How SoilStack predicts last-spring-frost dates

01What this is

Every SoilStack zone page shows a section called NOAA Frost Probability with three dates: an earliest, a typical, and a latest. Those three dates come from a NOAA dataset called the 1991–2020 Annual/Seasonal Climate Normals — a 30-year statistical baseline that climatologists update once a decade. We refer to it as the Normals below.

This page documents exactly how those three dates are derived, what each one means, what the underlying data covers, and where it doesn't. It also documents the one subtle place where NOAA's labeling convention runs opposite to most gardening literature — a detail we surface here so anyone digging into the source data understands why the numbers we display might look like they're in reverse order.

The frost dates themselves are NOAA's. SoilStack's contribution is the presentation layer: aggregating thousands of weather stations into per-USDA-zone composites, translating the probability percentiles into plain-English headlines, and surfacing the citation chain so any AI engine, journalist, or researcher can verify our numbers against the primary source in two clicks.

What “last spring frost” means here. Throughout this page (and on every SoilStack zone page) “last spring frost” refers specifically to the last calendar date the minimum temperature reaches 32°F or below — the freezing point of pure water and the temperature NOAA uses for its primary frost climatology element (TMIN-PRBLST-T32F). This is a hard scientific threshold, not a biological one: some crops can be damaged at temperatures above 32°F depending on duration of exposure, humidity, wind, and the crop's individual cold sensitivity. These dates are the freezing-threshold benchmark, not a universal “safe to plant” signal. NOAA also publishes probability dates at five other thresholds (16°F, 20°F, 24°F, 28°F, and 36°F) for stations and applications that need different cutoffs — see Section 2 for the full element catalog.

02Data sources

Primary: NOAA NCEI 1991–2020 Annual/Seasonal Climate Normals

Published in 2021 by NOAA's National Centers for Environmental Information. Dataset identifier gov.noaa.ncdc:C01619. Covers approximately 15,000 U.S. weather stations and reports last-spring-frost (and first-fall-frost) probability dates at six temperature thresholds — 16°F, 20°F, 24°F, 28°F, 32°F, and 36°F — for nine probability percentiles each. Released under the NOAA Open Data Dissemination license (public domain). Update cycle: every ten years.

We pulled this dataset via NCEI's Access Data Service in May 2026. Across the high-priority subnetworks (USC, USW, USS — Cooperative Observer, Weather Service first-order, and SNOTEL stations), 6,877 stations returned usable frost data on the first ingestion pass.

Supporting: US Census Bureau 2024 ZCTA Gazetteer

To map weather stations to USDA hardiness zones we needed ZIP-code centroids. Those come from the U.S. Census Bureau's 2024 ZIP Code Tabulation Area (ZCTA) Gazetteer — the canonical public-domain ZIP-to-coordinate reference. 33,791 ZCTA centroids were used to build a ZIP-to-station crosswalk via great-circle (Haversine) distance with inverse-distance-squared weighting. Median nearest-station distance across all U.S. ZIPs ended up at 11.85 km.

Supporting: USDA 2023 Plant Hardiness Zone Map

The Plant Hardiness Zone Map released by USDA in 2023 defines which 5-digit ZIP code maps to which hardiness zone (e.g. 47403 → Zone 6b). We use that mapping to aggregate per-ZIP frost composites into per-zone composites.

Coming in V2: NWS Gridpoint Forecast

A V2 of the Frost Prediction Engine will blend the climatological baseline above with the National Weather Service's 7-day Gridpoint Forecast. Methodology for that blending will land here when V2 ships — see Section 5's Coming next: live forecast blending subsection for the planned approach.

03How NOAA calculates frost-date probabilities

Here is the procedure NOAA applies to every station in the network, plain-spoken:

1. Take 30 years of daily minimum-temperature observations from a single weather station — for the current product, calendar years 1991 through 2020.
2. For each of those 30 years, find the last day the station's daily minimum temperature reached or fell below a chosen threshold. For the 32°F threshold — the consumer-canonical definition of "frost" — that's the last spring day each year that frost was recorded.
3. That gives a sample of last-frost dates across the 30-year record. Rather than reading percentiles straight off those raw dates, NOAA uses a statistical resampling (“bootstrapping”) procedure: it generates 10,000 simulated annual cycles of daily minimum temperature drawn from the record, which smooths out the jaggedness of a short 30-year sample.
4. From those simulations it derives, for each calendar date, the probability that the last spring freeze has not yet occurred — reported at the 10%, 20%, … 90% levels. That set of probability dates is the last-frost probability curve for the station.

