Weather Trader forecasts one thing above all: the daily high and low at a weather station,
the number that the climate record prints and that temperature markets settle on. This page says how that
number is built, how wide its range is, how it becomes a price, and how it has done. Each step names the
page where you can check it.
How the weather models themselves are scored (500 hPa skill, regimes, the Forecast Fork) is on
the model method page.
step 1
Model runs
Ten global and regional models, every cycle, read at each station.
step 2
Station MOS
Each model's own error at each station is learned and removed.
step 3
Consensus
The corrected forecasts are blended by how well each has done at that lead.
step 4
Range
The blend's own past error sets the 10–90% range.
step 5
Price
The range becomes the odds of any outcome, on each market's rule.
step 6
Score
Every issue is kept and scored against what the station recorded.
1The target
what counts as the high and the low
The forecast is for the official daily high and low of the local standard day, midnight to midnight
without daylight saving, in whole degrees Fahrenheit. That is the value in the National Weather Service
climate report, and it is the truth every station forecast here is trained on and scored against.
About 1,190 of the 2,411 forecast stations have that official record; the scores below use those.
Markets do not all settle on that value. Section 5 covers the difference.
2Station MOS: each model, corrected at each station
model output statistics
A weather model's grid value is not the airport thermometer. It runs warm at one station and cool at the
next, and its error changes with cloud, wind, the time of the high and the season. MOS learns that
error from the past and removes it.
One method for every model. A regression predicts each model's error (observed minus forecast)
from the model's own fields: its high or low, the hour it occurs, cloud, wind and humidity near the peak,
the temperature of the air above the surface, and onshore flow at coastal stations.
Regional equations with a station term. Stations are grouped into 33 regions. Each region's
equation is pulled toward the national one where it has little data, and each station keeps its own
recent offset.
Updated every day. Recent days count most (half the weight is from the last 30 days), so the
equations follow the season.
No hindsight. A day enters training only after its official value is published, and every score
is from a forecast made with equations that had not seen that day.
The models: ECMWF AIFS and the AIFS ensemble, ECMWF IFS, GFS, ICON, the UK Met Office global model, Canada's
GEM, HRRR, RRFS and a WeatherNext 3 station product for today and tomorrow. Each has its own page under
Station MOS. A model version the equations were not trained on is held back until
they are retrained.
3The consensus: one high and one low
a blend weighted by measured skill
No single corrected model is best at every lead. The MOS Consensus blends
them, with the National Blend of Models as one more member.
Weights come from the record. For each lead day, and separately for highs and lows, the weights
are the ones that would have given the smallest error over the last 60 scored days. They cannot be
negative and they sum to one.
Refit every week, using only forecasts whose official values were already published.
Stale members drop out. A member whose newest run is over 30 hours old is left out and its
weight goes to the rest.
In practice the weight goes where the skill is: the WeatherNext 3 product and the short-range models on
the first day, and the two ECMWF AI products from day 1 outward. A member with a weight of zero is still
shown on the station page, but it does not move the number.
Why blend at allA blend is only worth having if it beats its
best member. The Lead Board tests that at every lead, on the days both forecast.
It has, by 0.1 to 0.7°F, at every lead from the same day to day 13.
4The range
the 10–90% band
Each consensus value carries a 10–90% range: the official value should fall inside it four days
in five. The width is the consensus's own past error at that lead, treated as a bell curve around the
forecast. It widens with lead, from about ±2.5°F tomorrow to about ±7°F at day 12.
Known limitThe width is one number per lead for the whole
country. It does not yet narrow for an easy day in Miami or widen for a frontal day in Denver. Fitting it by
region and by how much the members disagree is the next change to this step.
5From forecast to price
settlement rules, odds and the comparison with the market
A temperature contract pays on a station's high or low for one day. The odds of any outcome are the share
of the forecast range that falls on the paying side, in whole degrees.
Each market reads the thermometer differently, and the difference is worth most of a degree:
Kalshi
The official climate value for the local standard day, the same truth as section 1.
Polymarket
The highest or lowest hourly airport report over the local calendar day.
ForecastEx
The same hourly reports; a contract pays if the value is above (or below) the strike.
The gap
The official high runs about 0.75°F above the highest hourly report, and the official low about 0.55°F below the lowest. The forecast is shifted by each station's measured gap before it is priced on a market that uses the reports.
These rules are re-checked against the settled contracts twice a day, on the Venues
page. Once the day has started, the high cannot finish below what the station has already reported, so the
range is cut off there and the forecast leans on the newest observations and short-range runs.
Three forecasts, for nowThe site currently has three blends
of the same models. The MOS Consensus described above is the station forecast and has the longest
record. The Live Desk prices markets from a second blend built for the 59 market stations, which
also uses observations during the day. Outcome Odds, the Screen and the station pages still use an
older, simpler blend of seven sources. They agree to within a degree or two on most days. A direct test between them on the same stations and days is the next piece of work, and the pages will move to whichever does best.
Positions on this site are paper. No money is traded, and nothing here is advice.
6The record
how it has done, updated daily
Every issue is archived as published and scored against the official value. The table shows the average
miss in °F by lead day: the consensus, the National Blend of Models, and the station's 1991–2020
normal for the date. Smaller is better.
Daily high · average miss, °F
Daily low · average miss, °F
The full board, with every member and the test against the best of them, is the
Lead Board.
Against the markets.The forward record against market prices is on
The Edge. The record is short, and it differs by market: the forecast has done
better against thinly traded prices than against the most active ones.
7What is not done
read the numbers with these in mind
No winter yet. The AI-model products have been scored from spring through early autumn. Cold-season
skill is unmeasured.
Early highs. On days when the high comes before mid-morning, usually behind a front, every product
here does worse than a simple bias correction. A fix is planned.
The range is national (section 4).
The market record is weeks old, not seasons. A result on a few weeks of settled contracts is not a verdict.
The Lead Board's first rows come from re-running each product's history under the no-hindsight rule;
the rows since each product went live are the forecasts exactly as issued. The board says which is which.
Changes to any step are listed on What's New. Source commit
–.