The First Flight of the Day

Summary

Fly early is the oldest piece of travel advice there is, and it is correct. It is also measured on a statistic that quietly excludes the flights it should be warning you about. An on-time rate is calculated over flights that operated: a cancelled flight has no arrival time, so it is not counted late — it is not counted at all. Across 17,321,677 scheduled departures from the 64 airports FlyerIntel covers, departures-weighted throughout, a flight scheduled to leave in the first bank of the day, 05:00–06:59, arrived on time 91.2% of the time — the best of the six bands the day divides into. It was also cancelled 1.35% of the time, against 1.08% for a mid-morning departure: 1.25 times as often. Charge those cancellations back to the hour that was supposed to fly and the first bank still wins, by 7.9 points over mid-morning instead of 8.2. The advice survives. What does not survive is the assumption that one number describes the risk.

An American on-time rate has a denominator that is easy to miss. Following the Department of Transportation's own method, a flight counts as on time if it arrives within fifteen minutes of schedule, and the rate is taken over the flights that arrived. A cancelled flight never arrives, so it has no arrival delay and cannot be counted late. It leaves the calculation entirely. Every on-time percentage in America is built that way — the DOT's, every airline's, every booking site quoting one, and FlyerIntel's own. It is the right arithmetic for the question "how late do flights run", and it is the wrong arithmetic for the question a passenger is actually asking, which is "will I get there". Those two questions have the same answer for most of the day. They do not have the same answer at six in the morning. This report puts a second measure alongside the first and keeps both on one basis: one scheduled departure from a covered airport, placed in a bucket by its scheduled local departure time. The on-time rate divides by the flights that operated. The whole-schedule rate divides by the flights that were sold — so a cancellation, and a diversion, count against the hour that was supposed to fly.

The two charts above are drawn from one population and they do not agree about the morning. Punctuality is a slide: 92.2% of flights scheduled out at 05:00 arrived on time, and the figure falls hour after hour to 69.0% at 20:00 before the last departures of the night recover slightly. Cancellation is a valley. It is lowest at 11:00, at 1.05%, climbs all afternoon to 2.31% at 22:00 — and it is already 1.31% at 05:00 and 1.36% at 06:00, before the day has begun. The best hour to be scheduled and the safest hour to be scheduled are not the same hour. Published rankings of departure times, this site's own included, are drawn from the first curve alone.

A network figure can be built two ways and they are not interchangeable. Averaging each airport's rate for an hour, one airport one vote, treats a hundred departures from a small airport as equal to twenty thousand from a hub. Weighting each airport by the departures it actually operated describes what happens to passengers. Every figure in this report is departures-weighted, and the difference is not cosmetic. Taken over one identical set of 1,163 airport-hours, each with at least 120 departures, the unweighted average makes 23:00 the 2nd-worst hour of the day for cancellation at 2.56%; weighted by departures the same hour is 1.57% and ranks 8th, because the airports that still have departures on the board at that hour are mostly small ones with thin schedules. The choice even reverses the order of the two hours this report is about: unweighted, 05:00 looks worse than 06:00 (1.42% against 1.35%); departures-weighted it is better (1.31% against 1.36%). Neither is wrong. They answer different questions, and only one of them answers the passenger's.

All five columns are departures measures taken at the origin airport over the same population. The last column counts a scheduled flight as a success only if it operated and arrived within fifteen minutes; cancellations and diversions both count against it.

This is the test the report exists to run, and it is worth reporting the result plainly even though it is undramatic. On the ordinary measure the first bank beats a mid-morning departure by 8.25 points and an evening departure by 21.56. Charge every cancellation and every diversion back to the hour that was scheduled to fly, and the first bank beats mid-morning by 7.95 points and the evening by 21.63. The morning premium over mid-morning shrinks by 0.30 of a point. Against the evening it grows by 0.06. So the cancellation penalty on the first bank is real, it is measured over 1,402,268 departures, and it costs the first bank a small fraction of a large advantage. Anyone who wanted this report to say "actually, do not fly early" would have to overstate a number that does not support it. Fly early. The finding is not that the advice is wrong; it is that the advice has a shape nobody has drawn, and the shape says the first bank buys punctuality with a slightly higher chance of the flight not existing.