Repeat across approximately 15,000 stations and you have the dataset SoilStack draws from. The full procedure, with all the statistical caveats around station completeness, missing-data interpolation, and quality-control flagging, is documented in NOAA's Normals Calculation Methodology (2020); the dataset itself is cited as Palecki et al. (2021) — see Section 9.

04What P10, P50, and P90 mean for planting decisions

Most gardening references give a single "last frost date" per location. That's a useful headline number — it's the median, the date by which half of years see their final frost — but it conceals the actual year-to-year variability. The P10/P50/P90 framing surfaces that variability directly.

On a SoilStack zone page, we label the three percentiles in gardener-friendly English:

• Earliest — the date by which roughly 10% of years have already seen their last frost. Planting tender crops this early will succeed in about 1 in 10 years and lose to a late frost in the other 9. Aggressive gardeners with row cover and a willingness to replant target this date.
• Typical — the median. Half of years have their last frost on or before this date. This is the date most one-number "last frost" references give. Reasonable target for unprotected tender crops in an average year.
• Latest — the date by which roughly 90% of years have seen their last frost. Waiting until this date cuts the frost risk to tender crops to roughly 1 year in 10 — low, but not zero — at the cost of compressing your growing season. Conservative gardeners in cold-hardy climates target this date.

05The percentile convention (read this if dates look backwards)

NOAA and the gardening literature use the same probability framework but label their percentiles from opposite sides. NOAA's calculation methodology defines last-spring-freeze probability dates from the remaining-risk side — the probability that the last spring freeze occurs on or after a given date. SoilStack displays the same dates from the complementary gardener-facing side — the probability that frost is already done by that date. This section explains the mismatch so anyone comparing a SoilStack zone page to a NOAA data file sees why labels flip.

Convention 1 — "probability frost still to come" (NOAA NCEI)

NCEI's dataset is labeled in terms of probability that frost is still to come. Under this convention:

• P10 means “10% chance frost is still to come on or after this date.” That's the latest plausible last-frost date.
• P50 means “50% chance frost is still to come.” The median.
• P90 means “90% chance frost is still to come.” That's the earliest plausible last-frost date.

Convention 2 — "probability frost is done by" (gardener-facing)

Gardening references, university extension publications, and most consumer-facing frost calculators label things in the opposite direction — probability that frost is already done:

• P10 means “10% chance frost is already done by this date.” The earliest plausible last-frost date.
• P50 means “50% chance frost is already done.” Same median.
• P90 means “90% chance frost is already done by this date.” The latest plausible last-frost date.

The two conventions describe the same physical reality — the distribution of last-frost dates over a 30-year period — from complementary angles. The median (P50) is the same in both. The two endpoints swap labels.

SoilStack stores values in Convention 1 (matching the NCEI source attribution chain in our database) and displays them in Convention 2 (matching gardener expectations). The conversion happens in exactly one place in our code — the per-zone aggregation step — and is documented inline alongside the conversion itself. The result: anyone tracing the station-level inputs back to NCEI sees parity with the source data, with SoilStack's interpolation and display-convention conversion documented separately.

Coming next: live forecast blending (V2)

The methodology above describes V1 of the Frost Prediction Engine — pure climatological baseline. A V2 release is planned that blends this baseline with the National Weather Service's 7-day Gridpoint Forecast: as the calendar approaches your zone's frost window, the predicted curve will narrow if NWS shows no sub-freezing temperatures in the remaining window, and the page will say so explicitly. The math for that blending will be documented here in full once V2 ships. This section will then split into V1 baseline + V2 blending, with this paragraph replaced by the locked specification.

