The warehouse

Where every number comes from, and how it was checked.

Every scheduled flight of the reporting carriers from January 2015 to July 2026 went through named quarantine rules (flagged and counted, nothing deleted) into a dbt warehouse of 18 models and 206 tests. The metric layer defines each metric once over the flights and once over the shipped mart and compares them at 4 grains.

The quarantine report

Rule	What it catches	Rows	Share	Excluded from
duplicate key	a second row with the same carrier, date, flight number, origin, destination and scheduled departure	4	0.00%	everything
actual without scheduled	an actual departure, arrival or elapsed time with no scheduled time to measure it against	2	0.00%	everything
impossible time	a negative taxi time, or an air time or an actual or scheduled elapsed time at or below zero or longer than the longest a domestic flight can take	31	0.00%	everything
elapsed mismatch	actual elapsed time that differs from taxi out plus air time plus taxi in by more than the tolerance	15	0.00%	everything
status conflict	a row flagged both cancelled and diverted, which cannot both have happened to one flight	13	0.00%	everything
missing actual	a flight neither cancelled nor diverted that reports no departure delay, no arrival delay, no actual elapsed time or no air time	974	0.00%	everything
cause mismatch	reported delay causes that do not sum to the arrival delay within the tolerance	119	0.00%	the cause shares in chapter 6
tail missing	no tail number on the row	338,715	0.46%	the rotations of chapter 4
tail format	a tail number that is not a valid FAA registration (N, a digit 1 to 9, at most two trailing letters) even after a missing leading N is restored, such as an airline's own fleet number	1,142,859	1.54%	the registry join; the rotations keep the row, since the value still names one aircraft
revision	a row that a later download of the same month changed (the later kept, the change counted)	0	0.00%	nothing; the earlier version is the one that is replaced

Source: BTS Reporting Carrier On-Time Performance (real); population: every row of every monthly file in the window; as of 2026-10-01.

The quarantine report by carrier and month

Rows flagged by each rule, by carrier and month; a row can trip more than one rule

Source: BTS Reporting Carrier On-Time Performance (real), the warehouse's mart_quarantine.

The metric layer, reconciled

3,501 cells compared across 4 grains, 0 disagreements.

Metric	Over the flights	Over the mart
scheduled flights	count(*)	sum(scheduled)
cancellation rate	avg(cancelled::int)	sum(cancelled) / sum(scheduled)
completion factor	1 - avg(cancelled::int)	1 - sum(cancelled) / sum(scheduled)
diversion rate	avg(diverted::int)	sum(diverted) / sum(scheduled)
on time rate	avg((arr_delay < 15)::int) filter (where not cancelled and not diverted)	sum(on_time) / sum(flown)
mean arrival delay	avg(arr_delay) filter (where not cancelled and not diverted)	sum(arr_delay_sum) / sum(flown)
median arrival delay	quantile_disc(least(greatest(arr_delay, -60), 300), 0.5) filter (where not cancelled and not diverted and arr_delay is not null)	median_from_histogram
mean scheduled block	avg(crs_elapsed) filter (where not cancelled and not diverted)	sum(crs_elapsed_sum) / sum(flown)
mean actual block	avg(actual_elapsed) filter (where not cancelled and not diverted)	sum(actual_elapsed_sum) / sum(flown)
mean padding	avg(padding)	sum(padding_sum) / sum(padding_flights)
padding share	sum(padding) / sum(crs_elapsed) filter (where padding is not null)	sum(padding_sum) / sum(padding_block_sum)
reported late aircraft share	sum(cause_late_aircraft) filter (where cause_ok) / sum(cause_late_aircraft + cause_carrier + cause_nas + cause_weather + cause_security) filter (where cause_ok)	sum(cause_late_aircraft_sum) / (sum(cause_late_aircraft_sum) + sum(cause_carrier_sum) + sum(cause_nas_sum) + sum(cause_weather_sum) + sum(cause_security_sum))
reported carrier share	sum(cause_carrier) filter (where cause_ok) / sum(cause_late_aircraft + cause_carrier + cause_nas + cause_weather + cause_security) filter (where cause_ok)	sum(cause_carrier_sum) / (sum(cause_late_aircraft_sum) + sum(cause_carrier_sum) + sum(cause_nas_sum) + sum(cause_weather_sum) + sum(cause_security_sum))
reported nas share	sum(cause_nas) filter (where cause_ok) / sum(cause_late_aircraft + cause_carrier + cause_nas + cause_weather + cause_security) filter (where cause_ok)	sum(cause_nas_sum) / (sum(cause_late_aircraft_sum) + sum(cause_carrier_sum) + sum(cause_nas_sum) + sum(cause_weather_sum) + sum(cause_security_sum))
reported weather share	sum(cause_weather) filter (where cause_ok) / sum(cause_late_aircraft + cause_carrier + cause_nas + cause_weather + cause_security) filter (where cause_ok)	sum(cause_weather_sum) / (sum(cause_late_aircraft_sum) + sum(cause_carrier_sum) + sum(cause_nas_sum) + sum(cause_weather_sum) + sum(cause_security_sum))
reported security share	sum(cause_security) filter (where cause_ok) / sum(cause_late_aircraft + cause_carrier + cause_nas + cause_weather + cause_security) filter (where cause_ok)	sum(cause_security_sum) / (sum(cause_late_aircraft_sum) + sum(cause_carrier_sum) + sum(cause_nas_sum) + sum(cause_weather_sum) + sum(cause_security_sum))
inherited share	sum(inherited_minutes) / sum(arr_delay_pos)	sum(inherited_minutes_sum) / sum(arr_delay_pos_sum)

Source: BTS Reporting Carrier On-Time Performance (real); population: every flight in the window, through the metric layer; as of 2026-10-01.

The recovery study

Every estimator, run on simulated networks where the truth is known, 12 conditions with 20 seeds each, before it touched a real flight.