← All posts
Weather·August 27, 2026Council was wrong

AI council favored 29°C for Madrid’s August 26 high — 30°C won

What the market asked

The Polymarket event was straightforward on paper: what would be the highest temperature in Madrid on August 26? Resolution rested on a single, station-specific number—the highest integer °C in the NOAA/NWS Western Region Headquarters timeseries Temp column for LEMD on the local calendar day of 26 August 2026, once the data had stabilized. Each integer degree was effectively its own outcome, so the market was a narrow weather call rather than a broad directional bet.

What the council concluded

Only Grok and GPT participated; Grok acted as chairman. The council’s formal action was SKIP, with an overall probability reading of 41% and confidence of 68%. Its stated most-likely outcome was “Will the highest temperature in Madrid be 29°C on August 26?” at 42%. That pick became the graded line.

The published reasoning was explicit about both the resolution rules and the decision to stand aside:

Council recommends SKIP. Confidence: 68%. Resolution criteria are unambiguous and station-specific: the market pays on the single highest integer °C appearing in the NOAA/NWS Western Region Headquarters timeseries Temp column for LEMD on 26 August 2026 (local calendar day), after data stabilize once…

In short, the models understood the measurement standard, saw no edge sharp enough to act, and still surfaced 29°C as the modal forecast when forced to name a most-likely bucket.

What the market itself was pricing

The council’s internal figures put the leading outcome (29°C) near 42% and the broader probability mass around 41%. That is a relatively flat distribution for a single-degree weather market: the models did not treat any one temperature as dominant, which is consistent with the SKIP recommendation. No stronger consensus price for 30°C (or any other degree) appears in the graded record, so the council’s own 29°C lean is the clearest signal we have of how the AI side was reading the day-ahead setup.

What actually happened

The market resolved to 30°C: “Will the highest temperature in Madrid be 30°C on August 26?” The council’s graded pick of 29°C was therefore incorrect. council_was_correct is false; the Brier score on the graded basis (most-likely) is 0.1764, which matches a 42% probability assigned to an outcome that did not occur.

A one-degree miss on a station-specific daily max is common in short-range temperature markets. The council had flagged the resolution source correctly and had chosen not to bet; the error surfaces only because the report is graded on the most-likely label even when the action is SKIP.

Takeaway

This is a clean illustration of why Prediction Council publishes losses. The models correctly identified unambiguous, station-level criteria and prudently recommended SKIP at 68% confidence. Their residual modal call—29°C at 42%—was simply off by one degree from the verified LEMD high. In thin weather markets, that gap is enough to turn a cautious pass into a graded miss. The Brier number is modest, the process was transparent, and the record now shows the miss rather than burying it. Honesty about near-miss temperature calls is the point of the exercise.

AI-generated analysis for informational purposes only. Not financial advice.

Every council report is graded against the real outcome and published — the good calls and the bad ones.

AI council favored 29°C for Madrid’s August 26 high — 30°C won — Prediction Council