TAF statistics: groups, change indicators and validity
This study counts how often change groups and other elements appear in TAFs, using snapshots of the NOAA Aviation Weather Center TAF cache. Identical forecast texts are counted once, elements are found with text patterns, and each share is the number of matching forecasts divided by all distinct forecasts in the sample.
Key points
- TAF snapshots are downloaded less often than METAR snapshots, and a forecast text seen in several snapshots is counted once.
- Change groups (TEMPO, BECMG, FM, PROB30, PROB40) are found with text patterns, so each share is the proportion of forecasts that contain at least one.
- Validity length is computed from the first day-hour pair of each forecast; the month change is approximated and values above 48 hours are ignored.
- The sample leans toward North America and mixes standard and amended forecasts, so it is not a worldwide picture.
- The tables describe a data sample, not an operational product.

Dataset: 121,001 METARs from 5,526 stations and 12,960 TAFs, NOAA Aviation Weather Center, 18 November 2026 to 19 November 2026. Generated 2026-11-19.
| Feature | Share | TAF |
|---|---|---|
| TAFs containing at least one TEMPO group | 52.1% | 6,748 |
| TAFs containing at least one BECMG group | 37% | 4,793 |
| TAFs containing at least one FM group | 31.2% | 4,044 |
| TAFs with forecast maximum temperature (TX) | 19.2% | 2,491 |
| TAFs forecasting cumulonimbus (CB) | 19.1% | 2,475 |
| TAFs containing PROB30 TEMPO or PROB40 TEMPO | 16.3% | 2,109 |
| TAFs containing PROB30 | 15.9% | 2,063 |
| TAFs containing CAVOK | 11.3% | 1,464 |
| TAFs forecasting thunderstorms (TS) | 11.1% | 1,439 |
| TAFs containing PROB40 | 10.2% | 1,328 |
| TAFs that are amendments (AMD) | 7.4% | 959 |
| TAFs containing NSW | 5.5% | 707 |
| TAFs with a forecast QNH group (QNH....INS) | 3.8% | 489 |
| TAFs with a low-level wind shear group (WS) | 2.5% | 323 |
| TAFs that are corrections (COR) | 1.1% | 140 |
| TAFs that are cancellations (CNL) | 0% | 0 |
| Validity | Share | TAF |
|---|---|---|
| 24 h | 53.3% | 6,901 |
| 9 h | 15.6% | 2,021 |
| 30 h | 14.6% | 1,891 |
| 12 h | 5.9% | 766 |
| 23 h | 1.2% | 155 |
| 22 h | 1% | 133 |
| 8 h | 1% | 131 |
| 21 h | 0.7% | 97 |
What does the TAF study measure?
The TAF study measures how often specific groups and indicators appear in a sample of TAF texts and how long those forecasts are valid. It counts strings. It does not evaluate whether a forecast was right, and it does not rank airports or forecasters. The tables on this page are generated from the same dataset as the METAR study, so they always match the method described here.
The first table lists each tracked element with the share of forecasts that contain it and the raw count. The second lists the most frequent validity lengths in hours. For a decoding of each group, start from how to read a TAF. The companion study on observations is on the METAR statistics page.
Where do the forecasts come from?
The forecasts come from the TAF cache file of the NOAA Aviation Weather Center, an XML file that its data page lists as all current TAFs, refreshed once a minute. A collection script downloads that file on the first iteration of its hourly loop and then on every sixth iteration, so about every six hours. Each download is saved unchanged with a timestamp in its name, and a failed download is logged and not retried.
Because a TAF stays valid for many hours, six hours between snapshots still captures most forecasts, but a forecast that is issued and replaced between two downloads can be missed. The page on METAR and TAF data sources and APIs describes the cache and the other ways to obtain the same data.
How are duplicate forecasts removed?
Duplicates are removed by comparing the text. The analysis script extracts each forecast from the XML with a regular expression that captures the raw text field and the station identifier, then collapses all runs of whitespace to a single space. The normalized text is used as the key of a map, so a forecast that is still current in the next snapshot is counted once.
