground: on the same band the late flight loses 2.44 min taxiing, so it still finishes 0.19 min worse against its schedule than the on-time flight on the same route. 2 of the 4 aircraft airborne longer, not shorter, and 51.0% of the flights in this comparison are on flights under 750 miles, where there is no recovery to be had at all. The honest summary is not "pilots never make up time". It is that the amount available to be made up is a rounding error next to the delay it is supposed to fix, and is spent again before the aircraft has left the airfield.

The same standardisation, split by operating carrier instead of pooled, for carriers with at least 20 qualifying route cells. Read across a row, not down a column: the level of a carrier's block-time margin reflects how it builds a schedule and which routes it flies, and is not a ranking. The last three columns are what this table is for.

A pooled result of almost exactly zero invites the suspicion that it is two real effects cancelling, so it is worth checking carrier by carrier as well as by route length. It is not. Every one of the 11 carriers gives back less schedule when it leaves an hour late than when it leaves on time — the same direction every time. All of them taxi out longer when they are late. The airborne column, by contrast, has no direction at all: 6 carriers are airborne less when late and 5 are airborne more, and the largest movement in either direction is 1.27 min (Republic Airways). A real operational practice of flying faster when behind schedule would show up as one sign repeated down that column. What is there instead is noise around zero, in an industry where every carrier has the same incentive and the same fuel bill.

The same quantity across the day makes the mechanism hard to miss, once one hour is handled honestly. 23:00 is not an ordinary hour of the schedule: the mean sector there is 1,673 miles against 887 across the day as a whole, because what departs then is the red-eye bank, and route length is the largest single driver of how much schedule a flight carries. Both versions are reported here rather than one. Across all 19 hours the correlation between an hour's margin and its taxi-out is 0.54 and the correlation with its mean route length is -0.63 — on those two numbers alone route length is the better explanation and the taxi story should be dropped. Excluding the red-eye hour, the route-length correlation collapses to -0.02 across the remaining 18 hours while the taxi-out correlation holds at 0.56. Within the ordinary working day, then, a flight scheduled out at 05:00 gave back 7.61 min of its schedule and one scheduled at 16:00 gave back 4.47 min, and the shape of that curve tracks taxi-out rather than distance. This is an association across 18 hours, not a measurement of cause, and it rests on excluding an hour — which is why the excluded hour, its size and the figures with it left in are all printed above. It points the same way as the standardised comparison, from a different direction: the slack in a schedule is consumed by the airport surface, and the surface is worst at the hours when flights are already running late.

Every column is a departures measure over flights that left this airport, including the taxi-out column. Airports with at least 5,000 completed departures, ranked by how much of their schedule their flights give back. The mean route length column is there because it is the confound for the rest of the table.

Two housekeeping results, both of which happen to be undramatic and are reported anyway. First, the weighting. Averaging the 64 airport averages together, one airport one vote, gives -5.72 min; weighting each airport by the departures it actually operated gives -5.74 min. On this particular quantity the choice barely matters, which is worth stating precisely because on other quantities in this dataset it decides the answer outright. Every network figure in this report is the departures-weighted one. The late-versus-on-time gap behaves the same way: 2.14 min weighted against 2.26 min unweighted, and 59 of the 64 airports show the late departure recovering less than the on-time one. Second, an explanation that looked good and is mostly false. JFK (New York) gives back the most schedule of any covered airport at 8.01 min, and it has a taxi-out of 27.2 min; HNL (Honolulu) gives back the least at 4.14 min on a taxi-out of 16.4 min. The tempting reading is that airlines pad their schedules where the ground is slow, and across all 64 airports the correlation is -0.30, which looks like support. But the airports with slow surfaces also fly longer average sectors (-0.19 between margin and mean route length), and route length is the dominant driver of margin. Restate every airport's margin and taxi-out onto one common route-length mix — the national one, the same standardisation used earlier in this report — and the correlation falls to -0.17 across the 39 airports that qualify in every one of the 4 route-length bands. Computed separately inside each band it is -0.10 on under 750 miles, -0.25 on 750–1,499 miles, 0.01 on 1,500–1,999 miles, -0.25 on 2,000 miles and over, which is the more informative result: not only is it weaker than the headline number, it is not the same strength twice. Roughly half of the apparent relationship was route mix and what is left is weak and unstable across route lengths. Something may genuinely be there. It is not enough for this report to claim that airlines pad schedules where the ground is slow, so that claim is not made.

Share of flights that pushed back this many minutes late and still arrived within fifteen minutes of schedule, at one-minute resolution over the whole period. A minute is shown only where the band recorded at least 120 flights at exactly that departure delay. The second column is the last whole minute at which the share is still above half; by the following minute it is below.

The practical version of all of this is a single number. Across every covered airport, a flight that pushed back 22 minutes late arrived on time 50.05% of the time — the last whole minute of delay at which the odds are still on your side, and the point at which a departure delay stops being absorbed and starts being inherited. Route length moves that line, but much less than the folklore suggests — from 21 minutes on flights under 750 miles to 25 on flights of 2,000 miles and over, a difference of 4 minutes. What route length really buys is not a later crossing but a longer tail beyond it. At 30 minutes down, 15.1% of flights under 750 miles still arrived on time against 39.2% of flights of 2,000 miles and over; at 45 minutes down it is 0.54% against 11.92%. A transcontinental flight does not give you meaningfully more grace before the odds turn against you. It gives you a real chance of recovery in a situation where a short hop has none — which is the same finding as the airborne table above, seen from the passenger's side.

Four limits, in descending order of how much they could matter. First, and most importantly, this report measures flights that finished. A flight that gives up and lands somewhere else leaves the population, and it does so more often the later it departed: 0.21% of flights pushing back on time were diverted against 0.57% of those pushing back an hour or more late. Those are small numbers, but they are removed disproportionately from the group that is already doing worst, which means the recovery figures for late departures are if anything flattering. Second, the airborne column is derived from three reported fields rather than observed directly, and it inherits whatever the carrier reported for taxi-out and taxi-in; it is also block time minus taxi rather than wheels-off to wheels-on as an air traffic system would define it. Third, the scheduled block time carries each route's timezone offset, so it is only ever compared within a route and never quoted as an absolute — and within a route it still varies, because carriers retime flights across three years, which is why it is a measured column in the decomposition rather than an assumed constant. Fourth, this dataset contains no winds, no aircraft type, no cruise altitude and no flight plan. It can say that airborne time does not change when a flight is late; it cannot say whether a particular crew asked for a shortcut and got one.

The slack in an airline schedule is real and it is generous — 5.74 min on the average flight, 73.3% of flights finishing inside their block time. Treat it as a standing discount rather than a rescue service. It is already priced into the departure delay you are looking at on the app, it does not get larger because your flight is in trouble, and on the evidence here it gets smaller. Three things follow. If you are watching a delay grow and doing connection arithmetic, the honest expectation is that you will land about 3.85 min better than the departure delay you can see, not the half hour the announcement implies — and past 22 minutes down, more than half of these flights arrive late. If the leg in trouble is a short hop, assume no recovery at all; the aircraft is airborne no less time than usual and will taxi longer at both ends. And if you have a choice between two itineraries with the same total delay exposure, the one whose long sector is the one at risk is the better bet, because that is the only place in this data where an aeroplane measurably makes up time. None of this predicts an individual flight. It describes what happened to 17,073,131 of them.