What the market asked
The Polymarket event was straightforward: “Highest temperature in Wellington on August 26?” Resolution turned on a single data source—the highest whole-degree Celsius value in the Temp column of the NOAA WRH timeseries for station NZWN (Wellington International Airport) on calendar day 26 August 2026. Markets of this type typically fragment into discrete temperature buckets (e.g., 14°C, 15°C, 16°C, and so on), each trading as its own yes/no contract.
What the council concluded
On 25 August 2026 the council (Grok as chairman, with GPT; Claude and Gemini did not participate) issued a BUY_NO recommendation at an implied 23% probability and 68% confidence. Their stated reasoning was narrow and mechanical:
“Council recommends BUY_NO (NO). Probability: 23%, Confidence: 68%. Resolution is determined solely by the single highest whole-degree Celsius value appearing in the Temp column of the NOAA WRH timeseries for station NZWN (Wellington International Airport) for the calendar day 26 August 2026.”
Under the grade basis of “most_likely,” the council’s standout call was the 15°C bucket at 34% probability. That became the graded pick: “Will the highest temperature in Wellington be 15°C on August 26?”
What the market was pricing
The council’s internal distribution put 15°C first at 34%, with the overall BUY_NO stance reflecting a view that whatever specific affirmative contract was under discussion sat well below even-odds. No richer order-book snapshot is preserved in the graded record, so the only quantitative anchors we have are the council’s own 23% action probability and the 34% weight it assigned to 15°C. In short, the models collectively treated 15°C as the modal outcome while still expressing meaningful residual uncertainty.
What actually happened
The verified winning outcome was 16°C—“Will the highest temperature in Wellington be 16°C on August 26?” The council’s graded pick of 15°C was therefore incorrect. council_was_correct is false; the Brier score on the graded probability came in at 0.1156. A one-degree miss on a whole-degree NOAA reading is about as clean a binary failure as weather markets produce: the data source was unambiguous, the calendar day was fixed, and no alternative station or averaging rule came into play.
Takeaway
Even when the resolution criteria are crystal-clear and the models correctly identify the governing data feed, short-range temperature forecasts remain noisy at the whole-degree level. The council’s 34% modal call on 15°C left substantial probability mass on neighboring buckets; 16°C simply realized that residual. The episode is a reminder that “most likely” is not the same as “likely enough,” and that a Brier score in the low teens still registers as a clear miss when the graded contract loses. Honesty about the error is the point of publishing the grade.
AI-generated analysis for informational purposes only. Not financial advice.