Three counting mistakes that put you in the wrong temperature bucket (checked on 1,153 resolved market-days)
17 November 2026 · ISOTHERM
Daily-high temperature markets look simple: one airport, one day, one number. But the market does not resolve on "how hot it was." It resolves on a specific set of airport reports, read by a specific rule. Count the reports a little differently and you land one bucket away, with a number that looks perfectly reasonable.
I rebuilt the daily high for every city with a Polymarket daily-high market from public airport reports, then compared my number with the range each market actually resolved to. Period: markets that closed between 17 October and 16 November 2026. That is 1,153 city-days across 48 cities. I only read which range won. No prices are involved anywhere in this.
Getting to 1,153 out of 1,153 took three corrections. Each one is a mistake I made first.
1. Counting only the top-of-the-hour reports
Many airports outside the United States publish a routine report (METAR) every 30 minutes, not every hour. Some archives file the half-hour ones under a different report type, so a query for "routine reports" quietly returns only the hourly ones.
If the day's peak falls on a half-hour report, you miss it.
- Hourly reports only: 1,102 of 1,153 city-days in the resolved range (95.6%)
- With half-hour routine reports: the 51 misses disappear
2. Counting US special reports
So I included the second report type everywhere. That fixed the 51 and broke 7 others, all at US airports.
US airports publish one routine report an hour. In between they issue special reports (SPECI) when conditions change quickly. Those specials can carry a higher temperature than any routine report that day, and they do not count. The trap: the archive I used strips the word SPECI from US reports, so "drop every line that starts with SPECI" removes nothing.
- Both report types everywhere: 1,146 of 1,153 (99.4%)
- Half-hour reports outside the US, hourly routine reports only in the US: 1,153 of 1,153
3. Truncating instead of rounding
A common explanation says the reading is truncated: 10.9 becomes 10. I tested it on the same days, using the 0.1 °C value from the report remarks where one exists, converting to °F first for °F markets.
- Rounded half up: 1,153 of 1,153
- Truncated: 1,094 of 1,153 (94.9%)
Truncating put the number in the wrong range on 59 city-days, about one day in twenty. At airports that only report whole degrees the two rules give the same answer, so the gap is concentrated where 0.1-degree values exist.
The rule that matched every day
- Take the airport named in the market rules and its local calendar day.
- Count routine reports. Outside the US that includes the half-hour ones. In the US, hourly routine reports only. Special reports never count.
- Use the 0.1 °C remark value when the report has one, otherwise the whole-degree temperature.
- Take the highest value. Convert to °F first if the market is in °F. Round half up to a whole degree.
What this does not tell you
Thirty days is thirty days. Rules change, archives get corrected after the fact, and a market can resolve on something other than what its rules suggest. So the comparison runs every week and every city-day is published, including any day that differs:
https://isotherm-82f.pages.dev/check.html
There is also a free single-file Python script and a 30-day Tokyo sample if you want to run the rounding check yourself:
https://isotherm-82f.pages.dev/kit.html
This is information about how numbers are counted. It is not financial or betting advice, and ISOTHERM is not affiliated with any exchange or prediction market.