Nobody Flies on an Average Day

Summary

Across 17,368,992 scheduled departures from the 64 airports FlyerIntel covers, 251,579 were cancelled — 1.45%, or one flight in 69, departures-weighted. That number is on this site, on every airline's report card and inside every booking tool that quotes one, and it describes almost none of the 1,096 days it was measured over. Rank each airport's own days by how much of its schedule it cancelled and the rate comes apart exactly. On its quietest half of days a covered airport cancelled 0.055% of what it had scheduled — one flight in 1,808. On its worst tenth it cancelled 9.9% — 178 times as much. Those worst days are 10.0% of the schedule and carry 68.4% of every cancellation in the period. Cancellation is not a hazard spread thinly across a year. It is a small number of days on which a large part of an airport disappears, and the rest of the calendar is close to clean.

Cancellation is published as a rate. FlyerIntel publishes one on every airport page, the Department of Transportation publishes one for every carrier, and every booking site that mentions reliability quotes something derived from them. Over the period behind this report that number is 1.45% for departures from the airports FlyerIntel covers: 251,579 cancellations out of 17,368,992 scheduled departures, one flight in 69. It is a correct number and it is reproduced exactly on the airport pages of this site. It is also, read the way a traveller naturally reads it — as the chance that this flight, on this day, will not operate — close to useless, and this report is an attempt to say precisely why and precisely by how much. The medium is the whole problem: a rate implies a hazard, something that ticks away evenly and occasionally lands on you. Cancellation does not behave like that at all.

Everything below is measured on a single population and a single basis: one record is one scheduled departure from one of the 64 airports FlyerIntel covers, and a cancelled flight is counted at the airport it was scheduled to leave. That is exactly the population behind the cancellation rate shown on each airport page, and the figures reconcile to it flight for flight. Nothing here is an arrivals measure and no figure mixes the two. Every network rate is departures-weighted from summed counts — total cancellations over total scheduled departures — rather than averaged across airport averages, and the difference is worth printing: the departures-weighted network rate is 1.448% while the plain average of the 64 airport rates, one airport one vote, is 1.431%. Days, unlike airports, are close to the same size as each other, so for the calendar the two agree: the departures-weighted rate across all 1,096 days is 1.448% and the plain average of the 1,096 daily rates is 1.448%. Every figure in this report, including the split-half, seasonality and persistence checks, is derived from one grid of 70,144 airport-days built in a single pass, so no two statements in it can be describing different sets of flights.

Each of the 1,096 days in the period is one observation: the cancellations recorded that day across all covered airports, over the departures scheduled that day. 74% of the days in the period were quieter than the published rate that is their own average. On 1 of the 1,096 days not one scheduled departure from a covered airport was cancelled; on January 25, 2026, 7,251 of 15,281 were.

The chart above is the 1,096 days of the period ranked by the national cancellation rate that day, quietest first, cut into ten groups of 109 or 110 days, with each bar that group's share of every cancellation in the period. A rare event will always look lumpy in a short window, so the first thing to rule out is that this is arithmetic rather than aviation. It is not, and the margin is not close. If cancellations struck flights independently at the network rate of 1.448%, a day of the average size in this period — 15,848 scheduled departures — would have a cancellation rate with a standard deviation of 0.095%. The observed standard deviation across the 1,096 days is 2.84%: 30 times wider, an index of dispersion of 894 against the 1 that independence would give. The same point can be made without any statistics at all. The quietest half of the calendar is half the schedule, and under independence it would carry very close to half the cancellations, because chance alone moves a day only 0.095% either side of 1.45%. It carries 7.2%. The worst tenth of the calendar carries 52.3%, and the worst 11 days of the three years — one day in a hundred — carry 13.6% on their own.

The obvious explanation for a lumpy calendar is that winter exists. It is a real effect and it is not this one. Measured strictly inside single calendar months — so that January is only ever compared with itself and never with July — the concentration barely moves. Across 36 months averaging 30.4 days, the worst single day of its own month carried 19.6% of that month's cancellations, and the worst three days, a tenth of the month, carried 43.3%. Knowing the month tells you the general level. It does not tell you which days, and the days are where almost all of it is. What the days do have in common is the reason the operating carrier gave for the cancellation: the further up the calendar you go, the more of it is weather.

The cancellation reason is the single code the operating carrier files with the DOT, and every cancellation in this dataset carries one, so the four columns account for all of them. The rows are the same calendar ranking as the chart above, so the second row contains the first.

The calendar above is national, and a national day is not a thing anyone experiences. So the rest of this report ranks each airport's own days within that airport, quietest first, and cuts them at the halfway mark and at the ninetieth percentile. That produces an exact decomposition rather than a comparison: because the three groups tile the airport's calendar and every cancellation it recorded fell on exactly one of its days, the cancellations from the three groups over the airport's whole schedule must add to the published rate. They do, at every airport, and this generator refuses to publish the report if they ever fail to. Network-wide the arithmetic reads 0.028% plus 0.431% plus 0.990% equals 1.448%. Put as an experience rather than an identity: on its quietest half of days a covered airport cancelled 0.055% of the flights it had scheduled, one in 1,808; on its middle four days in ten, 1.07%; and on its worst day in ten, 9.9% — 178 times the quiet-day rate.

Each airport's own days, ranked within that airport. The last column sums to the published network rate of 1.448% exactly, because the three groups tile every airport's calendar. Note that this is not the national calendar used in the chart above: an airport's quiet day is not the country's quiet day, which is why the two sets of shares differ.

