Weather markets are the best-calibrated thing on Polymarket
Almost nobody noticed, but daily temperature markets have quietly become the largest recurring category on Polymarket. A scan of open events on July 27 counts 117 live weather markets carrying about $2.89 million in 24-hour volume, ahead of every geopolitics and macro market on the board. Hong Kong alone did $244,000. Seoul did $193,000. And unlike almost everything else traded on a prediction market, these settle every single day against a thermometer. That makes them the one place where a crowd's forecasting ability can be graded, in public, on a schedule, with money on the line.
So it got graded. Across 189 settled markets in 20 cities, the crowd turns out to be almost perfectly unbiased about tomorrow. On the year, the same crowd is more than twice as aggressive as the statisticians. Both halves of that are worth sitting with.
What these markets actually resolve against
This is the part most people skip, and it is the part that determines whether you can price them at all.
A daily temperature market resolves to the highest reading recorded at one named airport weather station, expressed in whole degrees Celsius. Not a city average, not a forecast, not a regional index. One station, one number, rounded to an integer.
- Seoul settles on Incheon International, station code RKSI.
- Tokyo settles on Haneda, RJTT.
- London settles on London City Airport, EGLC.
- Milan settles on Malpensa, LIMC.
- Shanghai settles on Pudong, ZSPD.
Those stations publish routine aviation weather observations, called METARs, every 20 to 30 minutes, and the temperature field in a METAR is already a whole number of degrees. So the resolution source is not a private feed or a proprietary index. It is the same public observation stream that pilots use, and it updates several times an hour, all day, for free.
That matters more than it sounds. It means the settlement value is being assembled in public, in real time, in front of everyone holding a position. There is no waiting for an agency to publish. By late afternoon local time, most of the answer already exists.
Finding one: on tomorrow, the crowd is unbiased
The scoring method is simple. For each settled market, take the state of the book at roughly the midpoint of its life, before the day's weather had made the answer obvious. The highest-priced bucket at that moment is the favorite. The error is the distance in degrees between the favorite and the bucket that actually won.
Across 189 markets between July 18 and July 27:
- The favorite hit the exact degree 83 times, or 44 percent. These ladders carry roughly eleven buckets, so random guessing lands near 9 percent.
- 52 markets settled hotter than the favorite. 54 settled cooler.
- Mean error: negative 0.04 degrees.
That last number is the finding. A crowd that systematically believed the world runs hotter than the forecast would leave a warm lean in the residuals. A crowd anchored on cool historical averages would leave a cold one. There is no lean. The misses are almost exactly symmetric, and the average of all 189 errors rounds to zero.
An unbiased price is not a curiosity. It is the precondition for treating a market as a measurement rather than a sentiment gauge. A book that leaned two degrees warm would still be interesting, but you would have to correct for it before using it. This one you can read directly.
Finding two: accuracy tracks the weather, not the traders
The aggregate hides an enormous spread. Hit rates by city run from 89 percent down to 20 percent.
- Madrid 89 percent (8 of 9), mean error -0.11
- Paris 78 percent, Shanghai 70 percent, New York 63 percent
- Moscow, Shenzhen, Singapore, Hong Kong: 50 percent
- London, Mexico City: 33 percent. Chengdu, Taipei, Tokyo: 30 percent
- Toronto 22 percent. Beijing, Guangzhou, Seoul: 20 percent
The tempting read is that some crowds are sharper than others. That is not what is happening. Madrid in late July sits under a stable heat dome, where the same number prints day after day and the market simply reads the pattern. Seoul in late July is monsoon season, where a rain band arriving three hours early moves the daily high by several degrees. Beijing and Guangzhou sit in the same convective mess.
So the accuracy spread is tracking the predictability of the underlying weather, which is exactly what a well-functioning price should do. Where the physics is legible, the price is sharp. Where the physics is chaotic, the price spreads across buckets and admits it.
For anyone actually trading these, that is the useful cut. The edge is not in the stable cities. Madrid is solved, and a 89 percent favorite leaves nothing on the table. The uncertainty, and therefore any genuine edge, lives in the volatile ones.
Finding three: the annual market is a different animal
If the daily markets are neutral, the annual one is not.
The market on where 2026 will rank among the hottest years on record carries $3,210,235, which makes it one of the largest climate positions in any public market. As of July 27 it prices:
- First place: 44.5 percent
- Second place: 54.5 percent
- Third place: 2.0 percent
- Fourth or lower: under half a percent, combined
Read that distribution again. The market has assigned 99 percent of its probability mass to the two hottest outcomes in recorded history, and left essentially nothing for the possibility that this year is merely warm.
