Winter

Accumulation guidance

Snow odds for the window

Share of ensemble members reaching each total.

Day by day, within the window

Accumulation plume

Precipitation type and profiles

Precipitation type & transitions

Loading deterministic models…

Surface temperatures

Observations

How signals are set

HIGH 50%+ chance of 4" or more.

ELEVATED 40%+ chance of 1", or 20%+ chance of 4".

MARGINAL 10%+ chance of 1", or 30%+ chance of any measurable snow.

LOW None of the above.

Chances are shares of ensemble members, weighted by ensemble. They describe the model guidance, not a calibrated probability.

Precip type methods

Model type is each model's own WMO weather code (rain, snow, freezing rain/drizzle). The WMO code set models use has no sleet / ice-pellet category, so models cannot say "sleet" here. NWS wording can, and is shown in its own row.

Profile type (experimental) is the original Bourgouin energy method as NCEP runs it operationally (UPP routine CALWXT_BOURG, after Bourgouin 1992/2000). It is not the revised Bourgouin method (Birk et al. 2021). Layer energy is g·Δz·ln(T/273.15) J/kg, integrated on 25 m steps: surface warm energy, the summed energy of every warm layer aloft (PA), and the cold energy beneath the highest warm layer (NA). PA under 2: the surface warm energy decides (under 5.6 snow, over 13.2 rain, between a rain/snow mix). Otherwise NA > 66 + 0.66·PA gives sleet (melted to rain if the surface warm energy is over 13.2), NA < 46 + 0.66·PA gives freezing rain if the surface is below 0 °C (else rain), and between them a freezing rain/sleet mix. UPP draws randomly in transition zones; here they are labelled as mixes. How it is checked: GitHub Actions compiles the original UPP Fortran (pinned commit) and runs it on the same layer inputs as this code, comparing melting and refreezing energies and categories, threshold cases included (tests/upp/). A separately written JavaScript port is also used for quick offline cross-checks.

Three different questions. Algorithm agreement: the energy sums and decision tree match the original UPP code on the tested layer inputs. Profile representation: turning a model's sparse levels into those layers is an approximation. The primary reading uses layers linear in height (25 m); a second reading uses step layers between level midpoints. Across the 842 test profiles the two give different categories in 12 and refreezing energies differ by up to 1,280 J/kg. When they disagree for a model-hour it is marked REP-SENSITIVE and both diagnoses are shown. Forecast skill: neither check says which type will actually fall. That needs the observation log. Agreement with NWS is a comparison, not verification.

Coverage checks (provisional). Model profiles are sparse pressure levels, so passing these checks does not mean every warm or cold layer was resolved; failing them means the result would be guesswork. Profile type needs the surface, a first level within 1500 m, data through at least 3000 m and a sub-freezing top. Kuchera separately needs the column through 500 hPa with no gap over 1800 m. The largest gap between levels overlapping the lowest 3 km is shown, and marked COARSE above 1000 m, whatever the model. These thresholds are provisional. Profile type shows only when the model has ≥ 0.1 mm/h of precipitation and does not see moisture. Model counts are not probabilities.

Profiles: GFS, HRRR, ICON, GEM and UKMO have 8–10 levels up to 500 hPa; ECMWF has 6 (1000, 925, 850, 700, 600, 500); NBM and NAM have none, so they show model type only. Rain/snow split: a model's "rain" counts only when both its rain and showers fields are reported; a missing field is unknown, never zero.

About these numbers

Snowfall is each member's own model snow, which Open-Meteo reports at a fixed 7:1 depth-to-water ratio (its documented conversion). Cold, fluffy events can run higher and warm, wet ones lower, and it doesn't account for melting on warm ground.

A member only counts toward the odds and percentiles if it has every hour of the window. A member with gaps can only give a lower bound ("at least"), so it is listed separately and never decides a threshold. Its plume line stops where its data stops.