What the market asked
The Polymarket event was straightforward: what would be the highest temperature in Manila on June 8? The market offered discrete outcomes in whole degrees Celsius. Resolution depended on the official daily high, making it a classic short-horizon weather contract where late-day observations and forecast consensus usually dominate.
What the council concluded
All four models—Gemini, Grok, Claude, and GPT—participated, with Grok acting as chairman. The council’s graded pick (most-likely basis) was 33°C, assigned a 27% probability. Separately it issued a high-confidence BUY_NO at a 1% implied probability on the complementary side of the distribution. Confidence on the overall stance sat at 95%.
The published reasoning stated:
“Council recommends BUY_NO (NO). Probability: 1%, Confidence: 95%. Gemini’s real-time METAR evidence (33°C already recorded) combined with shifted council views (Grok/Claude now favor 33-34°C) and external forecasts converging on 31-34°C confirm the daily high will e”
In short, live station data already showing 33°C, plus model agreement clustering in the low-to-mid 30s, led the group to treat 33°C as the mode and to treat anything outside the narrow band as extremely unlikely.
How the probabilities looked
The council’s own distribution put only 27% on its modal bin (33°C). That is not a high absolute probability; weather markets with 1°C resolution routinely spread mass across several adjacent outcomes. The 1% BUY_NO figure and 95% confidence reflected conviction that the high would land inside the 31–34°C envelope already signaled by METAR and the forecasts, not certainty on any single degree. No separate exchange mid-market snapshot is recorded in the report, so the council’s 27% stands as the figure that was graded.
What actually happened
The winning outcome was 32°C. The council’s most-likely pick was therefore wrong. The Brier score on the graded prediction was 0.0729—respectable in absolute terms because probability mass had been spread, yet still a miss on the mode that the grading rule evaluated. One degree separated the call from reality.
Why the miss occurred
Two factors stand out. First, the METAR print of 33°C was treated as strong evidence that the daily maximum had already been reached or would hold; in practice the official high settled one degree lower, illustrating how station reports, rounding, and the precise definition of “daily high” can diverge late in the day. Second, once the models coalesced around 33–34°C, the residual probability on 32°C was under-weighted relative to the eventual outcome. Forecast convergence reduced perceived variance more than the data justified. The truncated reasoning itself shows the council stopping at the observation of 33°C rather than fully stress-testing the possibility that the official figure could still print lower.
Takeaway
A tight real-time observation plus multi-model agreement is usually enough for weather markets; here it was not. When bins are only 1°C wide, even high-confidence “already reached” evidence can leave the modal call one tick off. The council’s low Brier score shows the broader distribution was not reckless, but the graded most-likely outcome was simply incorrect. Calibration on the exact degree remains the harder problem.
AI-generated analysis for informational purposes only. Not financial advice.