Every US flight since 2003

150+ million scheduled US domestic flights, 2003–2025, cleaned and repacked. Free to take.

What this is

The US DOT Bureau of Transportation Statistics "Reporting Carrier On-Time Performance" table — the record airlines file with the federal government every month. It's a US government work, in the public domain. We cleaned it, kept the columns the show uses, and repacked it into one zstd-compressed parquet file per year, because the raw BTS download is hundreds of zipped CSVs.

What's in it

23 parquet files, one per year from 2003 to 2025, about 1.7 GB total (roughly 51–91 MB each). Columns include flight date, carrier, flight number, origin and destination airport, scheduled and actual departure/arrival times, delay minutes, cancellation and diversion flags, and the five official delay-cause buckets: carrier, weather, NAS, security, and late-arriving aircraft. One caveat from the fine print: reporting only covers carriers with at least 0.5% of domestic scheduled revenue, so small regional operators are absent.

What the show did with it

EP 01 was built on this table: the day replay, the storm meltdown, and the airport cheat sheets all run on these records, and the five delay-cause buckets are what the episode's argument rests on. The built artifacts live on the receipts page; the raw material is here.

Query it yourself

DuckDB reads parquet straight off the URL — no download needed. The ten most late-departing origins of 2025:

SELECT origin, round(100.0*avg((dep_delay>=15)::int),1) AS pct_late
FROM read_parquet('https://data.mathvsvibes.com/bts/bts_2025.parquet')
WHERE cancelled=0 GROUP BY origin ORDER BY pct_late DESC LIMIT 10;

Dataset post · dataset published 2026-07-07, posted 2026-08-06. Twenty-three years of flight data, give or take a pandemic.