Data and model status

Current coverage

• 32,024 Census places—including cities, towns, villages and CDPs—with local routes.
• 33,640 Census ZIP Code Tabulation Areas; 33,043 have 2020–2024 ACS population estimates.
• 4,551 valid NOAA station normals and four-station lineage for every place.
• 4,550 stations with independent December 25 observations from 1991–2025.
• 896 major place routes prebuilt; every smaller place and ZIP area remains available on demand.

Current model

Calibrated boosted spatial model using target-excluded NOAA/GHCN-D neighbors, USGS terrain and geographic context; promoted after sealed temporal, regional and spatial-block tests. A challenger can only replace it after clearing the predeclared accuracy, calibration and geographic gates.

Independent station test

Tree complexity was selected using 1991–2015 for fitting and 2016–2020 for tuning. A recent-period calibration used only pre-test data. The untouched 2021–2025 period contained 12,588 station/Christmas outcomes. NOAA scored 0.129416 on the Brier scale, versus 0.125306 for the selected challenger— 3.18% better.

• Selected model: monotonic histogram gradient boosting with 23 leaf nodes, after six locked complexity candidates.
• NOAA calibration error: 7.51%; challenger: 6.80%.
• A simpler calibrated spatial logistic model also passed at +2.19%; boosting improved the sealed score further.
• All Census-region holdouts were positive (worst +0.08%) and all five blocked spatial folds were positive (worst +0.43%).
• Activation rule: At least 1% Brier skill, no worse calibration, and no Census region or spatial fold below -1% on untouched 2021-2025 observations.
• Decision: promote boosted spatial v4.

Repaired SNODAS benchmark

The official NSIDC zero-repair mask was applied to the documented 2014–2018 coastal defect, allowing all 21 Christmases from 2005–2025 to be evaluated. Across 403,277 place/Christmas outcomes, NOAA scored 0.124222 and the selected SNODAS blend scored 0.124736—0.41% worse.

• Source: NOAA NOHRSC SNODAS at NSIDC, dataset G02158 version 1.
• SNODAS modeled snow depth >= 25.4 mm at 06:00 UTC on December 25.
• Alaska and Hawaii remain outside the masked SNODAS grid.
• SNODAS is a modeled analysis; GHCN-D supplies the independent station observations used for the stricter v2 decision.

Next model work

Archive NBM and global ensemble forecasts for lead-time calibration; continue spatially blocked climate-model challenges. GEFSv12 snow-water-equivalent reforecasts informed conservative 20–30% live weights after positive skill at every tested lead on five sealed Christmases. Operational ensembles provide direct snow-depth probabilities, so this transfer remains documented and the daily archive continues.