Bts Airline On Time Performance Data: Dataset Structure and Field Coverage

BTS's TranStats portal indexes 21 aviation databases, and the Airline On-Time Performance collection is the one built for reliability questions. Every record is one flight: the tables log a row per departure for certificated U.S. carriers across the United States domestic network, refreshed monthly and reaching back to October 1987. At roughly 5-7 million flight rows per year, a single decade of history means tens of millions of rows to work with.

Two properties separate this record from every mirror of it. It is the reporting source: carriers file to BTS under Form 234, so downstream copies inherit whatever the official tables say rather than correcting them. And it is the only version that keeps growing - a mirror captures one year and stops, while this archive adds another few million rows every month.

Yes. The Kaggle mirror 2015 Flight Delays and Cancellations (5.8M U.S. Flights), published by the USDOT organization under commercial delivery terms, compresses calendar year 2015 into three files totalling about 200 MB zipped: flights.csv with 5,819,079 rows (about 586 MB uncompressed), airlines.csv holding all 14 reporting carriers in 359 bytes, and airports.csv listing 322 airports with coordinates in roughly 24 KB. The underlying records originate from BTS Form 234 on-time reporting, so the mirror is the official stream in a friendlier wrapper rather than a different dataset.

Each flights.csv row carries what delay modellers reach for: airline, flight number, origin and destination, scheduled and actual times, delay minutes, and cancellation and diversion flags. With a carrier table measured in bytes and airports joined on coordinates, feature engineering starts immediately - group by airline, airport pair or hour and the labelled benchmark is ready without touching the portal. The trade-off is temporal: one calendar year, frozen at its February 2017 publication, so no post-2015 aircraft, route or schedule changes appear anywhere in it. Prototype on it, then extend back through the official archive when a model needs more than one year of seasonal patterns.

A repeatable build, using only sources named above: