How Wrong Is Your Nearest Weather Station About Frost? We Measured It

The finding

Nearly every frost date tool answers the same way: it finds the weather station nearest to you and hands you that station's dates. frostdatesbyzipcode.com says so plainly — "For each ZIP code, we find the nearest NOAA weather station within 100 miles using a spatial grid lookup."

That is a reasonable thing to do, and to their credit they do not hide it. The same page warns that "actual frost dates can vary by 2–4 weeks in either direction depending on weather patterns," and it names the culprit exactly: "microclimates (elevation, urban heat islands, proximity to water) can also shift frost dates significantly."

But read that first warning again — it is about weather, about how one year differs from the thirty-year average. It is not a statement about how wrong the station substitution itself is. That second error is a different thing and it is present in every answer the method gives, in an average year as much as a strange one.

Climate scientists have studied this kind of spatial error for decades, so we are not claiming to have discovered it. What we could not find is any frost date calculator that publishes an accuracy figure for its own lookup — a number telling you how far off the answer on the screen is likely to be. The problem gets named in a caveat. It does not get measured, and it never reaches the gardener as a number. So we measured it.

Read the middle row again. Roughly one station in seven disagrees with its own nearest neighbor by two weeks or more about when the last frost comes. If a website hands you a neighboring station's date, that is your chance of being two weeks out before anything else has gone wrong.

And it is a floor, not a ceiling. We compared stations to other stations, which is the friendliest possible version of the test — both ends are real thermometers with thirty years of record. Your garden is not a weather station. It sits somewhere in between, and nothing measures it.

Being closer barely helps

The obvious fix is a nearer station. We checked whether that works, and it mostly does not.

The median never moves off 5 days. Cutting the radius from a hundred miles to ten takes the two-week failure rate from 13.8% only as far as 13.1%.

Be careful how hard you lean on that table, and we would rather say so than let someone else catch it: US weather stations are mostly clustered, so 2,966 of our 6,191 pairs are already within ten miles of each other and the bottom row is nearly the same set as the top one. The rows are not independent samples. What the table does support is the modest claim, and it is the one that matters: getting a closer station buys you much less than you would expect. A station ten miles away still has better than a one-in-eight chance of being two weeks wrong about your frost.

What ten miles can hide

Averages undersell this. Here are the closest pairs of stations in the country that disagree the most — every one of them near enough that any calculator would happily hand you either answer.

The Grand Canyon pair is the picture of the problem. Bright Angel sits on the rim, Phantom Ranch sits on the canyon floor a mile below it, they are 7.7 miles apart, and NOAA's own record puts their last frost dates 114 days apart. Same canyon. A season and a half between them.

"But I live somewhere flat"

This is the fair objection, and it is the one we most wanted to be true, because if the problem were only mountains then most gardeners could ignore it. It is not only mountains.

Below are the worst disagreements in the country between stations that are close together and effectively at the same elevation — under 60 meters of height between them, so nothing here is a peak-versus-valley trick.

Sullivans IS, SC and Charleston City, SC sit 4.5 miles apart with 2 meters of height between them, and their last frost dates are 37 days apart. That is a barrier island and the city across the water from it — no hills involved, and far from the mountain west.

Closer still: Beaverton 2 SSW, OR and Nature Park-Beaverton, OR are 3.1 miles apart — the same town — and disagree by 49 days. You could walk between them in an hour. A calculator would hand you whichever one it found and call it your answer.

Elevation is the loudest cause but it is not the only one. Cold air is heavy — it runs downhill at night and pools in low ground, so a dip of a few feet can hold frost that the rise beside it never sees. Water moderates: several pairs above are a coastal town against a town a few miles inland, and the ocean keeps one of them frost-free for weeks longer. None of that is visible to a lookup that knows only which station is closest.

The second problem: your zone is answering a different question

The most common way to get a planting date is to look up your USDA hardiness zone. It is also a category error, and it is worth being precise about why.

A hardiness zone is defined by the average annual extreme minimum winter temperature. That is it. It is a measure of how cold your worst winter night usually gets, and it exists to answer one question: which perennial plants will survive here. It is genuinely good at that job.