06How we compute per-zone composites

NOAA publishes data per weather station. A USDA hardiness zone covers many ZIP codes, and each ZIP code is served by several nearby stations. To go from station-level data to zone-level data we apply three aggregation steps, in this order:

Step 1 — Station-to-ZIP crosswalk

For every U.S. ZIP code, find the nearest weather stations by great-circle (Haversine) distance to the ZIP centroid, and weight each station's contribution to that ZIP by inverse distance squared. A station 5 km away counts four times as much as one 10 km away. Stations beyond 150 km of a ZIP are not used.

The crosswalk runs once per data refresh and produces a per-ZIP lookup: for every U.S. ZIP, here are the top contributing stations with their weights. Median nearest-station distance across all covered ZIPs is 11.85 km; 95% of ZIPs are within 26 km of a usable station. 191 ZIPs (chiefly bush Alaska) have no usable station within 150 km and are excluded.

Step 2 — Per-ZIP composite

For each ZIP, combine its contributing stations into a single composite frost prediction. Frost dates are aggregated using a weighted circular mean — the standard technique for averaging calendar dates, which handles year-boundary wraparound correctly (relevant for warm zones where last frost falls in January or February of the same calendar year). Numeric values like growing-season length are aggregated with a weighted arithmetic mean.

If a contributing station happens to be missing a specific measurement (some stations record temperature thresholds others don't), the remaining stations' weights are renormalized for that element rather than backfilled with neighbors. The result: honest nulls when no coverage exists; no quiet interpolation.

Step 3 — Per-zone composite

For each USDA hardiness zone, average the per-ZIP composites across every ZIP assigned to that zone in the USDA 2023 map. This is an equal-weight average across ZIPs — ZIP-level weighting already accounted for station distance in step 1, so the zone-level step doesn't re-weight by station count.

Reference values per zone (current build)

The 10 USDA hardiness zones currently covered, with their composite last-spring-frost percentiles displayed in gardener convention — Earliest = gardener P10 (10% of years have already had last frost), Typical = P50 (median), Latest = gardener P90 (90% of years have already had last frost). These are the same values appearing on each zone's page. Click any zone code to see the full chart.

Warm zones (9A and 9B) show last-frost dates in January, February, and early March of the same calendar year — those climates see their final frost in winter, not spring. The aggregation handles this year-wrap explicitly via a 365-day circular calendar internally; the dates above are correct.

6.5How accurate the interpolation is (cross-validation)

Documenting a method is one thing; measuring how close it lands is another. To put a number on the accuracy of the station-to-location interpolation in Step 1 above, we ran a standard leave-one-out cross-validation across every in-scope weather station.

What we tested

The procedure mirrors what the engine does for a real ZIP that has no weather station of its own — it estimates from the neighbors. For each of the 5,867 NCEI stations in zones 5A–9B, we removed that station from the network, re-derived its location's last-spring-frost dates (Earliest / Typical / Latest at 32°F) using only the remaining stations, and compared the estimate to the station's own published normals. Because a held-out station never contributes to its own estimate, the comparison is honest: it measures how well the inverse-distance method fills a gap, which is exactly the job it does in production. Across all stations and all three percentiles that yields 17,601 predicted-versus-actual comparisons.

The result

For context, the interpolation error is smaller than the P10–P90 frost-date spread shown in the zone composites — within a single zone the Earliest-to-Latest window is a month or more (see Section 4). That does not make the estimate a year-specific forecast, but it does mean the interpolation error is modest relative to the climatological frost window the product is built to summarize — though, as the table below shows, that accuracy varies by region.

Accuracy by zone

Accuracy is not uniform across the country: it is tightest in the cold and temperate zones and widens in the warm southern zones. That pattern is consistent with winter-timed freeze events, sparser effective station coverage, coastal gradients, and microclimate effects, but the table below should be read as an empirical error summary rather than a full causal diagnosis. The figures are the held-out-station errors grouped by the zone each station falls in — a regional breakdown of interpolation error, not a measurement of the zone-wide composite. Mean absolute error by zone:

About one comparison in nine missed by more than two weeks; these outliers cluster overwhelmingly in the warm zones and in the high-elevation and coastal microclimates described in Section 7. We report them rather than trim them — a station in a frost pocket or on a mountainside is genuinely hard to predict from its neighbors, and the larger error there is real information, not noise to be hidden.