Two consequences follow:
- An amendment is a different text from the forecast it replaces, so both can be present in the sample. The same applies to a correction. The amendment, correction and cancellation page explains these indicators.
- The count is a count of distinct forecast texts, not of station-days. An airport that issues a forecast every three hours contributes more texts than one that issues it every six hours.
How are TAF elements detected?
Each element is detected with a regular expression applied to the whole normalized text, since a TAF has no remarks section comparable to the METAR one. The table below lists the families of patterns.
| Element | What a match means |
|---|---|
| TEMPO | The token TEMPO appears as a separate group |
| BECMG | The token BECMG appears as a separate group |
| FM | The letters FM followed by six digits (day, hour, minute) |
| PROB30, PROB40 | The token appears as a separate group |
| PROB with TEMPO | A probability indicator immediately followed by TEMPO |
| AMD, COR, CNL | The token appears as a separate group |
| NSW | The token appears as a separate group |
| TX or TN group | A maximum temperature group, as described on the temperature groups page |
| Wind shear | The letters WS followed by a height and a wind, as on the wind shear page |
| CAVOK | The token appears as a separate group |
| CB | CB preceded by three digits (a cloud layer with cumulonimbus) |
| Thunderstorm | TS, with optional intensity sign and VC, as on the thunderstorm page |
| Altimeter group | QNH followed by four digits and INS, a pattern used in some national TAFs |
A share is the number of forecasts that contain the element at least once, divided by the number of distinct forecasts. A forecast with three TEMPO groups counts once. For that reason the shares answer the question "how many forecasts use this", not "how many times is it used".
What do the change groups mean for the counts?
The change groups are counted as presence indicators, not as numbers of groups. Their meanings differ: TEMPO describes fluctuations that last less than an hour each time and less than half of the period, BECMG describes a lasting change within a period, and FM starts a new self-contained segment at a stated time. The page on TEMPO, BECMG and FM compared gives the rules and examples.
PROB30 and PROB40 express a probability of 30 or 40 percent for an event, and are used alone or with TEMPO. Two rows track them separately, with a third for the combined form. A high share for one of these rows reflects the practice of the forecasting offices in the sample, which depends on national rules, so no comparison between countries should be drawn without regional detail. The sample's regional weight is described on the METAR statistics page.
How is validity length computed?
Validity length is computed from the first pair of day and hour groups in the forecast, written DDHH/DDHH. The script reads the start day and hour and the end day and hour, then calculates the duration in hours as the difference in days times 24 plus the difference in hours.
Two adjustments apply:
- If the result is zero or negative, which happens when the validity period crosses a month change, 720 hours are added. This assumes a 30-day month, so the result is approximate in months of 28, 29 or 31 days.
- Any result above 48 hours is ignored, since it signals a pattern that matched something else than a validity period.
The table shows the eight most frequent lengths. The share is relative to all valid lengths, not to all forecasts. Typical values such as 9, 12, 24 or 30 hours depend on national rules, as the validity period page explains.
How should the tables be read?
The tables should be read as frequencies within the sample. Four points help:
- A row is not a rate of use per airport. Large airports issue more forecasts, so they weigh more than small ones.
- Items typical of some countries, such as the QNH group or TX and TN groups, have low shares because few countries use them.
- Optional groups, such as the icing and turbulence groups described on the military TAF groups page, are not tracked in the tables.
- A short collection period mixes only the forecasts valid during that time, so seasonal patterns are not visible.
What are the main caveats?
The main caveat is coverage. The cache holds the forecasts that NOAA receives and makes available, and North American stations weigh heavily. A second caveat is the text-based method: a regular expression cannot know whether a token is a group or part of free text, and a malformed forecast can be counted wrongly. A third caveat is timing: forecasts are valid for a period, so a forecast that appears in a snapshot may be partly outdated at the time of analysis.
A schematic report
A schematic TAF shows which elements the patterns find. The example below is fictitious: the station is real, the values are not.
