How to trade weather prediction markets
Daily temperature markets are the largest recurring category in prediction markets, at roughly $2.9M a day across 117 markets. They look like the easiest thing on the board, because everyone has an opinion about the weather and everyone can check a forecast. They are not easy, and the reasons are specific rather than vague. This guide is the whole checklist, in the order it matters, with the measurements behind each step.
1. Read the rulebook before the forecast
A temperature market resolves against four fields, and getting any of them wrong can make you right about the weather and wrong about the trade.
The station. One named airport, not the city. London settles on London City, not Heathrow. Denver does not settle at Denver International and Dallas does not settle at DFW. We read all thirty nine rulebooks and four of them settle somewhere most people would not guess.
The provider. Most settle on Weather Underground's daily history page for that station. A minority settle on the National Weather Service. Whichever it is, that company's posted number is the answer.
The rounding. Usually whole degrees, so bucket edges sit on half degrees and 33.5 goes up.
The day. The station's own local calendar day, which is why these close at local midnight rather than at a shared hour. A market that looks like it has fourteen hours left may have four.
2. Work out how much of the day is already over
This is the step almost nobody takes and it changes everything about what you are doing. Across 13,185 city-days at 33 stations, the day's high had already been recorded by 1pm local in 53.7 percent of cases, by 3pm in 85.7 percent, and by 4pm in 93.5 percent.
So for most of the trading window you are not forecasting. You are reading a result that has already happened and waiting for the rest of the book to notice. The hour by hour curve is here, and it moves by city: Los Angeles locks in about four hours before Paris.
3. Use the forecast that is right for the hour
The settling provider gives you two forecasts for the same question, an hourly curve and a daily high, and they usually disagree by about a degree. Which one is more accurate depends on how close you are to the day's peak, and they swap places.
| Distance from the peak | Hourly curve lands in bucket | Daily high lands in bucket |
|---|---|---|
| More than 6h before | 43% | 12% |
| 4h to 3h before | 41% | 45% |
| 2h to 1h before | 24% | 49% |
| 1h before to peak | 40% | 58% |
In the morning the hourly curve is clearly better. Approaching the peak the daily high overtakes it and finishes more than twice as accurate. The curve does not get worse so much as it stops getting better, because it reports values on the hour and the true maximum rarely lands on one. Full measurement here.
4. Read the nearby airports, but correct them
The other stations around a city update on the same schedule and are free to read, so it is tempting to treat a neighbour ticking up as a preview. Measured across 1,824 readings in fourteen cities, the hottest neighbouring station averaged 1.6C above what the settlement station finally printed and matched its bucket only 16 percent of the time.
The offset is city specific and it comes from where the airports are rather than from today's weather. San Francisco's settlement station runs nearly five degrees below its nearest neighbour because it sits on the bay. The table for all fourteen cities is here.
Use the field for direction. When every station around a city turns over and starts falling, the settlement station usually follows, and knowing that an hour early is worth something. Do not use the field for level.
5. Check the book before you size
Temperature ladders open wide and tighten. We measured live spreads from open to close: they start at 14 to 16 cents and come in to one or two, but not evenly across the ladder. The buckets nobody believes in stay wide all day, which matters because those are exactly the ones a contrarian wants.
Size to the spread you can get out at, not the one you got in at.
The four traps, in one place
Trading the city instead of the station. Consumer weather apps forecast a city. The market settles on an airport that is often cooler than the city that shares its name.
Trading a forecast after the day is decided. After about 4pm local you are usually trading a fact, and paying forecast prices for it is the wrong way round.
Reading the neighbours as level. A degree and a half of average offset is bigger than the bucket you are trading.
Assuming close means midnight your time. These close at the station's local midnight. Check the timezone.
A free board that runs this checklist. SmartX publishes weather.smartx.io, covering fourteen cities. For each market it reads the settlement station, puts a probability on every bucket against the traded price, and shows the working underneath: the anchors behind the number, how far the station has climbed today, and what the nearby field is doing. It is free to read with no account, which makes it the fastest way to do steps two through four without opening five tabs.
Is there actually an edge here
Honestly, less than the category's reputation suggests. Temperature markets are the best calibrated recurring category we have measured, which is a compliment to the market and a problem for anyone trying to beat it. Prices track settlement closely and the obvious mispricings get taken quickly.
What remains is structural rather than meteorological. Knowing which station settles the market, knowing how much of the day has already happened, and knowing which of the two forecasts to weight at the hour you are looking, are all things the market prices imperfectly because most participants skip them. Having a better opinion about the weather than the crowd is a much harder way to make the same money.