Donny-Guo/airline-delay

U.S. Airline On-Time Performance and Delay Causes, 2009-2018

This dataset contains backend-ready Parquet data derived exclusively from the 120 monthly Reporting Carrier On-Time Performance archives published by the U.S. Department of Transportation, Bureau of Transportation Statistics (BTS), for January 2009 through December 2018.

It provides two configurations:

• aggregate (default): monthly reporting-carrier and destination-airport aggregates for dashboards, APIs, and analytical queries.
• flight_detail: one enriched record per official source row, partitioned by flight year and month for drill-down and audit use.

This is a derived community dataset, not an official BTS publication. BTS is the authority for the underlying flight records.

Dataset summary

Item	Value
Coverage	2009-01 through 2018-12
Calendar months	120
Official flight-detail rows	61,589,291
Aggregate rows	153,587
Aggregate columns	21
Detail columns when read as a Hive dataset	41
Detail Parquet files	120
Flights arriving at least 15 minutes late	11,437,324
Records marked cancelled	975,445
Records marked diverted	145,675

No passenger or traveler-level personal information is included.

Repository layout

The expected Hub layout is:

README.md
data/
├── aggregate/
│   └── airline_delay_causes_2009_2018.parquet
└── detail/
    ├── year=2009/
    │   ├── month=1/
    │   │   └── *.parquet
    │   └── ...
    └── year=2018/
        └── month=12/
            └── *.parquet
docs/
├── DATA_DICTIONARY.md
├── FRONTEND_DATA_QUALITY_GUIDE.md
└── OFFICIAL_DATA_AUDIT.md
metadata/
├── data_quality.json
├── official_all_months_audit.json
└── official_sources.json

Source and provenance

The sole fact-data authority is the BTS TranStats PREZIP archive. Each of the 120 official monthly ZIP files was preserved locally and recorded in a source catalog with its official URL, archive member, byte size, CRC, SHA-256 digest, normalized-file SHA-256 digest, and row count.

The production process:

1. selects the 27 flight fields used by this dataset from each official archive;
2. applies explicit production types without imputing missing operational data;
3. enriches carrier and origin/destination airport codes using historical, date-effective BTS dimension tables;
4. preserves all official rows without candidate-key deduplication;
5. writes monthly Hive-partitioned detail Parquet;
6. calculates the monthly carrier/destination aggregate; and
7. validates the published files before release.

Carrier and airport names are historical values effective on the flight date, not names normalized to the present day.

Configurations

aggregate (default)

One row represents one:

year + month + reporting carrier + destination airport

The natural key is (year, month, carrier, airport).

Column	Type	Meaning
year	int16	Flight calendar year.
month	int8	Flight calendar month, 1-12.
carrier	string	BTS reporting-carrier code.
carrier_name	string	Historical carrier name effective during the period.
airport	string	Destination airport code.
airport_name	string	Historical destination label, formatted as city: airport name.
arr_flights	int64	All official records in the group, including cancelled and diverted records.
arr_del15	int64	Records with arr_delay >= 15.
carrier_ct	double	Fractional delayed-flight equivalents attributed to the carrier.
weather_ct	double	Fractional delayed-flight equivalents attributed to extreme weather.
nas_ct	double	Fractional delayed-flight equivalents attributed to the National Aviation System.
security_ct	double	Fractional delayed-flight equivalents attributed to security.
late_aircraft_ct	double	Fractional delayed-flight equivalents attributed to a late-arriving aircraft.
arr_cancelled	int64	Records marked cancelled.
arr_diverted	int64	Records marked diverted.
arr_delay	int64	Arrival-delay minutes summed over records delayed at least 15 minutes.
carrier_delay	int64	Carrier-cause minutes for records delayed at least 15 minutes.
weather_delay	int64	Extreme-weather minutes for records delayed at least 15 minutes.
nas_delay	int64	National Aviation System minutes for records delayed at least 15 minutes.
security_delay	int64	Security minutes for records delayed at least 15 minutes.
late_aircraft_delay	int64	Late-arriving-aircraft minutes for records delayed at least 15 minutes.

The five *_ct values are fractional attributed-flight equivalents, not independent integer flight counts. For one flight delayed 100 minutes, with 60 carrier minutes and 40 NAS minutes, the contributions are carrier_ct=0.6 and nas_ct=0.4. For a group of delayed flights, the exact definition is:

cause_ct = round_half_up(sum(cause_minutes_i / arr_delay_i), 2)

This is not the same as multiplying arr_del15 by the aggregate delay-minute share sum(cause_minutes) / sum(arr_delay). The first measure gives every delayed flight a total weight of one before splitting it among causes; the second weights longer delays more heavily. For 2018-01 / DL / ATL, the carrier minute share is 27.3250%, while the carrier flight-equivalent share is 23.5204% and carrier_ct=626.11.

Each *_ct value is independently rounded to two decimals. The build uses an exact rational fallback at half-cent boundaries so results such as 2.225 are deterministically stored as 2.23, independent of binary-float summation order.

The BTS definitions for the five broad delay-cause categories are documented in Understanding the Reporting of Causes of Flight Delays and Cancellations.

flight_detail

One row represents one official BTS source row enriched with historical carrier and airport attributes. It includes:

• flight identity and route;
• scheduled and actual departure/arrival times;
• departure, arrival, taxi, elapsed, airborne, and distance measures;
• cancellation and diversion status;
• the five BTS delay-cause minute fields;
• date-effective carrier metadata;
• date-effective origin and destination airport metadata; and
• Hive-derived year and month partition columns when read as a dataset.