TAF EGLL 191100Z 1912/2018 24012KT 9999 SCT025
BECMG 1916/1918 27008KT
TEMPO 1920/2006 4000 SHRA BKN012CB
PROB30 TEMPO 2006/2012 3000 BR
FM201200 20005KT CAVOK
For this text the validity pattern reads 1912/2018, which gives 30 hours (day 20 minus day 19, times 24, plus 18 minus 12). The text contains BECMG, TEMPO, PROB30 and the combined PROB with TEMPO, FM followed by six digits, CAVOK and a CB layer. It contains no AMD, COR, CNL, NSW or TX groups. Each row of the table would count this forecast once, whatever the number of groups.
In real reports
In real reports, forecasts differ widely in length and in structure. A short TAF for a small airfield can have a single line and no change group, while a long one for a major airport can have five or six segments. Some countries add national groups, temperature groups, or text that the patterns were not written for. Amended forecasts repeat the full text with a new issue time, so a single forecast period can leave several distinct texts in the sample.
The tables also reflect operating practice. Validity length, the use of probability groups and optional groups depend on national rules and on the rules of each forecasting office. For the origin of those rules, see the regulations page.
Method and limits
This section summarizes the method and its limits so that the tables can be checked and reproduced.
- Source: TAF cache of the NOAA Aviation Weather Center (XML, gzip), described on its data page as all current TAFs updated once a minute. The file is downloaded on the first hourly iteration and every sixth one after it, then saved unchanged.
- Reproduction: run the collection script, then the analysis script, which reads every saved file and writes the data file used by the tables. Both are in the repository of this site.
- Deduplication: by normalized text, with whitespace collapsed. Forecasts without a station and a raw text field are skipped.
- Detection: regular expressions on the whole text. A share is the number of forecasts that contain the element at least once.
- Validity: first
DDHH/DDHHmatch; the duration is computed from days and hours, 720 hours are added when the result is not positive, and values above 48 hours are dropped. - Not represented: the sample is not representative of worldwide practice. It favors North America, depends on the collection period and misses forecasts issued and replaced between two downloads.
- Sources read: the Weather Center data API page for the cache files, WMO-No. 782 (2025) for the code, and the NWS and FAA material listed above for US practice.
This page describes a data sample. It is not an operational source and must not be used for flight planning or any safety decision.
Training and reference only. For flight planning, always use official briefing sources (your national weather service, AIS or approved provider).
Frequently asked questions
Where do the TAF figures on this site come from?
They come from the TAF cache file of the NOAA Aviation Weather Center, an XML file that its documentation describes as all current TAFs, refreshed once a minute. A script downloads it every sixth hour of the collection loop and stores it unchanged. The analysis then reads the saved files and counts distinct forecast texts.
Why is the TAF sample smaller than the METAR sample?
A TAF is issued only a few times a day, and each one stays valid for many hours, while a METAR is issued every hour or half hour. The TAF cache is also downloaded less often, so each distinct forecast appears only a limited number of times. A smaller sample means the shares move more when new data arrive.
Does a high share of TEMPO mean forecasters rely on TEMPO more than on BECMG?
Not necessarily. The table counts forecasts that contain at least one group of a given type, not how many groups each forecast holds. Local practice, airport size and forecast length all influence the result. The page on TEMPO, BECMG and FM explains how the three indicators differ in meaning.
Can the validity table be used to check the length of a given TAF?
No. The validity table is built from the first day and hour pair found in each forecast and approximated at month changes. It describes the distribution of lengths in a sample. To read the validity of a real TAF, use the decoding rules and the current official text.
Sources
- NOAA Aviation Weather Center. Data API and cache files (METAR, TAF and other products)
- World Meteorological Organization (2025). Aerodrome Reports and Forecasts: A Users' Handbook to the Codes (WMO-No. 782), 2025 edition
- Federal Aviation Administration. Aeronautical Information Manual, Chapter 7, Section 1: Meteorology
- National Weather Service. NWS Instruction 10-813, Terminal Aerodrome Forecasts