What this does and doesn't establish

This tests the station-to-location interpolation — how well the inverse-distance step (Step 1 above) reconstructs a point from its neighbors. It does not separately validate the zone-wide composite average (Step 3), an individual ZIP centroid, or a specific backyard, all of which carry their own additional uncertainty. Nor does it independently validate NOAA's underlying station normals: NCEI is the sole published source for those, and NCEI's own records already use neighboring stations to fill gaps, so the figures describe how faithfully SoilStack reproduces NOAA's numbers at a withheld point, not an independent check of NOAA. The cross-validation is reproducible from the published data via the frost:cross-validate and frost:analyze-cross-validation commands, and the full per-station residuals — observed and predicted dates, signed and absolute error, contributing-neighbor count, and nearest-neighbor distance for every station tested — are published as a downloadable CSV: frost-cv-residuals.csv.

07Limitations

What this dataset does well, what it doesn't, and where it should be cross-checked with local knowledge:

It's a 30-year baseline, not a single-year forecast

The dates on this page describe the distribution of last-frost dates across 30 years (1991–2020). They are not a prediction of when last frost will occur in any one calendar year. For a specific year's forecast you need a short-range weather prediction; the V2 blending described in Section 5 is the intended next layer.

Microclimate variation is not captured at the zone level

If your garden sits at the bottom of a frost pocket, against a south-facing brick wall, or 800 feet up a hillside above your town, your actual last-frost date can be one to three weeks off the zone-wide composite. The dataset is calibrated to weather stations sited per WMO observation standards, which is roughly "open, level, away from heat sources" — not the typical backyard. The composite is a good starting estimate; your own yard's record over multiple seasons is the ultimate authority for your site.

Station density varies by region

Eastern U.S. zones have dense station coverage. Bush Alaska, high-elevation western U.S., and parts of the desert Southwest have substantially sparser coverage. The "Contributing ZIPs" column in the reference table above tells you how much ZIP-level coverage each zone composite has; a lower number means the zone composite is based on fewer mapped ZIPs and should be interpreted with more caution.

Data revision cadence is 10 years

NCEI updates the Climate Normals once per decade. The current product is the 1991–2020 release; the next refresh covering 2001–2030 will arrive around 2031. In a warming climate, an official 1991–2020 baseline may lag current conditions in some locations — a consequence of using decadal 30-year climate normals, not a site-specific current-year forecast.

No NOAA endorsement

NOAA publishes the source data and does not endorse SoilStack's presentation of it. The dataset is in the public domain under NOAA's Open Data Dissemination license; SoilStack's aggregation and presentation are our own work.

08Data freshness

Several timestamps relate to a SoilStack frost page; here's what each one means.

• The underlying NCEI dataset covers 1991–2020 and was published in 2021. It is refreshed by NOAA once per decade.
• Our ingestion timestamp — when we pulled the dataset into our database — is May 2026 for this V1 release.
• The per-zone "Composite generated" line on each zone page reflects when we last recomputed the zone composite from the ingested station data. That changes only when the aggregation logic itself changes or when station data is re-ingested.
• This methodology page's "Updated" timestamp — the June 16, 2026 stamp at the top — reflects when the methodology document last changed. That's separate from the data refresh cycle. See Section 11 for the methodology version history.

Because the underlying NCEI dataset only refreshes once per decade, SoilStack frost pages do not regenerate daily. The data is stable climatology, not a daily forecast. When V2 lands and we begin blending in the live NWS forecast, the per-zone pages will gain a second, faster-moving timestamp; this methodology page will document the dual-cadence model when that happens.

09Citation chain

The NCEI 1991–2020 Normals product is published as a dataset (identifier gov.noaa.ncdc:C01619; dataset citation: Palecki, Durre, Applequist, Arguez, & Lawrimore, 2021), with its production methods described in NOAA's Normals Calculation Methodology (2020) and the peer-reviewed literature behind the established normals algorithms. We list the load-bearing references below.

Primary source and methodology documentation

• Palecki, M., Durre, I., Applequist, S., Arguez, A., & Lawrimore, J. (2021). U.S. Climate Normals 2020: U.S. Annual/Seasonal Climate Normals (1991–2020). Dataset, NOAA National Centers for Environmental Information (identifier gov.noaa.ncdc:C01619). ncei.noaa.gov/products/land-based-station/us-climate-normals
• NOAA National Centers for Environmental Information (2020). Normals Calculation Methodology 2020. Documents the production methods for the 1991–2020 Normals, including the bootstrapped frost-freeze probability procedure.

