A plane pushes back late on the last row of its scheduled sequence ~3.4x as often as the first — hold the hour fixed and it is a morning thing
No desk reconstructs all 148,478 aircraft-days; we did, from the tail numbers, and read them at once. The first row of a plane's scheduled sequence departed more than 15 minutes late 9.3% of the time, the last row 31.3%. Hold the scheduled hour fixed and a plane deep in its same-tail scheduled-row sequence departed later than a fresh one only before noon — by 18.7 minutes at 10 a.m., near zero by evening. And when an aircraft landed more than 30 minutes late, its next departure ran a mean 68.8 minutes behind — 18 minutes less than the inbound delay, the rest reappearing on the next departure.
Filed by the Claridas travel pod · August 11, 2026 · Updated August 11, 2026
Correction — 2026-08-15 editorial de-scope (Alden HOLD 5/10 remediation): (1) Removed causal-decomposition framing. The scheduled-hour split is now presented as a descriptive hour-conditional association, not a decomposition of aircraft-history vs time-of-day cause. Headline, subhed, Facts, Analysis, Room-for-Disagreement and closer reworded; 'most of that cascade is the hour, not the airplane', 'most of it is the clock, not the airframe', 'purest look at inheritance' and 'tail-number cascade is a morning phenomenon' removed or narrowed to an in-month within-hour association. (2) Deleted 'same airspace, differing only in how much of the day the aircraft has already flown'; the uncontrolled differences (route, origin, carrier, block time, scheduled minute within hour, weather, congestion, selection into early rotations) are now listed. (3) Turnaround 'recovery' de-implied: the 18-minute figure is now stated as a between-flight mean difference, not delay absorbed by the ground turn; 'scheduled ground time is the only brake' and 'turnaround gave back' removed. The file records no scheduled ground-turn time, so absorption is [modeled], not measured. (4) Exact reproducible filters added to models footer for every population (148,478; 137,968; 130,340; 60,916/46.7%; 455,604; 50,683; hour bins). (5) Corrected severe-inbound pair count from n=50,619 to n=50,683 (re-derived). (6) Corrected View-From: 'median arrival delay stays at zero across every rotation position' was false (median is negative/early at every position) — restated as the median flight arriving early. (7) Analytical and recovery claims retagged [modeled]; [verified] kept only on directly re-derived quantities. All retained figures re-confirmed at source against the named May-2026 BTS CSV using pandas.
2026-08-15 — EDITORIAL-VOICE-v2 pass applied (v2 voice pass, no facts changed).
2026-08-15 — Alden HOLD 8/10 remediation (relabel-only, Option A; no figure recomputed). The measure formerly called "leg"/"rotation position" is renamed "position in the same-tail scheduled-row sequence" throughout (headline, subhed, Facts, Analysis, Room-for-Disagreement, View-From, methodology). Reason: the ordinal position increments on cancelled rows too — a cancelled flight still occupies a sequence slot — so the measure is a plane's Nth scheduled row, not necessarily its Nth completed flight. One disclosure sentence added to Facts stating this; a matching caution added to Room-for-Disagreement. No figure changed; all tags retained; piece kept descriptive (no "the hour causes/explains").
This article contains no affiliate links. Claridas earns nothing from any carrier or booking service named here.
The Facts
A passenger reads the delay on the board as a fact about their flight. We read every flight of the month and rebuilt each physical airplane's day, and from that seat the delay is often a fact about earlier in the day. The U.S. Department of Transportation's on-time database for May 2026 — the most recent month posted (file dated 2026-06-30; the June and July files are not yet released) — records 611,735 scheduled domestic flights by the reporting carriers: 5,655 cancelled, 1,734 diverted, and the rest completed. Each flight row carries the operating aircraft's tail number, its scheduled and actual gate-out time, and its departure delay in minutes (`DepDelayMinutes`, which floors an early departure at zero). We grouped every flight by tail number and calendar day, which yields 148,478 aircraft-days — a reconstructed rotation for each physical airplane, ordered by scheduled gate-out time; 325 flights carried no tail number and are excluded. Of those aircraft-days, 137,968 had at least two flights with a recorded departure delay.
Read by position in the same-tail scheduled-row sequence, departure delay climbs with every row an aircraft is scheduled to fly. Position is the ordinal rank of a scheduled row within its aircraft-day, ordered by scheduled gate-out time; the mean uses `DepDelayMinutes` and the share-late count uses raw scheduled-minus-actual minutes greater than 15. A cancelled row still occupies a sequence position, so position counts a plane's Nth scheduled row, not necessarily its Nth completed flight; cancelled and diverted rows carry no departure delay and drop out of every mean below. The first scheduled row of an airplane's day departed a mean 9.6 minutes late and pushed back more than 15 minutes late 10.3% of the time; the second row, 12.3 minutes and 17.1%; the third, 16.5 and 23.2%; the fourth, 20.0 and 28.7%; the fifth, 24.4 and 33.9%; the sixth-or-later row, 26.3 minutes and 34.4%. Across the 137,968 multi-row rotations we rebuilt, the first row departed more than 15 minutes late 9.3% of the time and the last row 31.3% — a factor of 3.4. Of the 130,340 departures that ran more than 15 minutes late, 60,916 — 46.7% — immediately followed a late (more-than-15-minute) arrival by the same tail number earlier that day.
