The market prices a 99°F day at LAX at 67%. Our model says 8%.

What the contract actually asks

Kalshi's contract settles on the daily maximum temperature The Weather Company reports for Los Angeles International Airport (CLILAX), for the climate day of Oct. 4 measured in local standard time — Oct 4 08:00Z to Oct 5 08:00Z. “99° or above” pays if the reported high is 99°F or more. PropBetEdge verifies every settlement against the National Weather Service's Daily Climate Report for the same station, stored separately.

That precision matters here. The airport sits on the coast; a few degrees of marine influence separate an ordinary warm day from a 99-degree reading, and the contract does not care what happened anywhere else in the basin.

Where the forecasts disagree

Three forecast systems were in the record at the model's data cutoff (Oct 3 17:00Z). The National Blend of Models — the guidance the current model version uses — put the Oct. 4 high at 95°F with a blend spread of ±3°F. GFS MOS, the older station guidance, said 97°F. The National Weather Service's own gridded forecast, issued Oct 3 19:21Z, showed 99°F — stored for context, not used as a model input. The 1991–2020 normal for the date is 76°F; every system expects an unusually hot day.

The model does not average these numbers. It takes the National Blend's high and asks how far the reported temperature has historically landed from that guidance at this exact station and lead time: 2.8°F is one standard deviation of that error over 1,819 past forecasts. Reaching 99°F from a 95°F blend needs a miss of 4°F or more in the warm direction — something the station's own record says happens rarely. That is the whole of the 8%.

The full distribution

The model's most likely bucket is 95° to 96° at 36%. The market's is 99° or above at 67%. In other words, the two are not arguing about whether it will be hot; they disagree about the last few degrees.

How the number got here

The first forecast published for this contract, at 18:31 UTC on Oct. 3, came from model version 1.0, which uses GFS MOS: 24%. Version 1.1, built on the National Blend after it beat GFS MOS in a holdout of 56,668 station-days (log loss 2.4312 vs 2.5827), published 8% at 18:45 UTC from the same data cutoff. The difference is the guidance, not new weather information — the subject of a separate PropBetEdge analysis.

One snapshot needs a footnote. At 19:15 UTC the engine could not fetch the National Blend run and re-published the GFS-only version (24%) before the blend returned. That record stays in the archive — nothing is deleted — and the engine now holds a station when a guidance fetch fails rather than downgrading the model.

The case for the market

A 59-point gap is not a signal to trust the model. Several things in the record point the market's way. The Weather Service's own forecast sits at 99°F, right on the line. GFS MOS is 2°F warmer than the blend. And the model's own documented limitations apply squarely: it is a pre-window forecast with no current-conditions input, and its research notes record that guidance can run cold in unusual heat regimes. A market can price a heat event's persistence; this model, by design, cannot see it.

The model's case is narrower and historical: at this station, blend misses of 4°F or more in the warm direction are uncommon. Which view is better calibrated is exactly what the scored record exists to answer — one contract will not settle it.

How this will be scored

The snapshots that count were fixed by rule before the outcome: FIRST_PUBLISHED (the 24% GFS-based forecast), and FINAL_PRE_RESOLUTION — the last forecast captured before the climate day began at Oct 4 08:00Z. No T-minus-24-hour snapshot exists, because the first forecast came less than a day before the window. Both PBE and the market are scored on the same snapshot with Brier score and log loss once the official value is stored.