Somerset Weather · Hyper-local Forecast

Two forecasts on this site

The one you read (Today, Hour by hour, 7 days, Why) is our statistical forecast, described on this page: the four weather models' average, corrected with the station's own record. The one on trial is our physics forecast: a simulation of the air, grass and soil at the mast, started from the station's readings and corrected the same way. It is made at the same times and graded on the same How we're doing; if it wins, it becomes the forecast. Which model made each number, the treatments being tried on the physics, and today's numbers from every model side by side: The models.

How this forecast works

1 · The big modelsFour global weather models: the American GFS, European ECMWF, German ICON and Canadian GEM, plus the National Weather Service's own forecast and 122 ensemble runs that show how much the models could be off.
→
2 · Their averageFor each high, low and 12-hour period, the four models are averaged at the station's exact location.
→
3 · The station correctionA correction learned from this station's own record adjusts that average for clouds, wind, humidity, season, how much the models disagree, and their recent errors.
→
4 · The forecastHighs and lows with ± ranges, rain chances, frost and freeze chances, gust and heavy-rain levels, and the hour-by-hour run.

The temperature correction is a straight-line formula, so every degree can be traced (see "Why ours is…" in the latest run in detail). Rain chances come from a probability model of the same inputs, and a small AI model on the station computer reads the National Weather Service's written forecast discussion and adds what it says about rain. Every piece has to pass a fair test before it is called tested.

Model version
2026-10-14
the date it was trained
Learned from history through
2026-10-13
917 days of highs it learned from
Next scheduled retrain
-
Friday, January 1, 2027 at 4 AM
Status report
-
rebuilt every hour

How we test

  1. Practice and grading are kept apart. A formula is built from one stretch of the station's past, then graded on a later stretch it was never built from, like a student who practices on one set of problems and is graded on another. A formula that only looks good on its own practice data is usually memorizing noise, so only the grade on set-aside data counts.
  2. The passing mark is written down first. Before grading, the Lab Notebook records exactly what counts as a pass: for example, at least 0.05° better than what we use now. The result is published either way, pass or fail.
  3. Luck is checked. A small gain can be luck. So the graded days are re-drawn at random thousands of times, and a gain has to hold in at least 9 of 10 re-draws.
  4. New days settle it. Some features then run "on trial": logged every day as the weather happens and judged again once enough new days have passed.

How accurate it is

Average error of our highs and lows, tested on October 1, 2025 to October 13, 2026 (a fair test: days set aside from building the model)
Days aheadHigh: our forecastModel averageImprovementLow: our forecastModel averageImprovement
1 (made that morning or evening) 1.6°2.0° 21% 1.7°1.9° 12%
2 (made the day before) 2.0°2.5° 22% 2.0°2.3° 13%
3 2.4°2.8° 15% 2.3°2.6° 10%
4 2.8°3.2° 13% 2.5°2.8° 9%
5 3.2°3.6° 10% 3.1°3.4° 7%
6 4.0°4.3° 7% 3.8°4.0° 5%
7 4.8°4.9° 3% 4.7°4.9° 4%

"Average error" is how far off the forecast was, on average, in degrees. Day 1 is the forecast made just before the high or low began: that morning for the high, that evening for the low. Day 2 is the one made the day before, and so on. The biggest gains are 1 to 3 days out, where local detail matters most; by day 7 the big-picture pattern dominates and there is less for a local correction to fix.

Rain chance: how much better than "typical for this time of year"
Days aheadDaytimeOvernight
138% 32%
236% 37%
331% 34%
428% 34%
526% 31%
625% 22%
711% 14%

0% would be no better than simply knowing how often it rains at this time of year; 100% would be perfect. Since October 2026 the rain chances also use the National Weather Service forecast discussion, which tested 3.4% better again over the next two days in the fair test (the Notebook has the test).

How the live forecasts are doing, every day, against the National Weather Service, WxSim and the weather models: the How we're doing. The biggest misses, and what was going on: Misses.