Each figure is the share of scheduled departures cancelled under that code. The three rows are computed inside the same 1,862 (airport, carrier, calendar month) cells — each with at least 200 departures in every band — and reweighted onto one common mix, so the bands describe an identical set of airports, airlines and seasons rather than three different ones.

Strip out the fact that different airports, different airlines and different months are busy at different hours, and one row of that table is almost flat. Cancellations the operating carrier filed against itself — the code the DOT defines as circumstances within the airline's control, which is maintenance, crew, cleaning, fuelling, loading — run 0.377% in the first bank, 0.355% mid-morning and 0.362% in the evening. That is a swing of 0.021 of a percentage point across the whole working day. Weather swings 0.449 points over the same cells and the air traffic system swings 0.213. The implication is worth stating slowly, because it is the opposite of the folk theory. Choosing an early departure does not reduce your exposure to a broken aeroplane or a crew that timed out; on this evidence that risk is roughly constant from breakfast to bedtime. What choosing an hour changes is your exposure to weather and to congestion, and those are the two things that build up over a day. The first bank's own excess over mid-morning is the same story read backwards: of the 0.184 points by which the first bank is cancelled more often, 86% is weather-filed. It is not the airline being slow to get started. Taken carrier by carrier rather than pooled, the difference between the two codes is a difference in consistency. Of the 6 largest carriers in these cells, the first bank's weather rate is above that carrier's own mid-morning weather rate at 6 of 6 — the same direction every time. Their own-code rate goes up at 4 and down at 2, with no carrier moving more than 0.162 of a point in either direction (Spirit Airlines, the largest), which is why they cancel to 0.021 pooled against 0.159 for weather. Individual carriers do diverge from each other later in the day, and the table below shows that; what none of them shows is a morning of its own making.

The same standardisation, split by operating carrier instead of pooled. Read down a row, not across the column: the level of a carrier's own-code rate reflects how it files as much as how it operates, and is not comparable between carriers. The shape across the three bands is what this table is for.

Everything in the previous two sections rests on the cancellation code, so it is worth being precise about what that field is. It is the operating carrier's own filing. The DOT collects the reason the airline gave; it does not adjudicate it. Read across carriers, the field is close to unusable as a measure of fault, and the site's own cause tables show it plainly. At ORD (Chicago), over 14,867 coded cancellations, SkyWest Airlines attributed 1.2% of its own cancellations to itself and Delta Air Lines attributed 51.5% — on the same airfield, in the same weather, over the same three years. Nor does accepting less blame go with cancelling less: SkyWest Airlines cancelled 2.43% of its departures there against 1.09% for United Airlines. None of that is a judgement of any of these airlines. They fly different aircraft on different missions from different concourses, and the point is exactly that the code cannot separate any of those things from a house style for filling in a form. Across all 112 airport-carrier pairs at 44 covered airports with at least 500 coded cancellations, the correlation between how much of its own cancelling a carrier attributes to itself and how often it actually cancels is -0.39 — the wrong sign for a fault ranking, and about the right sign for a difference in filing convention. This report therefore never compares one carrier's codes to another's. Every cause figure here is a comparison of one carrier against itself at different hours of its own day, inside one airport and one month, which is exactly the comparison a filing convention cancels out of. The assumption that remains is that a carrier files the same way at six in the morning as at six in the evening. That is not provable from this dataset, and it is the load-bearing assumption behind the finding above.

Each departure is classified by what happened at the same airport the previous evening. Only airport-days where the previous evening had at least 10 scheduled departures between 17:00 and midnight are used.

