How accurate is the model?
Every prediction is graded against the number that actually settles the market — the official NWS Climate Report (CLI) daily high. Not a blend of public models, not our own confidence. Here is the measured record over the last 120 days across 9 markets.
The morning call, when Kalshi brackets are widest and the edge is largest. We don't claim to out-forecast the National Weather Service here — on settled days our morning error matches the NWS daily high almost exactly. What we add is a calibrated probability across the brackets, and then the convergence above as the day fills in.
Why this window matters: at the midday read the book is still undecided — its favorite bracket averages 61¢ and settles right 59% of the time (973 days). A view is worth having while the price still moves.
Degrees are only half the story: what pays is naming the bracket that settles. Below is how often the bracket we named contained the official high, split by when we named it — and what that bracket cost at the time. A bracket caught late, at 90¢, is worth much less than the same bracket caught at dawn.
The remaining 5% we never called correctly — counted here rather than quietly dropped. Measured over 439 settled days.
The same day re-read at noon, once the morning observations have landed but the afternoon peak is still ahead. Sharper than dawn — but the book has started to move too, so treat it as a revision, not a fresh edge.
| Market | Morning miss | Midday miss | Within ±1°F (midday) | Market favorite (midday) | Days |
|---|---|---|---|---|---|
| HoustonKHOU | 1.34°F | 1.31°F | 59% | 60¢ | 119 |
| AtlantaKATL | 1.44°F | 1.42°F | 48% | 53¢ | 119 |
| San FranciscoKSFO | 2.28°F | 2.27°F | 29% | 53¢ | 119 |
| MiamiKMIA | 1.19°F | 1.26°F | 55% | 60¢ | 118 |
| Los AngelesKLAX | 2.13°F | 1.20°F | 63% | 74¢ | 118 |
| New YorkKNYC | 1.61°F | 1.42°F | 37% | 64¢ | 118 |
| ChicagoKMDW | 1.48°F | 1.31°F | 54% | 60¢ | 118 |
| DallasKDFW | 1.51°F | 1.24°F | 52% | 62¢ | 91 |
| AustinKAUS | 1.51°F | 1.69°F | 41% | 61¢ | 91 |
These describe the stations, not our model. Every figure below is the public NWS morning forecast measured against the official settled high, so you can reproduce them without taking our word for anything. Ranked by how much the weather itself moves.
| Market | Moves per day | Forecast busts ≥3°F | Obvious bracket right | Days |
|---|---|---|---|---|
| ChicagoKMDW | 4.43°F | 24% | 57% | 119 |
| New YorkKNYC | 4.03°F | 28% | 47% | 119 |
| San FranciscoKSFO | 3.82°F | 41% | 31% | 119 |
| AtlantaKATL | 2.72°F | 15% | 61% | 119 |
| Los AngelesKLAX | 2.29°F | 42% | 39% | 119 |
| AustinKAUS | 2.29°F | 24% | 62% | 91 |
| HoustonKHOU | 2.19°F | 17% | 60% | 119 |
| DallasKDFW | 2.08°F | 20% | 66% | 91 |
| MiamiKMIA | 1.64°F | 10% | 71% | 119 |
Read the two middle columns separately — they are not the same thing. “Moves per day” is how far the settled high travels between consecutive days: raw weather volatility. “Forecast busts” is how often the public morning forecast lands 3°F or more from the truth: how predictablethe place is. A station can be calm and hard to forecast at the same time — Los Angeles barely moves day to day, yet its forecast busts about as often as San Francisco's. That combination is the most useful thing on this table, and any single “difficulty score” would hide it.
We are not telling you to avoid the volatile ones. A quiet market is quiet for everybody: where the obvious bracket is right three days in four, it is priced that way, and there is little left on the table. Spread is where the payoff lives. What costs traders money is spread they did not expect — so the point of this table is to know which kind of day you are buying before you buy it, not to steer you to the calmest ticker.
Method: each settled day compares the model's logged prediction to the official NWS CLI high for that station. Morning rows use the ~8 AM call with a live forecast anchor; midday rows use the station-local noon read and exclude days the observation feed was down (the μ would be computed blind). Backfilled and untrainable rows are excluded — the same gate the model's own learning uses. “Market favorite” is the average price of the Kalshi book's own top bracket at that same read, from our logged snapshot of the book — the honest measure of how much of the answer was already public when we published ours. We deliberately do notpublish a pre-close (5 PM) number: that read lands after the day's peak, when the market has already repriced, so scoring ourselves on it would flatter the model on exactly the days an earlier call had already cost someone money. Catch prices are the ask on the bracket we named at the moment we named it; where we never recorded a price we show “—” rather than substituting another number. Updated hourly.