Updates during the day Tested

Once a period has begun, the forecast keeps updating every hour with how warm or cool the station is running. That update passed a test on a year of days set aside for grading; these checkpoints keep checking it on real days, against the forecast made before the period began.

CheckpointMade before it beganModels only, thenBefore it began, never past the high or low so far Updated with the station (shown)National Weather Service, thenDays scored
Today's high at 11 AM1.3°1.2° 1.3°1.2°1.7° 32
Today's high at 2 PM1.3°1.1° 1.3°1.0°1.7° 32
Tonight's low at 9 PM1.5°1.5° 1.5°1.3°2.8° 31
Tonight's low at 11 PM1.5°1.5° 1.5°1.3°2.8° 32

Average error in degrees, same days in every column. "Models only" is the hourly rebuild from the newest model runs without looking at the station. In the fair test (October 2025 to October 2026) the update cut the error on the afternoon high and on the small-hours low by about a quarter against "never past the high so far"; the exact numbers, the test and this live check are in the Lab Notebook.

What it's built from

Station history the forecast learned from
1.87 million
one-minute readings since February 1, 2023 (Weather Display)
The station's full record
4.48 million
one-minute readings since October 2017, the same Davis throughout (Cumulus, then Weather Display); used for the climate normals, station records and the lab's research
Station readings logged
11,794
every 5 minutes since October 10, 2026
Forecast runs saved
4,157
model, National Weather Service and outlook snapshots
Forecasts we've issued
980
hourly runs, 131,295 numbers logged for scoring
Forecast discussions read
2,009
1,973 from history + 36 live by the AI model
Database on the station computer
108.8 MB
everything collected so far
What the station computer savesSavedSinceLatest
Global model runs923October 13, 2026 November 19, 6:51 PM
Upper air & sunshine runs334October 14, 2026 November 19, 5:13 PM
Ensemble runs (122 members)170October 14, 2026 November 19, 4:41 PM
HRRR runs766October 14, 2026 November 19, 6:39 PM
NWS gridpoint forecasts484October 13, 2026 November 19, 4:11 PM
NWS text forecasts788October 13, 2026 November 19, 6:08 PM
WxSim runs197October 13, 2026 November 19, 5:38 PM
National surface analyses300October 14, 2026 November 19, 5:38 PM
NWS alert checks19October 16, 2026 November 19, 7:03 PM
Storm outlooks (day 1)176October 16, 2026 November 19, 6:39 PM

A new copy is saved only when the content changes, so the counts are real updates, not repeats.

Growing for the next version

Ensemble archive: 36 days of the full 122-run ensemble saved, of about 180 needed before calibrating it to this station can be tested fairly.

What's proven, what's on trial

Every feature on this site, and how far it has been tested. "Proven" means it passed a fair test. "On trial" means it is being checked on new days as they happen. "Experimental" means it is shown for interest and not yet tested. Badges link to the test in the Lab Notebook.

FeatureStatusWhere
Highs and lows, with their ± rangesProvenToday, 7 days
Rain chances, with the National Weather Service's written discussionProvenToday, 7 days
Today's high and tonight's low, adjusted liveProvenToday
The 6-hour nowcast (the station's error and the upwind airports)ProvenToday, Hour by hour
Today's high and tonight's low updated with the station once the period has begunProvenToday, 7 days
Frost and freeze chances; gust and heavy-rain levelsProvenToday, Significant weather, Freeze and growing season
The hour-to-hour shape between high and lowModels' ownHour by hour
Our physics forecast, corrected with the station record: the candidate to become the forecastOn trialThe models, Physics, Scoreboard
Treatments of the physics forecast, one change eachOn trial On trial On trial On trialThe models
The physics forecast started from the station's latest reading every hour (the half in progress restarts)On trialThe models, Physics
Physics model: day high (with the crop calendar), inside the main modelProvenPhysics
Physics model: snow on the ground insulates the grass, inside the main modelProvenPhysics
Overnight update started saturated when the station is saturated (fog)Proven On trialPhysics
Night physics 1G: an evening layer that starts shallow and deepens (the shape of the night)On trialPhysics
The physics on every ensemble member: odds of the low, frost, freeze and saturationOn trial On trial On trialPhysics
Rain, snow and freezing rain in the physics modelNot provenPhysics
Grass temperature estimate at nightExperimentalHour by hour, Freeze and growing season
Harvest hours, drying, spray windows, your-corn drydownExperimentalIn the field
Significant weather: which kinds are in play, from the ingredientsExperimentalSignificant weather
Snow on the ground estimateExperimentalGround (in winter), Snow

