opentemporal/AirportDailyDelay

Airport Daily Delay

This dataset converts daily departure-delay rates and observed airport weather events into the format defined by the OpenTemporal data repository. It contains nine independent airport time series: SFO, SJC, OAK, JFK, LGA, EWR, LAX, DCA and IAD.

1. Source

Flight records come from the U.S. Bureau of Transportation Statistics Reporting Carrier On-Time Performance monthly files. Weather observations come from the Iowa Environmental Mesonet ASOS/METAR archive. Both are U.S. public-domain data; attribution to IEM is appreciated. The collected source files and source-level sequences are under BTS/AirportDelay/.

The flight measure covers reporting U.S. domestic carriers rather than every domestic and international flight at an airport. Fields used are FlightDate, Origin, DepDel15 and Cancelled. Weather fields used are observation time, sustained wind, gust, visibility, cloud layers and present-weather codes.

2. Value

One value is the share of valid observed departures at an airport on one local calendar day that departed at least 15 minutes late:

count(DepDel15 == 1) / count(DepDel15 in {0, 1})

Cancelled flights and flights without a valid DepDel15 are excluded from the denominator. For example, SFO on 2022-01-01 had 81 delayed departures among 275 valid observations, so its value is 0.29454545.

3. Multivariate time series

None. Each airport is one independent single-series group named departure delay rate. Weather events concern that airport only. The airport code is stored in metadata.json, not in dataset rows.

4. Grid, patch, example

The grid is daily from 2022-01-01 through 2026-08-31, 1,704 steps per airport. patch_len = 1, because the value itself is daily and observed weather can begin on any day. One example contains 28 patches, four weeks, and the stride is 7 patches, one week. With one series and one value per patch, patch_text_len = 2 and each example is 56 characters.

5. Normalization and clip

The 1st and 99th percentiles are computed once per airport over its entire 1,704-day stretch. quant_scale = 100.0. The normalized clip is 4, the largest integer strictly below 5 / sqrt(patch_len) for a one-step patch. No value is clipped. Across airports, the 1st percentile ranges from 0.028 to 0.067 and the 99th percentile from 0.392 to 0.670. The maximum decoded error in delay-rate units is 0.00153; exact per-airport quantiles and errors are stored in dataset/metadata.json.

6. Gaps

An isolated missing daily rate is forward filled. Longer gaps would remain missing and prevent affected examples from being built; the current source has no longer gap. Five values are filled across the nine series. The low-denominator rates at EWR and IAD are retained as source observations.

7. Events

The source has six detailed observed types. They are reduced to three short texts so each can be learned repeatedly:

detailed observation	event text
thunderstorm, snow or freezing rain	storm
sustained wind at least 15 m/s or gust at least 20 m/s	strong wind
visibility at most 1,000 m or ceiling at most 150 m	restricted visibility

Multiple source observations of one type on the same local day collapse to one event. When categories coincide, only one text is kept in the fixed priority storm, strong wind, restricted visibility; this avoids rare joined texts. An event is placed at the boundary before its local day, because the series value aggregates departures from that same day and the weather can affect them immediately.

Vocabulary counts in examples:

text	train	test	distinct source days
storm	4,030	699	1,248
strong wind	1,358	169	396
restricted visibility	3,558	419	1,031

Measured on distinct source days, the mean delay-rate change against the previous day is +0.076 for storm, +0.038 for strong wind, and +0.026 for restricted visibility; medians are +0.066, +0.026 and +0.016. The direction survives serialization unchanged because the build clips no values.

8. Split

The cut is 2026-01-01. Train examples end on or before the cut; test examples start on or after it. No example straddles the cut. Train starts are laid out backward from the cut so the final train example ends exactly there.

9. What the model sees

One row contains 28 days of one airport's departure-delay rate, normalized to that airport's own range. The model does not see the date, airport code, flight counts or weather magnitudes. It sees zero or more of the three event texts at daily patch boundaries and learns daily and weekly continuation and the change in delays around disruptive observed weather.

10. Output and figures

dataset/ is a Hugging Face DatasetDict with train and test. Every row has exactly serialized, patch_text_len, events and positions. dataset/metadata.json stores sources, parameters, quantiles, airport-to-row mapping, starts, event counts and measured event effects.

Current validated build:

	train	test
rows	1,845	279
examples per airport	205	31
rows with events	1,618 (88%)	222 (80%)
events per row, mean	4.85	4.61
most events in one row	18	15

The saved dataset and metadata occupy about 0.5 MB.

11. Pipeline and open points

Run from this folder with the repository environment and the included tokenizer wheel on PYTHONPATH:

PYTHONPATH=tstokenizer-2.2.0-py2.py3-none-any.whl conda run -n base python build.py

The script reads the collected daily and event sequences, fills isolated gaps, serializes each airport, attaches reduced daily weather events, validates every row, writes dataset/, and reopens it for a final contract check.

Open points: revisit the 28-day window, weekly stride, event reduction and contemporaneous event placement after a first training run.