Your frost dates, computed for your exact spot

What just shipped

Our free frost date lookup got a new engine this week.

Until now it worked the way most frost tools do: find your nearest NOAA weather stations (NOAA is the US government's weather agency), show you their numbers. Good stations, honest odds — but the answer was always about the station's spot. Not yours.

Now the primary answer is computed for your spot. Type an address and the tool reads frost odds for the roughly 800-foot square you're standing in, corrected for that square's own elevation. Behind it: 26 years of daily low temperatures (1995–2020) from NASA's Daymet — a dataset that reconstructs daily weather for every square kilometer of the continent — sharpened from that 1 km grid to ~250 m squares using elevation data from USGS, the US Geological Survey, then precomputed for the entire lower 48. The result is one big baked frost map we call the surface; your lookup reads 36 bytes out of its 6.32 GB in storage (the raw baked file runs nearer 8 GB before the ocean is trimmed away).

Your nearest stations still show up with every answer — as the cross-check, with the distance listed and the agreement measured in days. When the surface and the station disagree, the tool says so and says by how much.

It's free, there's no signup, and it stays that way. Try it with your address.

Why ZIP lookups get frost wrong

The standard method behind almanac-style frost lookups is simple: snap your ZIP to the nearest weather station and hand you that station's dates. On flat ground that's mostly fine. In terrain, it falls apart — and NOAA's own stations prove it.

Cache Valley, Utah. Two NOAA stations, one valley. The Utah State University station sits on a bench — a raised shelf of ground — at 4,790 feet. The Trenton station sits on the valley floor, about 335 feet lower. Per NOAA's 1991–2020 climate normals (the official 30-year averages), the floor's frost-free season is 45 days shorter than the bench's. Forty-five days, inside one valley. And the twist: the bench is higher and warmer, because cold air drains downhill at night and pools on the floor.

North Ogden and Eden, Utah. North Ogden sits on a bench at the mouth of Ogden Valley. Eden sits in the valley basin behind it. NOAA's nearest stations put North Ogden's frost-free season at 186 days and Eden's at 124 — a 62-day gap between neighboring towns.

A single ZIP-snapped date can't carry any of that. Whichever station your ZIP lands on, somebody nearby is getting a date that's weeks wrong for their yard.

The accuracy numbers, published

Every frost tool asks for trust. We'd rather show you the test.

Here's how we measure: bake the surface, hide real NOAA stations from it, then score the surface's best-guess (median) date predictions against what those hidden stations actually recorded — a hidden-station test. The baseline gets the same test — a nearest-station lookup, the standard method. Errors are in days: MAE in the table is mean absolute error, the average size of the miss. "Complex terrain" means real relief within a kilometer of the station. And the "10–90% band" is the date range the tool shows around its best guess — the span it expects about 8 in 10 real outcomes to land inside; "held" means they did.

The claim we're comfortable making, exactly as measured: in complex or station-sparse terrain — the places where frost actually gets weird — the surface's median date runs about 5 days more accurate than a nearest-station or ZIP lookup, with probability bands that held in 92–94% of cases. Where a good station sits close to you, the two mostly agree, and the tool shows you that agreement instead of asking you to pick a side.

Two pilot regions aren't the whole country. A national validation run against all ~7,300 NOAA normals stations is crunching right now; the numbers on the tool page swap the day it lands, whichever direction they move. The full method and the provenance of every number live on the methodology page.

What it can't do yet

Saying what shipped means saying what didn't.

• It can't see cold-air pooling. That 45-day Cache Valley spread? The new surface captures zero of it — we measured, and we're telling you anyway. Elevation correction can't catch a cold pool, because the cold spot is the low one and the warm spot is the high one. Same story in North Ogden vs. Eden: the surface separates them by 18–20 days of that real 62-day gap. A cold-air-drainage correction is in the works as the next stage — it ships only if it beats the current surface in this same held-out-station test.
• It's climatology — long-run odds — not a forecast. These are the odds over 26 years, not a prediction for this year. The week you plant, check the live forecast. Every time.
• There's no frost alert yet. That's the plan for the GardenTrack app, in beta now: it knows your garden's odds from this same surface, watches the weather, and gives you a heads-up when frost's in your local forecast. Until that ships, it's a roadmap item, not a feature.

That last one is personal. I once put over a hundred seed starts outside in early April because it had been warming up — and didn't check the weather. One frost, overnight, and months of work were gone. A heads-up on my phone would have saved all of it. That's a big part of why GardenTrack exists.

What's next

This blog is the build log. We'll keep publishing what we build, what breaks, and the numbers as we measure them — including the misses.

We also plan to open up the data itself, starting with the cleaned 7,305-station NOAA freeze dataset we validate against. That release isn't live yet; when it is, it'll be announced here.

If you want the war story behind this launch — the tile-naming bug that nearly poisoned 3.3 million cells of this surface, and the paranoid check that caught it — that's the next post.