It says nothing about spring. Two gardens can share a zone — same worst winter night — and part company completely on when that last frost lets go, because how cold your coldest night gets and how long your cold season runs are different things. So when a chart says "Zone 7: plant tomatoes April 15," the zone did not supply that date. Somebody attached a spring calendar to a winter measurement, and the join is where the error gets in.

Both problems have the same shape. A wide-area average is true about the area and wrong about the spot, and the gardener is standing on the spot.

What we do instead, and where we lose

We do not look up a station. We take 26 years of daily low temperatures for your actual location, count the real last-frost day in each of those years, and read the odds straight off that count — corrected for your ground's true elevation at about 250 meters of resolution. The method, the validation and the failures are all on our frost methodology page, and the station file this study is built from is a free download.

Measured against 7,006 held-out NOAA stations, that lands at about 6 days of median-date error nationally, against 7.2 days for a nearest-station lookup. Where terrain is complex and stations are sparse — the rural US, mountain valleys, the places this whole page is about — the gap widens to 8.5 days against 13.1.

Now the part the rest of this category leaves out. In the South, our surface measured 0.21 days behind a plain nearest-station lookup. On flat ground with a good station nearby, a station lookup is a strong opponent and we do not beat it. Along marginal-frost stretches of the Pacific coast every method we tested is weak, ours included. And our own dates are odds from a 26-year record for a 250-meter cell — not a promise about next Tuesday.

That is the honest shape of it: the standard method carries an error that never reaches the gardener as a number, we put a number on it, and where we beat it we can show by how much — including where we do not.

Get the frost odds for your exact spot — free

How this was measured

NOAA's U.S. Climate Normals give each weather station its official 1991–2020 freeze probabilities. We publish that file as a free CSV. For this study we took every station carrying a median last-spring-frost date, found its nearest neighboring station by great-circle distance, and compared the two dates.

• 6,949 stations carried a usable median last-frost date.
• 6,935 nearest-neighbor pairs were formed and measured.
• 744 pairs were excluded from every headline number. A pair only counts if both stations have a complete record and actually freeze in at least 90% of years. Where frost is irregular, a "median last frost date" is a median taken over years that include no frost at all — it swings on a couple of mild winters and it is not something a gardener can act on. Every figure above is the remaining 6,191.
• Day differences are circular, so a pair straddling New Year is two days apart rather than three hundred and sixty-three.
• 14 stations found no neighbor inside the search window and were dropped. Those are the most isolated stations in the country — exactly where a nearest-station lookup does worst — so leaving them out makes this result more conservative, not less.

Does the finding depend on where we drew the line?

That 90%-of-years rule is our judgment call, and a filter that manufactures its own conclusion is worth nothing. So here is the same computation without it.

The median does not move and the two-week rate shifts by a point. We went with the stricter filter because it is the more defensible one, not because it flattered the result — it makes our number slightly smaller. Both cuts are in the JSON.

What this study does not show. It measures how much two stations disagree with each other, which is a proxy for how wrong a substituted station is — not a direct measurement of error at your garden. Nobody has a thermometer in your garden, which is the whole difficulty. It also uses median (50%) dates only; the spread at the 10% and 90% ends is wider still. And a large disagreement between two stations does not tell you which of the two is closer to your ground.

Check it yourself

We would rather be checked than believed. The input file is our free public download, the computation is one script with no dependencies, and the full result — including the marginal pairs we excluded — is the JSON below.

• The full result as JSON (7 KB) — every figure on this page.
• The input dataset — NOAA 1991–2020 station freeze normals, 7,305 stations, free CSV.
• The source data is NOAA NCEI U.S. Climate Normals 1991–2020, US public domain. This study is CC BY 4.0 — use it, cite it.

Source: NOAA NCEI U.S. Climate Normals 1991-2020, freeze probabilities (our open release)
Method: For every station with a median last-32F date, find the nearest other station and compare the two dates. Headline figures count a pair only when both stations have a complete 18-date record AND freeze in at least 90% of years. This is the error a nearest-station lookup inherits — a LOWER BOUND, since a real garden sits between stations, not on one.

Questions