Limits worth knowing

Going deeper: the latest run in detail (every input, the forecast discussion, why each number is what it is, run-to-run changes, recent verification) and the model charts (hour by hour, and the upper air).

Retraining

Four times a year the station computer refreshes everything (the station's minute logs, the models' past forecasts, the forecast discussions), checks the station history for glitches like September 2026's false "rain days", re-runs the tests and trains a new model. The new model only replaces the current one if it does at least as well on the days since the last training, days neither model has seen. Otherwise the current one stays. The old model is always kept, so going back is one step.

Upcoming
History
WhenOutcomeTest on new daysWhy
October 16, 2026 Kept current - not enough new days yet for a fair test (3 of 45)
October 16, 2026 Kept current - not enough new days yet for a fair test (3 of 45)

Glossary

80% range (±). The range the actual value fell inside four times out of five when tested.

Average error. How far off a forecast was, on average, in degrees, ignoring whether it was too warm or too cold.

Model average. The four global models (the American GFS, European ECMWF, German ICON and Canadian GEM) averaged at the station. "Our forecast" is that average after the station correction.

Ensemble. A model run many times with slightly different starting points; the spread of the runs shows the uncertainty.

Nowcast. A forecast for the next few hours that leans on what the station and nearby airports are reading now.

Frost. Ice forming on grass, roofs and windshields. It needs the grass at 32° or colder, which on a clear, calm night can happen with the 5-foot air at 36° or warmer.

Freeze. The air at 5 feet reaching 32°. A killing freeze is about 28° for a few hours, enough to end the season for tender crops.

Grass temperature. The air right at the tips of the grass blades, measured by the station's ground sensor (or estimated at night). On clear, calm nights it can run 10° to 20° below the air the station measures 5 feet up.

Air up the mast. A second air sensor 115 inches up the station mast (from October 23, 2026), against the station's own sensor at 5 feet. When the higher one reads warmer at night, the air is warming with height: a surface inversion, the still, layered air of a frost or fog night, and the condition that holds spray drift close to the ground.

Dew point. The temperature at which the air becomes saturated; dew or frost forms when the grass cools to it.

Wet-bulb temperature. How cold the air gets when water evaporates into it. Falling snow can survive to the ground in air a few degrees above freezing when the wet-bulb is near 32°.

Temperature inversion. Air near the ground colder than the air above, common on calm, clear nights; it traps smoke, fog and spray drift.

Delta T. The air temperature minus the wet-bulb: how quickly spray droplets evaporate.

925 mb, 850 mb, 500 mb. Pressure levels about 2,500, 5,000 and 18,000 feet up, where forecasters watch the air mass.

CAPE, lifted index, K index, Total Totals. Measures of how unstable the air is, which is how much energy is available for thunderstorms.

Growing degree days (86/50). Each day's average temperature above 50°, with highs capped at 86°: the heat units crops develop on.

Physics model. Our own model of the heat budget at the station mast (sun, sky, ground, air), run beside the forecast.

Reanalysis (ERA5). A best reconstruction of past weather from all observations, used for normals and for checking.

HRRR, NBM, MOS. A fast-updating high-resolution American model; the National Blend of Models; and Model Output Statistics, the National Weather Service's statistical corrections of model output.