Split every morning by what happened at that airport the evening before, and the first bank's disadvantage almost vanishes on the good days. After an evening on which the airport cancelled not one departure between five and midnight, the first bank was cancelled 0.44% of the time and mid-morning 0.40% — 0.04 of a point apart, which is nothing. After every other kind of evening the same two bands were 2.97% and 2.12%, 0.85 points apart. The whole of the first bank's cancellation penalty lives in the mornings that follow a night that went wrong. Two explanations fit that and this dataset cannot separate them. One is carry-over: the aeroplane and the crew that were meant to operate the 06:10 are in the wrong city, and the flight is cancelled overnight rather than at the gate. The other is simple persistence: a storm that closed an airport at nine in the evening is frequently still there at six the next morning, and the two bands are not independent draws from the weather. Distinguishing them would need aircraft tail numbers to trace a specific airframe from one day to the next, and the DOT on-time record does not contain them. What can be said without choosing between the two is the part a traveller can act on: the first departure of the day is a reliable flight on an ordinary morning and an exposed one on the morning after a bad night, and you can check which kind of morning you are about to have the evening before.

Airports with at least 5,000 scheduled departures in each of the three bands, ranked by the first bank's whole-schedule on-time rate. The on-time columns are separated by far more than their sampling error at every airport here. The cancellation columns are much closer to theirs: at 23 of these 59 airports the gap between the first bank and mid-morning sits inside twice its own standard error, so read those columns as a pattern across the table before reading them as a fact about one airport.

On the whole-schedule measure the first bank is the best band of the day at every single airport compared: it beat mid-morning at 59 of 59 and the evening at 59 of 59, by as little as 1.7 points at ANC and as much as 15.9 at SFO. There is no airport in this data where flying later in the day is the reliable choice. Cancellation is where the picture breaks up. The first bank was cancelled more often than mid-morning at 45 of the 59 airports and less often at 14. Not every one of those gaps is separable from the airport's own sampling error, and it matters which are: at 34 airports the first bank is worse by more than twice the standard error of the difference, at 2 it is better by that margin, and the remaining 23 are gaps too small for this data to call either way. LGA has the widest at 0.82 points and is comfortably clear of its own error bar. But no single airport is what carries the finding — the count is: 45 of 59 pointing the same way, where an even split would put about 30 on each side. Which airports they are is also not random. The first bank pays most at airports that see winter and convective weather, and pays nothing, or is rewarded, at the dry western fields. That is an association between climate and the size of the effect, consistent with the cause codes, and not a measurement of cause.

Calendar months pooled across the whole period, departures-weighted.

The first bank costs more in July than in any other month — 2.26% against 1.47% mid-morning, 0.80 points apart — and least in November, where the gap is -0.08. It runs against the first bank in 11 of the twelve calendar months. The months where it bites are the two disruption seasons rather than one: the winter, and the convective summer. That is what makes the aircraft-out-of-position and the weather-persistence explanations so hard to pull apart — both seasons produce evenings that strand aeroplanes and mornings that are still bad.

Take the early flight. That has not changed and this report does not overturn it: across 1,402,268 first-bank departures the whole-schedule on-time rate is 89.8% against 68.2% in the evening, and the gap holds at every airport measured. Three things are worth adding to it. First, the number you are choosing on is not the number you have been shown: the ordinary on-time rate flatters the early departure by 0.30 points against mid-morning, because it silently drops the flights that were cancelled. Second, the thing an early departure protects you from is weather and congestion, not the airline — on this evidence a mechanical or crew cancellation is about as likely at dawn as at dusk, so if that is the risk you are worried about, the departure time is not the lever. Third, the first bank is conditional in a way the rest of the day is not. On an ordinary morning it is the safest flight on the board. On the morning after an evening that went badly at your airport it is the most exposed part of an already bad situation, and that is knowable the night before. None of this predicts any individual flight. It describes what happened to 17,321,677 of them.