How frost-date probabilities work

Two dates, two directions

NOAA’s National Centers for Environmental Information publish frost-freeze probability dates in the 1991–2020 U.S. Climate Normals. The calculation methodology for those normals says the dates represent the likelihood that a given minimum temperature will be experienced before or after a given date. For the beginning of the growing season, that is the probability the last spring freeze occurs on or after the specified date. For the end of the growing season, that is the probability the first autumn freeze occurs on or before the specified date. The thresholds include 16°F, 20°F, 24°F, 28°F, 32°F, and 36°F. The probability levels run from 10% to 90% in steps of 10. This site shows 28°F, 32°F, and 36°F at 10%, 50%, and 90%, and it uses the 32°F set as the one required for a station to count as qualified.

The column names in the station file look symmetrical and they are not. ANN-TMIN-PRBLST-T32FP10 is the last-spring 32°F date at the 10% level. ANN-TMIN-PRBLST-T32FP90 is the 90% level. Because the probability is “on or after,” the 10% date falls later in the spring and the 90% date falls earlier. ANN-TMIN-PRBFST is the first autumn date, and the probability is “on or before,” so the 10% date is the earlier autumn date and the 90% date is the later one. If a website sorts these columns from “earliest” to “latest” without saying which probability sits on which side, it will swap the risk. The tables on this site label the direction in the column header.

Austin-Bergstrom, as a check

The station file for Austin-Bergstrom International Airport, USW00013904, is the worked example. For 32°F, the last-spring dates are February 20 at 90%, March 12 at 50%, and April 2 at 10%. Read that in the words of the methodology. There is a 90% chance the last spring minimum of 32°F or lower occurs on or after February 20. There is a 50% chance it occurs on or after March 12. There is a 10% chance it occurs on or after April 2. February 20 is not a safe planting date. It is the early end of the distribution: in most years the last freeze is still ahead of it. April 2 is the cautious end: a freeze that late or later is the uncommon case, about one year in ten under the normal. March 12 is the median. Half the probability sits on either side of it.

The first autumn 32°F dates at the same station are November 1 at 10%, November 16 at 50%, and December 5 at 90%. There is a 10% chance the first autumn 32°F occurs on or before November 1. There is a 50% chance it occurs on or before November 16. There is a 90% chance it occurs on or before December 5. An early freeze, on or before November 1, is the uncommon case. A first freeze by December 5 is the common case. If you are counting a fall crop back from “the first frost,” you have to say which of those three dates you mean. This site’s fall calculator makes you pick 10, 50, or 90. It does not hide the choice inside the word frost.

The growing-season lengths in the same file, ANN-TMIN-PRBGSL-T32, are 279 days at 10%, 252 days at 50%, and 223 days at 90%. The methodology says these are the lengths the growing season meets or exceeds at that probability. A 10% season length is a long season: only about one year in ten is that long or longer. A 90% season length is a shorter number that the season reaches or beats in about nine years out of ten. The 252-day figure is the median length, and it is the one people usually quote. The length on the page is NOAA’s published value. Subtracting March 12 from November 16 by hand is a different statistic.

Denver International Airport, USW00003017, shows why you cannot carry Austin’s calendar north. In the same product the last 32°F dates there are April 23 at 90%, May 5 at 50%, and May 19 at 10%. The first 32°F dates are September 27 at 10%, October 9 at 50%, and October 20 at 90%. The 50% growing-season length is 156 days. Houston Intercontinental, USW00012960, sits the other way: last 32°F on January 12 at 90%, February 13 at 50%, and March 12 at 10%; first 32°F on November 17 at 10%, December 9 at 50%, and January 11 at 90%; a 50% season length of 300 days. A city page does not use these stations unless they win the distance test. They are here so the column meanings can be checked against three climates. The nearest-station rule is on the methodology page.

Conditional dates, missing dates, and flags

The methodology says the computation is a conditional probability. That matters in warm climates where a 32°F day does not happen every year. The dates describe the freeze given that the temperature is reached at least once. If the probability of occurrence is under 10%, NCEI stores the special value −6666 and does not offer a date. At the other extreme, some northern or high stations have a chance of the threshold on every day of the year. Those can be stored as −4444. The file readme uses measurement flag V for year-round risk that is “too cold to compute,” flag M for missing, flag Y when there are not enough values, and flag Z when the result contradicted another value. This site drops any station whose 32°F dates carry those flags.

Completeness is a separate flag. S means the normal meets World Meteorological Organization standards with 24 or more years in the 30-year period. R means representative, with 10 or more years. P is provisional. E is estimated from neighbors. A qualified station on this site is S or R on all six 32°F probability dates, with at least 10 years used, and with the dates in the seasonal order the definitions require. The page prints the flag and the year count. If the annual probability of at least one day at or below 32°F is present and itself passes the same flag test, the page prints that percentage too, so a conditional date is not mistaken for a date that happens every year.

The cold season used for these normals runs from August 1 to July 30, which is why a “spring” date at a warm station can fall in January and a “fall” date can fall in December or even January. The words spring and autumn in the documentation are loose on purpose. The dates are kept in that August-to-July order. A December date is earlier in the cold season than a February date, even though December comes later on a January-to-December calendar.

What the probabilities are not

They are not a forecast for the coming season. A strong El Niño year, a late Arctic front, or a mild March can sit outside the 10-to-90 band. The band describes the 1991–2020 distribution at that station after NCEI’s bootstrapped simulations, which the methodology says were used because raw first and last dates jump around from year to year. They are not the temperature in your yard. The station may be at an airport, on a different elevation, and several miles from the bed. The page prints the miles for that reason. They are not the same thing as a National Weather Service frost advisory. Those products are defined in the frost, freeze, and hard freeze guide. NCEI’s 36°F, 32°F, and 28°F rows are thresholds in a climate normal. The Weather Service’s “hard freeze” is a forecast category tied to 28°F. The numbers can line up. The products should stay in separate sentences. They are not one definition.

They are also not a planting date by themselves. An extension guide that says “plant tomatoes one to eight weeks after the frost-free date” has not said whether frost-free means the 50% date or the 10% date. This site applies that kind of instruction to the median and says so, and it leaves the 10% and 90% dates on the page so you can see the spread. If you want the cautious tomato date at Bergstrom, you look at April 2, the 10% last-32°F date, and you still read the extension window rather than treating April 2 as a rule the guide did not write.

Sources

• NOAA NCEI, Normals Calculation Methodology 2020, frost-freeze probability dates and growing-season length: https://www.ncei.noaa.gov/data/normals-daily/1991-2020/doc/Normals_Calculation_Methodology_2020.pdf
• NOAA NCEI, Readme for by-station normals files, including PRBLST, PRBFST, PRBGSL, and measurement flags: https://www.ncei.noaa.gov/data/normals-annualseasonal/1991-2020/doc/Readme_By-Variable_By-Station_Normals_Files.txt
• Station access files, 1991–2020 annual/seasonal normals, including USW00013904, USW00003017, and USW00012960: https://www.ncei.noaa.gov/data/normals-annualseasonal/1991-2020/access/
• Product overview: https://www.ncei.noaa.gov/products/land-based-station/us-climate-normals