Supporting normals methodology

• Applequist, S., Arguez, A., Durre, I., Squires, M., Vose, R., & Yin, X. (2012). 1981–2010 U.S. hourly climate normals. Bulletin of the American Meteorological Society, 93(11), 1637–1640.
• Arguez, A., Durre, I., Applequist, S., Vose, R. S., Squires, M. F., Yin, X., Heim, R. R., & Owen, T. W. (2012). NOAA's 1981–2010 U.S. climate normals: An overview. Bulletin of the American Meteorological Society, 93(11), 1687–1697.
• Durre, I., Squires, M. F., Vose, R. S., Yin, X., Arguez, A., & Applequist, S. (2013). NOAA's 1981–2010 U.S. climate normals: Monthly precipitation, snowfall, and snow depth. Journal of Applied Meteorology and Climatology, 52(11), 2377–2395.
• World Meteorological Organization (2017). WMO Guidelines on the Calculation of Climate Normals (No. 1203).

Climate-trend context

• McCabe, G. J., Betancourt, J. L., & Feng, S. (2015). Variability in the start, end, and length of frost-free periods across the conterminous United States during the past century. International Journal of Climatology, 35(15), 4673–4680.
• Kukal, M. S., & Irmak, S. (2018). U.S. agro-climate in 20th century: Growing degree days, first and last frost, growing season length, and impacts on crop yields. Scientific Reports, 8(1), 6977.

10For developers and researchers

Every per-zone frost page emits a schema.org/Dataset JSON-LD entity with the full citation chain machine-readable. The entity ID pattern is:

where {zone} is one of 5a through 9b. Each Dataset declares isBasedOn → the NCEI source dataset, citation → the NOAA/NCEI dataset and calculation-methodology documentation, license → the NOAA Open Data Dissemination URL, and isPartOf → this methodology page (https://soilstack.net/frost/methodology#article), along with three PropertyValue entries for the P10/P50/P90 dates with the NCEI element keys documented.

This page emits a corresponding schema.org/TechArticle entity that lists the Dataset entities via mentions. An AI engine or knowledge-graph crawler traversing either node lands in a complete, self-consistent subnet for SoilStack's NOAA-cited frost data.

The cross-validation residuals behind Section 6.5 are published as a flat CSV at /data/frost-cv-residuals.csv — one row per station and percentile, with observed and predicted dates, signed and absolute error in days, the contributing-neighbor count, and the nearest-neighbor distance. It regenerates from the same frost:cross-validate run that produces the Section 6.5 figures.

Citing SoilStack

If you cite a SoilStack zone-page frost prediction in writing, please attribute the underlying data to NOAA NCEI and link this methodology page so readers can verify the chain. A workable short form:

Or you can cite NCEI directly for the underlying data and skip us — NOAA's dataset and methodology documentation handle the rigor on their own.

11Version history

This page is versioned. When the methodology itself materially changes, the version number bumps and a row is added below. Wording-only edits don't bump the version.

v1.1 — June 16, 2026

Empirical validation added. Section 6.5 reports leave-one-out cross-validation accuracy (median 5-day, mean ~7-day absolute error across zones 5A–9B) measured by holding out each NCEI station and predicting its last-frost dates from its neighbors, with the full per-station residuals (17,601 observations) published as a downloadable CSV for external auditing. The aggregation method itself is unchanged — this is a measurement of the existing method, not a modification of it.

v1.0 — May 21, 2026

Initial methodology release. Documents the Frost Prediction Engine V1 build: NOAA NCEI 1991–2020 Climate Normals as the sole data source, station-to-ZIP crosswalk via inverse-distance-squared Haversine weighting, per-ZIP composites via weighted circular mean, per-zone composites via equal-weight average. Documents the NCEI percentile convention bridge (Section 5).

Planned future versions: v2.0 will land with the V2 NWS short-range forecast blending (Section 5 will then split into V1 baseline + V2 blending). v2.1 will land with autumn first-frost extension (same methodology applied to autumn variables in the same NCEI dataset).