Run that decomposition airport by airport and the ranking everyone publishes turns out to be a ranking of one thing only. What an airport's quietest half of days contributes to its published rate varies across the whole country by 0.064 of a percentage point, from 0.000% to 0.064%. What its worst tenth contributes varies by 2.01 points, from 0.342% to 2.355%. Because the published rate is identically the sum of the three parts, the between-airport variance in it splits exactly between them: 0.4% of it comes from the quietest half of days, 23.8% from the middle four in ten, and 75.8% from the worst one in ten. The correlation between an airport's published rate and its worst-tenth contribution is 0.97; with its quiet-half contribution, 0.11. Take the thresholds away entirely and the same thing shows up in plain order statistics: the correlation between an airport's published rate and its own median-day cancellation rate is 0.09, and with its 90th-percentile day, 0.91. LGA (New York) has the highest published cancellation rate of any covered airport, 3.14%. On its median day it cancelled 0.33% of its schedule — a shade less than LAX (Los Angeles) on its median day, 0.35%, whose published rate is 0.77%. The two airports are indistinguishable on an ordinary day. They part company entirely on LGA's worst tenth, which cancelled 22.9% of what it had scheduled against 4.1% at LAX.

Every column is a departures measure over the same population. The published rate matches the cancellation figure on the airport's own FlyerIntel page. The last three columns add to the published rate exactly, at every row. The middle three describe the airport's own daily distribution: the day-level columns are coarser at airports scheduling few departures a day, where a single cancellation is a large share of the day, which is why the network figures are restated in the methodology over only the 47 airports averaging at least 100 departures a day.

If disruption arrives in days rather than as a rate, the practical question is whether those days announce themselves. They do. An airport is in its own worst tenth of days, by construction, about 10.0% of the time. On the day after one of those days it is in its worst tenth 42.7% of the time, across 7,038 consecutive-day pairs. The obvious objection is that bad days cluster in bad months, so the comparison is not argued away here but measured: on those same pairs, the share of that airport's days in that same calendar month that were in its worst tenth is 20.3%. Knowing the airport and the month lifts the odds from 10.0% to 20.3%; knowing that yesterday was bad lifts them the rest of the way to 42.7%. The reverse holds too: after an ordinary day the chance is 6.4%, below the 8.9% that the month alone would suggest. And the effect decays slowly enough to matter for rebooking, which is the one decision a traveller makes with this information. What it does not decay across is the map. On the days an airport was in its own worst tenth, the rest of the country — that airport removed from the arithmetic so it cannot be correlating with itself — cancelled 6.0% of its own schedule on average, against a median day of 0.49%. On 84% of those days the rest of the country was itself cancelling at or above the network's published rate, and on only 5% of them was it having an ordinary day or better. A bad day at an airport is mostly a bad day for the system, not a local failure.

A reasonable alternative reading of all of the above is that it describes carriers rather than places: that some airlines cancel in bulk when a day goes wrong and others trim flights steadily, and that an airport inherits the shape of whoever flies there. Cut the same days by operating carrier and that is not what comes back. Every one of the 9 carriers large enough to compare shows the same shape — a cancellation rate on the quietest half of the airport's days between 0.013% and 0.222%, and a rate on the worst tenth between 6.3% and 23.1%. What this report will not do is rank them. Restating each carrier onto a common national airport mix — the same direct standardisation used elsewhere on this site — moves the quiet-day figures by as much as 3.2 times, at Republic Airways, and 61% of the 36 pairs of carriers here are closer together than that. Where a carrier flies dominates what it looks like on a quiet day, and both columns are printed below so that is visible rather than asserted.

The day groups are the airport's, not the carrier's: a day is quiet or bad according to what happened at that airport across all carriers, and the carrier's own cancellations on those days are then counted. A carrier appears only where it qualifies at 20 or more airports with at least 500 scheduled departures inside the airport's worst tenth, and the raw and restated columns are computed over exactly the same cells so the only difference between them is the weighting.

Every claim made here about an airport is implicitly a claim that it will behave that way again, so each measure was recomputed on the first and second halves of the period independently, each half ranked on its own days, and correlated across the 64 airports. The published rate replicates at 0.75 and the worst-tenth contribution — the quantity this report rests on — at 0.77, so the structural claim is about something stable. The quiet-half and middle contributions replicate at 0.60 and 0.64. But the most quotable version of this finding does not survive. Measured on the worst 1% of an airport's days rather than the worst 10%, the contribution replicates at only 0.46 — a ranking that would reshuffle if the period moved. Whether an airport has bad days often is a property of the airport. Which handful of days blow up, and how catastrophic the very worst of them are, is not something three years of data can rank airports by, and this report does not try. Everything above is stated on the worst tenth for that reason.

Three things follow, and none of them is on a booking site. First, the cancellation figure attached to an airport or an airline is not the risk to your flight; it is that risk blended with a small chance of a very bad day. On its median day not one of the 64 covered airports cancelled more than 0.81% of its schedule — CLE (Cleveland) is the worst in the country on that measure — so the published spread between the best and worst airports mostly is not about your ordinary day. Second, what the published rate really measures is exposure to bad days, and that is worth knowing when the trip cannot absorb one: a wedding, a cruise, a single-connection itinerary in February. Third, and most usable: if your airport had a bad day yesterday, today is 4.3 times more likely than usual to be another one, and the elevation is still visible 7 days later, so rebooking onto tomorrow at the same airport during a disruption is a worse trade than it looks. Rebooking through a different part of the country is better but not a rescue: on the days an airport was in its own worst tenth it cancelled 9.9% of its own schedule while the rest of the country was cancelling 6.0% of its own, against 0.49% on a median day. These are days when the system, not the airport, is having the problem.