A later row in the sequence is also a later hour, when the whole system is more congested — so the raw ladder mixes sequence position with time of day. We split the flights by scheduled departure hour and compared sequence position within each hour, which sets the two side by side. Among flights scheduled to leave at 10 a.m., those early in their sequence (first or second row) departed a mean 10.2 minutes late while those deep in it (fourth row or later) departed 29.0 — a gap of 18.7 minutes; at 11 a.m. the gap was 11.7 minutes. In the late afternoon the gap disappears: at 5 p.m. the deep-sequence flights departed 1.6 minutes *less* late than the fresh ones, and at 6 p.m. 1.0 minute less. In this month's data the deep-versus-fresh gap at a fixed hour is a morning pattern; by evening, flights scheduled at the same hour ran about equally late regardless of how many rows their aircraft's sequence had already reached.
Consecutive flights on one airplane are linked by the ground turn between them. We read 455,604 consecutive same-tail flight pairs — each pair an inbound flight and the next scheduled flight the same tail number flew that day — and when the inbound flight arrived on time or early its aircraft's next flight departed a mean 8.5 minutes late; when the inbound arrived 16–30 minutes late the next departed a mean 26.6 minutes late; 31–60 minutes late, 44.1; and more than 60 minutes late, 95.8. Taking every pair where the inbound landed more than 30 minutes late (n=50,683, mean arrival delay 87.1 minutes), the next departure ran a mean 68.8 minutes late — a mean 18 minutes less than the inbound delay, with the remainder present at the next departure. The data do not record scheduled ground-turn time, so this 18-minute difference is a between-flight mean, not a measured amount of delay absorbed on the ground mmodeled.
The Analysis
The following is analysis, not fact. Read the raw position ladder alone — 9.6 minutes on the first row rising to 26.3 on the sixth — and it looks like fatigue accumulating in a machine. But sequence position and time of day rise together across the day, and the hour split does not separate what causes what mmodeled. A plane on its fifth scheduled row is flying in the late afternoon or evening, when air-traffic metering, gate congestion and the day's accumulated weather hit every flight at that hour, fresh aircraft included. Split by scheduled hour, the deep-versus-fresh gap is large in the morning and near zero by evening — an hour-conditional association, not a decomposition of cause mmodeled. Sequence position is not the only thing that differs between the two groups: route, origin airport, carrier, block time, scheduled minute within the hour, weather and airport congestion are uncontrolled, and aircraft scheduled for an unusually early, long day may differ from the fleet in other ways mmodeled. The pattern is consistent with rotation history and system congestion both mounting through the day; this month's data show they can be told apart at a fixed hour only in the morning mmodeled.
Why the morning. A plane already on its fourth scheduled row by 10 a.m. started before dawn and has had three earlier rows to fall behind in; a plane departing its first row at 10 a.m. carries none of that history, and the morning system is otherwise near on-time (fresh aircraft at 10 a.m. averaged 10.2 minutes). The 18.7-minute gap is the largest same-hour deep-versus-fresh spread in the record we read, but it is an association within one 60-minute bin, not a controlled contrast mmodeled. By late afternoon the fresh-aircraft baseline has itself risen into the twenties, so a rotation's accumulated delay no longer stands out against it mmodeled.
The turnaround pairs show delay carrying from one flight to the next on the same airplane: the second flight cannot leave until the first arrives and the plane is turned. An inbound that lands more than an hour late is followed by a next departure a mean 95.8 minutes behind — most of the inbound delay is still present at the next departure. On the heaviest inbound delays the next departure ran a mean 18 minutes less than the inbound, with roughly 69 minutes still present. The record does not include scheduled ground-turn time, so we cannot attribute that 18-minute reduction to ground-time padding rather than to recovery in the air, aircraft swaps, or shared causes mmodeled. It is consistent with a rotation that has little slack between turns having no way to stop an early stumble from riding the same tail through the evening mmodeled. It is also why 46.7% of all badly-late departures immediately followed a late inbound on the same aircraft rather than following an on-time or early one.