Now put the institutions next to it. The World Meteorological Organization forecasts 2026 at 1.44 degrees above the 1850 to 1900 baseline, placing it among the four hottest years, near but probably not above the 2024 record of 1.55 degrees. Berkeley Earth's running estimate puts the year at 1.47 degrees, with a 62 percent chance of second place and a 19 percent chance of first.
The market says 44.5. The statistical model says 19.
The nearby markets carry the same tilt. A separate book prices a 91 percent chance that July 2026 is the hottest July ever recorded, and another gives 94.5 percent to at least one month of 2026 setting a record. A third prices this July's temperature anomaly at 1.20 to 1.24 degrees with 87.5 percent confidence, which is a remarkably tight band for a variable that will not be published for weeks.
Two readings, and the data does not settle it
The first reading is anchoring. The last two years delivered records, recency is a powerful distortion, and traders extrapolate trends that statistical models deliberately damp. Under this reading the annual book is simply hot and gets corrected in January.
The second is that the market is absorbing something the models update too slowly to price. A strong El Nino is developing, with central estimates pointing at roughly 2.2 degrees of Nino-region warming by September. Markets update continuously. Seasonal forecasts update on a schedule. When conditions move fast, a continuously repriced book can lead a periodically republished model, and that is precisely the situation prediction markets are supposed to handle well.
The evidence available today does not choose between them. What it does do is make the question scored. In January this stops being a debate and becomes a settled fact, and whichever reading was right will be a matter of public record.
Why the calibration result is the important one
Heat has become a first-order economic variable. Heat-related deaths in the United States rose 117 percent between 1999 and 2023. A single heatwave lasting five days or longer costs a country roughly 0.2 percent of annual labour productivity, with effects that linger for up to two years. Lost work capacity from heat is projected to cost the global economy $2.4 trillion by 2030, and Europe's largest economies alone face an estimated $638 billion in heat-driven losses over the same period.
Construction, agriculture, logistics and energy carry most of that exposure. Every one of those sectors hedges fuel, currency and rates as routine business. Almost none of them can hedge the variable now doing the damage. Weather derivatives exist in institutional form, but they are large, bespoke and effectively closed.
Nobody in those industries needs to be told the planet is warming. What they have never had is a number: a live, public, continuously updating price for how hot it gets in their city this week. A price with a mean error of negative 0.04 degrees is the first version of that number that is worth referencing.
What to watch next
- Does the calibration survive a season change? July in the northern hemisphere is one specific regime. If the same symmetric error distribution shows up in October and January, the result is structural. If it does not, it was seasonal luck.
- Does the annual market converge toward the science, or the other way around? This resolves in January with no ambiguity about who was right.
- Does the city accuracy spread persist? If Madrid holds near 89 and Seoul near 20 over a longer window, that gap is structural, and it defines where a real edge lives.
Frequently asked
Are weather prediction markets accurate?
On this sample, yes. The favorite hit the exact degree bucket 44 percent of the time across 189 settled markets, against roughly 9 percent for random guessing, with a mean error of negative 0.04 degrees. Accuracy ranges from 89 percent in Madrid to 20 percent in Seoul and Beijing.
What do these markets settle against?
The official high at one named airport station, in whole degrees Celsius. Seoul resolves on Incheon, Tokyo on Haneda, London on London City, Milan on Malpensa. The underlying data is the public METAR observation stream, issued every 20 to 30 minutes.
Do traders overestimate how hot it gets?
Not on the daily markets. 52 settled hotter than the favorite and 54 cooler, which is as close to unbiased as this sample size can demonstrate. At the annual level it reverses sharply: 44.5 percent priced on a record year against a 19 percent statistical estimate.
Where is the edge?
In the volatile cities, not the stable ones. Madrid under a heat dome is already solved. Monsoon and convective regimes such as Seoul, Beijing and Guangzhou are where the market is genuinely uncertain.
Where to read them
These markets settle on Polymarket's order book. SmartX is a terminal on top of that book which puts the live probability next to what larger positions are actually holding, which matters on a question like this one where the sentiment is loud and the positioning is quiet.
Everyone checks the forecast. Almost nobody prices it. The gap between those two habits is where the information lives.