Same data, two guidance systems: switching to the National Blend moved up to 67 points of probability

What happened

At 18:31 UTC on Oct. 3, PropBetEdge published probability distributions for Sunday's high temperature in 7 cities using model version 1.0, which reads GFS MOS station guidance. At 18:45 UTC it published them again with version 1.1, which reads the National Blend of Models. Both sets used the same guidance cycle and the same data cutoff — Oct 3 17:00Z. Nothing about the weather changed between the two. Only the guidance did.

Because every forecast is an immutable snapshot, the pair is a controlled experiment that already happened in public: the same question, the same information, two guidance systems. Here is what it shows.

The guidance gap

The two guidance systems agreed within a degree in Miami and Chicago. They disagreed by four or more degrees in Philadelphia (GFS MOS 65°F, blend 69°F), Denver (GFS MOS 78°F, blend 83°F) and Austin (GFS MOS 79°F, blend 83°F). For contracts sold in two-degree buckets, four degrees is two whole buckets.

How much of each forecast moved

A useful single number is how much probability changed hands between buckets — the share of the distribution that moved (total variation distance). Zero means identical forecasts; 100 means no overlap at all.

Austin moved most: 67 points of probability changed buckets. Chicago moved least, 10 points. In all 7 cities the most likely outcome itself changed.

Austin, bucket by bucket

Austin shows the mechanism clearly. GFS MOS forecast 79°F; the blend forecast 83°F. Version 1.0 put 69% on “79° or below”; version 1.1, from the same data, put 9% there and moved its weight to 84° to 85° (32%).

Why the blend is the default

Version 1.1 was not adopted on a hunch. Both versions were evaluated on the same chronological holdout — 56,668 station-days from 2025-07-01 to 2026-09-30 — scoring the probability each assigned to the exact reported high. The National Blend version's log loss was 2.4312; the GFS MOS version's was 2.5827 (lower is better). A simple average of the two scored 2.4252 on log loss but worse on Brier score (0.13901 vs 0.13827), so the model uses the blend alone and keeps GFS MOS only as a fallback.

It is tempting to ask which version agreed with the market. Of the 5 cities where every bucket had a two-sided quote, v1.1 was closer to the market's distribution in 4. That is not evidence of accuracy — the market is not the outcome. The scored record, built from snapshots fixed before each outcome, is the only place that question gets answered.

The fallback, and the fix

The archive also shows what went wrong. In all 7 cities, a GFS-only version was published again at 19:15 UTC and 19:30 UTC, after the engine failed to fetch the National Blend run, before the blend version returned on the next cycle. The same data cutoff produced public numbers that flipped back and forth — exactly the kind of change that is not a forecast change. The engine now holds a station when a guidance fetch fails instead of falling back. The snapshots stay in the archive; nothing is rewritten. For scoring, the FIRST_PUBLISHED role for these contracts belongs to the v1.0 forecast — it was first — and FINAL_PRE_RESOLUTION to the last forecast before each city's climate day begins.

Takeaway

For daily temperature contracts, the choice of guidance can matter as much as the weather. When a PropBetEdge probability changes, the snapshot archive shows whether new data arrived or the model changed — and this page shows the size of the second effect on a day when it was the only thing that changed.