You are not trading the weather. You are trading one thermometer.
This is the mistake that costs new traders the most money in temperature markets, and it does not feel like a mistake while you are making it. You check a forecast, it is a good forecast from a serious provider, it says 34 degrees, and you buy 34. The day is 34 degrees. You lose. The market did not resolve against the weather. It resolved against one instrument at one airport, as reported by one company, rounded to whole degrees, over a calendar day defined in a timezone you may not have checked.
What actually resolves the market
Every daily temperature market names four things in its rulebook, and only these four count.
A source. Most of these markets settle on Weather Underground's daily history page. A smaller group settles on the National Weather Service. The rulebook says which. That provider's posted figure is the answer even in the rare case where it later looks wrong.
A station. Not the city, a specific airport. London settles on London City, not Heathrow. San Francisco settles at SFO, which sits on the bay. We wrote up the station question separately because it is the single biggest trap in the category.
A precision. Usually whole degrees, so 33.6 is 34 and the bucket boundaries sit on half degrees.
A day. The station's local calendar day, which is why these markets close at local midnight rather than at a shared hour.
The forecast providers, and what each one is for
None of the sources below is bad. They are built for different jobs, and only one of them is the job the market is scoring.
| Source | What it forecasts | Use for a temperature market |
|---|---|---|
| Weather Underground | Station level, and it publishes the daily history page most of these markets resolve against | The one that pays. Read the same station the rulebook names |
| NWS / NOAA | Official US forecasts and observations, station level | Settlement source for a minority of markets. Check the rulebook |
| ECMWF, ICON, GFS | Raw numerical models, grid cells rather than stations | Useful for direction and for spotting a day where models disagree. Not the settling number |
| Windy, Ventusky | Model output rendered as maps | Good for seeing a front arrive. No station level daily maximum |
| AccuWeather, phone apps | A city, blended and smoothed for consumers | The furthest from resolution. A city forecast is not a station forecast |
Why the gap is bigger than the bucket
The reason a city forecast fails here is geography rather than skill. Settlement stations are airports, and airports get built where cities are not: on coasts, on estuaries, on flat open ground outside town. Those places run cooler than the built up area that shares their name.
We measured the size of it across fourteen cities. Taking the hottest reporting station in the field around a city, the kind of number a trader notices and reacts to, it averaged 1.6C above what the settlement station actually printed, and it landed in the settling bucket only 16 percent of the time. San Francisco was the extreme: the station that settles ran nearly five degrees below its nearest neighbour, because marine air reaches the bay and does not reach inland. The full table by city is published here.
On a ladder where each rung is one degree, a source that is systematically a degree and a half off is not a slightly worse forecast. It is pointing at a different question.
Better model, wrong trade
There is a subtler version of this that catches experienced people. Suppose you run a genuinely superior model and it says the true maximum at that airport will be 34.4. The provider posts 34. You were right and you are paid. Now suppose the true maximum is 34.5 and the provider posts 34 anyway, because their station reported on a schedule that missed the peak by eight minutes.
Resolution does not care. You are not forecasting the atmosphere, you are forecasting what one provider will publish about the atmosphere, and those are almost the same thing but not quite. The gap is small and it is not zero, and it always favours reading the same source the market reads.
The practical version
Three rules cover most of it.
Read the settling source first, and read the station. Everything else is context. If your primary screen is a consumer weather app, you are one step removed from the question before you start.
Use the other providers for disagreement, not for level. When ECMWF, ICON and GFS spread out across three degrees, that is a genuine signal that the day is uncertain and the market should be wider than it is. When they agree, tightness is earned. Neither case tells you the settling number.
Know what time it is. The two forecasts on the settling source itself, the hourly curve and the daily high, swap places about three hours before the peak. Before the crossover, weight the curve. After it, weight the daily high.
A board that does this for you. SmartX runs a free public board at weather.smartx.io covering fourteen cities. It reads the settlement station for each market, puts a probability on every bucket against the traded price, and shows the derivation underneath: which anchors moved the number, what the nearby field is doing, and how far the station has climbed today. No account needed to read it.
Caveats
Rulebooks change and new cities get added with different sources. Read the resolution text on the specific market rather than trusting a rule of thumb, including this one.
The 1.6C figure comes from 1,824 readings across fourteen cities in the final six hours before close, taken over a few summer weeks. The direction is geography and should hold. The size will move with the season.
Research published by PredictionSignal. Nothing here is financial advice. Market prices move and any figure quoted was true when measured.