Not from the settlement page. This reads the raw routine reports straight from the station through api.weather.gov and rebuilds what the settlement source should show — take the routine report's tenths of a degree Celsius, convert, round half up, post it to the following hour. It is a reconstruction, verified 58 of 58 hours on two stations, not a copy. Forecast and backfill cells come from model blends and are not observations at all.
api.weather.gov keeps roughly seven days of observations and nothing older. Beyond that the grid has no station data and falls back to a model, shown as flat grey italic cells labelled backfill with a badge on the day. Those can miss the real reading by several degrees and should never be read as data — the settlement page keeps its own archive going back further, so it will disagree, and it is right.
The mean of the 24 hourly values for the day. Not the midpoint of the high and low. Rounded to a whole degree at the end.
A day counts only if the rounded average lands above the line, so it has to reach 91. Exactly 90 fails. That puts the true bar on the unrounded mean at 90.5.
Each station files one routine report near the end of every hour, and that report is the following hour's value. A 10:52 report is the 11a cell. Verified against the settlement source on two stations, 68 hours, exact.
The routine report carries tenths of a degree Celsius. That is what gets converted and rounded, not the whole-degree figure in the report body. The two disagree by a full degree on roughly a third of hours.
A 24-hour day clears the 90.5 line only if its hourly readings sum to 2172. Deg to go is that total minus what is already banked; the small figure under it is what every remaining hour has to average to get there. Green means the model blend for those hours is at or above that number, red means it is not.
Above means >, and values are read to two decimal places, so Yes needs 0.01" — one measurable tip. Trace counts as 0.00. Missing daily values count as 0.00. Both are a No. Total precipitation is liquid equivalent and includes melted snow, sleet and hail, so a winter market is not a rain market.
The rulebook defines the area by the NWS primary station and sums the daily precipitation values reported by the Source Agency — and an agency's daily value is its own local climate day. So the rain tab is pinned to station local and ignores the board's Eastern toggle. Window names the day and clock being priced. Last trading is one minute before that day ends.
Only the initial non-preliminary value resolves the market; a later correction does not supersede it unless the first release was itself flagged preliminary. That makes a dark station a directional risk rather than noise — if the daily value lands missing, it is a zero and no amount of backfilled data undoes it. Report risk in the model read is that discount, up to 25% for a station reporting almost nothing.
The point model asks does the bucket tip, not does the point get wet. Each member's cell total gives a wet-area fraction w = 1−e−S/S₀, and P(point ≥ 0.01") = w·e−0.254·w/S. A drizzle event with a 0.1 mm cell mean prices near 11%, not near certain — that is the trace case, and trace pays No.
Two independent reads, blended. The ensembles — ECMWF, GEFS, ICON, GEM, roughly 130 members — each carry a full run, so summing one member across the remaining hours preserves the timing correlation a probability figure throws away. The NBM read compounds the calibrated hourly PoP over Bursts effectively independent chances, spread across however many hours remain — so a 35% afternoon reads the same at 1pm as at 4pm instead of inflating with the number of boxes to multiply. NBM wt sets the mix.
A model cell is 10–25 km across; the bucket is one spot inside it. Each member's cell total is converted to a point chance through 1−e−S/λ, where λ widens with CAPE — layered rain wets the whole cell, a summer air-mass storm wets part of it. λ is shown per city, from 0.4 to 2.2.
The weights, λ and the correlation term are reasoned defaults, not fitted to these 28 stations. Log the fair value and the settled outcome daily and the reliability curve will tell you where it is biased. Until then treat it as a considered opinion, not a price.
Priced as its own probability, then combined: fair = already + (1 − already) × rest of day. The ladder, strongest first — the station's own hourly gauge, which is the only rung allowed to reach certainty; its 6-hourly precip group; hours of rain in present weather; and last, model analysis, capped at 0.80 and cut to a quarter of that when the station has been reporting all day and never once observed rain. A model saying rain fell over a station that never saw any is a cell that missed the field, not a settled market.
Expiration value is read at 10:00 AM ET on the expiration date, but once the payout event occurs expiration moves earlier — a settled Yes can pay before the window closes. Position accountability is $25,000 per strike per member.
Edge is fair minus your price in cents; net subtracts the trading fee, which peaks near 1.75¢ at a 50¢ price and falls toward the wings. A two-cent edge on a coin flip is not an edge.
A day with fewer than 18 reported hours does not count and breaks a streak. Hour counts turn red when a finished day is short.
The settlement page prints every station in Eastern regardless of where it sits. The rules say local time. Outside Eastern those are different 24-hour windows. Use Day on to see both, and confirm with Kalshi which one settles. Each row carries the station's own clock next to its call sign — Phoenix reads MST year round, so it sits three hours off Eastern now and two after the clocks change in November.
A station usually files its routine at the same minute, but a delayed or corrected report can land at :54 when it normally files :52. Those hours used to vanish and get backfilled with a model blend — a violet cell where a real observation existed. Now they are recovered and marked with a white underline, and the hour count shows them as +n r. They count toward the eighteen.
A blank on the settlement page means no report arrived before resolution. One that shows up later still lands in the archive this reads, so a filled hour here can be a permanent blank there.
Solid figures are reported observations. Anything violet is a weighted model blend standing in for hours that have not happened. It moves with every model run and it is not evidence.