See docs/DATA_DICTIONARY.md for the complete detail schema and types.

Important detail-partition behavior

The detail layout uses Hive directory partitioning. For example:

data/detail/year=2018/month=1/part-....parquet

The individual part-....parquet file physically stores 39 columns, including fl_date, but does not repeat year and month inside every row. The two values are encoded once in the directory path. A Hive-aware dataset reader turns the directory labels into two virtual columns, producing the documented 41-column detail dataset.

Consequences:

• Reading the whole directory with Hive partitioning exposes year and month.
• Opening one monthly Parquet file directly exposes 39 physical columns and does not expose separate year and month columns.
• This is a storage-layout property, not missing flight data. fl_date remains a physical column in every detail file.
• Hub consumers should verify whether their selected reader and the Dataset Viewer materialize Hive partition columns. If they do not, consumers can use fl_date, or the publisher can release a repackaged copy with physical year/month columns.

PyArrow example after downloading the repository:

import pyarrow.dataset as ds

detail = ds.dataset(
    "data/detail",
    format="parquet",
    partitioning="hive",
)

january_2018 = detail.to_table(
    filter=(ds.field("year") == 2018) & (ds.field("month") == 1),
)

DuckDB example after downloading the repository:

SELECT *
FROM read_parquet(
    'data/detail/**/*.parquet',
    hive_partitioning = true,
    union_by_name = true
)
WHERE year = 2018 AND month = 1;

Loading from Hugging Face

Replace YOUR_NAMESPACE/YOUR_DATASET with the final Hub repository ID.

Load the default aggregate configuration:

from datasets import load_dataset

aggregate = load_dataset(
    "YOUR_NAMESPACE/YOUR_DATASET",
    "aggregate",
    split="full",
)

assert aggregate.num_rows == 153_587

Stream flight detail without downloading all rows first:

from datasets import load_dataset

detail = load_dataset(
    "YOUR_NAMESPACE/YOUR_DATASET",
    "flight_detail",
    split="full",
    streaming=True,
)

first_record = next(iter(detail))

For backend dashboards and common trend queries, prefer aggregate. Use flight_detail for drill-down, custom metrics, validation, or audit workflows.

Integrity validation

This release was compared with all 120 locally preserved official monthly ZIP archives, not a sample. The audit extracted each original official CSV, applied the same production types, and compared all 27 production business fields plus derived year and month with the final detail Parquet using bidirectional EXCEPT ALL semantics.

Validation	Result
Months compared	120 / 120
Official rows compared	61,589,291
Local detail rows compared	61,589,291
Official rows missing locally	0
Local rows absent from official data	0
Exact duplicate rows across the normalized business fields	0
Unmatched carrier or airport dimensions	0
Missing core grouping keys	0

The complete machine-readable evidence is provided in metadata/official_all_months_audit.json.

Known source-data limitations

Exact equality with the source does not imply that every official field is complete. The following source conditions are deliberately preserved:

• 2 rows are missing at least one scheduled-time field;
• 640 non-cancelled rows have no reported departure-delay value;
• 351 completed rows have no reported arrival-delay value;
• 1 Frontier Airlines flight on 2017-08-22 reports 194 arrival-delay minutes but has no minutes assigned to any of the five BTS delay-cause fields;
• 4 narrower flight-candidate-key groups contain two different official records;
• 13 rows are marked both cancelled and diverted, a reportable operational combination rather than mutually exclusive status flags.

No missing value is imputed as zero. No official row is removed because of a candidate-key collision.

Important analytical consequences:

• A missing delay value means unreported or unknown, not zero and not on time.
• arr_cancelled and arr_diverted are independent counts and must not be added as if they were mutually exclusive.
• Do not derive an exact on-time count or on-time rate from the current aggregate alone; the aggregate does not yet expose the unknown-arrival count or the cancelled/diverted intersection required for a correct denominator.
• For delay-cause minute charts, compute the residual arr_delay - sum(five cause-minute columns) and label a positive residual Unreported cause. Do not redistribute it across the five reported causes.
• Do not infer an unknown cause-flight count from rounded *_ct columns.
• Do not pair a cause-minute percentage with a *_ct value as if one could be calculated from the other. For minute charts use sum(*_delay) / sum(arr_delay); for flight-equivalent charts use sum(*_ct) / sum(arr_del15).

See docs/FRONTEND_DATA_QUALITY_GUIDE.md for display rules and metric denominators, and metadata/data_quality.json for the complete machine-readable quality profile.

Intended uses

Suitable uses include:

• airline and airport delay dashboards;
• backend APIs for monthly delay metrics;
• carrier, airport, route, and temporal trend analysis;
• reproducible examples using DuckDB, PyArrow, Polars, pandas, or Hugging Face Datasets; and
• source-to-output audit and data-engineering demonstrations.

Users should not treat the five broad delay categories as exact incident-level root-cause narratives. They are reporting categories defined by BTS.

Attribution, reuse, and disclaimer

Please cite the U.S. Department of Transportation, Bureau of Transportation Statistics as the authority for the underlying records and link to the BTS TranStats PREZIP archive.

This derived dataset is provided without endorsement by BTS or the U.S. Department of Transportation. Users are responsible for validating that its scope and documented limitations are appropriate for their use case.