Kalshi's weather markets beat the public forecast in six of seven cities

In September, Alexander Crosier published a study of Kalshi’s markets on the next day’s high temperature in seven US cities. Each day the exchange lists a ladder of contracts on two-degree bands, and their prices add up to a forecast. The study found that an hour after the market opens, that forecast beats the National Blend of Models, the best single public forecast, by about 10 percent in root-mean-square error.

I wanted to know whether the result holds when I rebuild it from public data with my own code, and how it looks after the study’s data ends on 12 August 2026.

The method

I fixed the test and its pass rule before I read any data. For each city and day, I took the market’s prices at 11:00 Eastern on the day before, applied the paper’s liquidity screens and turned the ladder into one expected high. The forecast it races is the newest National Blend bulletin issued at least an hour earlier. The truth is the National Weather Service’s daily climate report, and days with an overnight high are out, as in the paper.

The result

Period	City-days	Market error	Forecast error	Market’s lead
The paper’s, 1 Jan 2025 to 12 Aug 2026	3,655	2.43°F	2.79°F	12.7%
After the paper, 13 Aug to 22 Oct 2026	454	1.78°F	2.46°F	27.7%

The paper gives 9.8 percent at this hour. My lead in its period is far larger than its standard error, with a Diebold-Mariano t of −7.98 on the daily differences. The market was closer to the real high on 56 percent of city-days.

City	Market’s lead
New York	18.0%
Los Angeles	17.7%
Chicago	16.3%
Philadelphia	11.6%
Austin	10.3%
Denver	10.1%
Miami	−1.4%

Miami is the one city where the forecast is as good as the market. The paper also finds its smallest lead there.

What could inflate it

The free archive keeps four National Blend runs a day, but the forecast is issued every hour. At 11:00 Eastern my bulletin is two to three hours old, and the paper uses the freshest one. An older forecast makes the market’s lead larger, so this can explain part of 12.7 percent against 9.8. I have not measured how much.

The months after the paper are a description, not a test. In those months Kalshi settles these markets on The Weather Company’s reading, and my truth is still the Weather Service’s report.

What was not tested

The paper’s other hours, its five other forecast products, the best blend of market and forecast, its years before 2025, and its finding that the market does not move when a new forecast lands. No second person has reviewed how I read the claims out of the paper, and its code is not public.