Temperature markets priced as well as public forecast models
Across 1,977 daily-temperature events on Polymarket, trading against ECMWF, GFS and Open-Meteo forecasts showed no edge: the best case was +$22 (±17) on 867 trades at two days out, and closer to resolution results were negative.
What we asked
Polymarket lists daily "highest temperature" and "lowest temperature" markets for many cities. Each event is split into baskets, such as a one-degree range, and one basket wins. Public weather models are free. We asked whether the gap between a published model forecast and a basket's price, one or two days before resolution, was large enough to trade profitably.
Data
- 1,977 closed daily events titled "highest/lowest temperature in [city]" from 23 September to 14 October 2026, across 51 cities and 21,747 baskets, with $46.5M of volume over the three weeks.
- Price history of each basket's YES token from the CLOB API. The price used is the mid, the midpoint between best bid and best ask, sampled every 30 minutes.
- Hourly forecasts from three models (ECMWF at 0.25 degrees, GFS and Open-Meteo's best_match) at lead times of 0 to 3 days, from the Open-Meteo previous-runs service. The forecast point is the city's geocoded location, not the weather station that settles the market.
- The truth is the basket that won. Forecast error was calibrated per city and per type (maximum or minimum) with leave-one-out, so a day is never used to calibrate its own forecast.
Forecast error across all cities, in degrees Fahrenheit: ECMWF had a mean absolute error of 2.1 on the observation day, 2.5 one day ahead and 2.7 two days ahead. GFS was worse by about 0.5. City biases reached ±3 degrees; for example, LaGuardia runs warmer than the model.
Result
The market was well calibrated on liquid baskets, measured 12 hours before local midnight of the observation day: a price of 0.07 won 13% of the time, 0.15 won 19%, 0.25 won 30%, 0.35 won 36%, 0.45 won 48% and 0.54 won 55%. Cheap baskets were slightly underpriced.
The test rule: buy a side when the model disagrees with the price by a margin of 0.10, pay the price plus 2 cents, and hold to resolution. Results are total profit on $1 per trade, with the uncertainty the report attaches to each total.
| Entry | Side | Trades | Win rate | Total on $1 per trade |
|---|---|---|---|---|
| 2 days ahead (36 h before) | YES | 273 | 12% | +$16 (±51) |
| 2 days ahead (36 h before) | NO | 867 | 69% | +$22 (±17) |
| 1 day ahead (12 h before) | YES | 686 | 9% | -$136 (±76) |
| 1 day ahead (12 h before) | NO | 1,354 | 61% | -$40 (±24) |
| Previous day's forecast, observation day +6 h | YES | 721 | 6% | -$350 (±69) |
| Previous day's forecast, observation day +6 h | NO | 1,378 | 54% | -$72 (±29) |
The two-day results are inside their uncertainty bands. The later entries are negative.
Why the edge disappears
On liquid baskets the price was no worse than the public models at any lead time we tried, and the closer to resolution, the further the market moved ahead of the forecast. We attribute this to traders using observations that arrive during the day, which forecasts do not yet reflect; we did not test that explanation separately. Beating the market would require a forecast better than ECMWF for the specific settlement station, which the models we used do not provide.
Limits of this measurement
- A data artifact we removed. The mid on thin baskets is unreliable: with a best bid of 0.001, the mid often reads 0.3 to 0.5. The sum of mids across an event was 3.0 when volume was under $5,000 and 1.03 at $50,000 or more. Our first run showed about +$2,000 on NO bets purely because of this. The final figures use only liquid cases: events with at least $20,000 of volume and baskets with at least $2,000 (about 960 events).
- A leak we excluded. A "same-day forecast" entry showed +$250, but the archive's last run for a past date includes model runs made after the entry moment. We did not count it.
- Forecasts are taken at the city's geocoded point, not the settlement station.
- The period is three weeks, so seasonal behaviour and weather regimes outside this window are not covered.
- We tested one rule with one margin (0.10) and a fixed 2-cent entry cost. Other rules and thresholds were not tested.
What it means for a trader
On liquid temperature markets, the price already contained what the ECMWF, GFS and Open-Meteo forecasts knew. The only positive totals we found sat within their own uncertainty, and entries made later were negative. We found no evidence of a mispricing that a free public forecast could capture.
Sources
- sourcePolymarket Gamma API
- sourcePolymarket CLOB API
- sourceOpen-Meteo
Polytrading is an analytics tool. It does not trade for you, hold your keys, or promise returns.