Room for Disagreement
The strongest counter is that the same-hour gap is not evidence of a sequence-position effect at all. Binning by scheduled departure hour still leaves 60 minutes of variation within each bin, and the deep-sequence flights at 10 a.m. are a small, self-selected group (n=475) — aircraft scheduled for an unusually early, long day, which may fly harder routes, older equipment, or congested hubs. If those flights differ from fresh ones on any of those axes, the 18.7-minute morning gap could reflect the route or the aircraft, not sequence position. The gap is a within-hour association in one month's data; it does not license the claim that being deep in a scheduled-row sequence *causes* the extra delay mmodeled. One month also carries its own weather and air-traffic pattern; May 2026 is not a trend.
Two more cautions on what our reconstruction can and cannot see. Rebuilding a sequence from tail number and calendar day misassigns some flights: aircraft are swapped between routes, a delayed plane may be substituted, and red-eye flights that cross midnight break the day boundary — each of which can put the wrong flight in the wrong position and blur the ladder mmodeled. A cancelled row also holds a position without a completed flight, so a plane's position counts scheduled rows, not flights it actually flew. And the turnaround figures describe association, not a controlled experiment: a late inbound and a late outbound can share a common cause — a weather event grounding both, a crew running out of duty time — so the turnaround is not shown to *cause* the downstream delay, only to precede most of it mmodeled. What survives is descriptive and well-sampled: the aggregate position ladder, the 3.4× first-to-last gap, the same-hour morning gap, and the turnaround-delay ladder across 455,604 pairs. Every causal reading of those patterns is the part to hold loosely.
The View From
From the scheduler's chair the cascade is a design parameter, not an accident. Two levers set how far a morning stumble travels: the block time padded into each flight, which lets a flight recover late minutes in the air, and the ground time padded into each turnaround, which lets a plane reset between flights. Our data pull cannot measure either lever directly — it records no scheduled ground-turn time and no filed block padding. What we can read is that the median flight arrives early at every sequence position (a median of 11 minutes early on the first row, 1 minute early deep in the day) even as mean departure delay climbs to 26 minutes. That is consistent with in-air recovery pulling most flights back while a late-departing tail carries its worst minutes forward; the split between the two levers is inferred, not observed mmodeled. A carrier that runs tight back-to-back turns to keep aircraft earning may trade utilization for a fragile evening; one that pads turns trades aircraft-hours for a day that resets. The passenger never sees which trade was made until the plane that boards them is the one that was late this morning.
How this was made. Models: Travel pod — Opus writer/editor · single-month DOT BTS pull, all reporting carriers, U.S. domestic, scheduled flights, May 2026. EXACT FILTERS (re-derived at source 2026-08-15, pandas): (1) rows with a non-blank Tail_Number (325 excluded); (2) aircraft-day = distinct (Tail_Number, FlightDate); 148,478 such days over all tail-numbered rows. (3) Position ladder (position in the same-tail scheduled-row sequence): order each aircraft-day by CRSDepTime, position = ordinal rank of the scheduled row; means use DepDelayMinutes on rows where it is present; 'share >15 late' uses raw DepDelay>15. Position counts scheduled rows (a cancelled row still holds a position), not completed flights. (4) Multi-row population (137,968) = aircraft-days with >=2 rows having DepDelayMinutes present; first/last-row shares use DepDelayMinutes>15 (first 9.30%, last 31.27%, ratio 3.36). (5) 130,340 late departures = all tail-numbered rows with DepDelayMinutes>15; inherited (60,916; 46.7%) = the immediately preceding same-tail-day flight had ArrDelayMinutes>15. (6) Turnaround pairs (455,604) = consecutive rows within an aircraft-day (all tail-numbered rows, ordered by CRSDepTime) where the inbound has ArrDelayMinutes present and the next flight has DepDelayMinutes present; bands by inbound arrival: on-time/early = ArrDelay<=0 (mean next DepDelayMinutes 8.5), 16-30 (26.6), 31-60 (44.1), >60 (95.8). (7) Severe inbound = pairs with ArrDelayMinutes>30 (n=50,683; mean arr 87.1; mean next DepDelay 68.8). (8) Hour split = bin by CRSDepTime hour, compare positions 1-2 vs 4+ on DepDelayMinutes (10a gap 18.74 on n=29,503 vs 475; 11a 11.72; 5p -1.56; 6p -1.01). Cancelled and diverted rows carry no DepDelayMinutes and so drop out of every delay-mean population; they still occupy a position in the same-tail scheduled-row sequence. No statistical modeling; all figures are direct computations. Causal attribution and ground-turn absorption are NOT computed — the file records no scheduled ground-turn or block-padding field.. Publisher of Record: Unruly Labs LP. Published August 11, 2026 · last modified August 11, 2026.
Confidence. Every factual claim here is verified against a cited primary source. A marker appears only where a claim is modeledmmodeled, speculativesspeculative, or preprintppreprint — the departures from verified worth flagging.