The first four pieces were mostly about speed: who sees the weather first, and how fast the price agrees. This one asks a slower question about the price itself. A bracket at 4¢ says it will pay about four times in a hundred. Does it?
The test
The desk reads every bracket of Kalshi's daily-high and daily-low ladders once an hour, in 20 cities. For each market it keeps one price at a fixed distance from the close: the midpoint of the bid and the ask at its last read 24 hours before the market closes, while the day being measured has not yet begun. Then it waits for Kalshi to settle the market and records whether the bracket paid. A ladder has six brackets and exactly one of them pays, so 599 ladders give 3,594 prices and 599 winners. Group the prices into bands, and a fair price is one whose band pays about as often as its average price says.
From 10¢ up, the price is right
The 1,697 brackets priced 10¢ or more a day out are as honest as the desk can measure. Those priced between 10¢ and 20¢, 14.7¢ on average, paid 14.4% of the time. Those between 20¢ and 50¢, 33.2¢ on average, paid 34.1%. Those at 50¢ or more, 60.0¢ on average, paid 60.1%. Together they returned 101 cents for every dollar of price, with a 95% interval of 100¢ to 103¢ when whole market days are resampled. Nothing in that half of the ladder is priced too high or too low by more than chance.
Below 10¢, it isn't
The other 1,897 brackets, a little over half of every ladder, were priced under 10¢ a day out. They paid 34 times. At an average price of 3.0¢, they should have paid about 57.
The cleanest test is the long shots priced 2¢ to 10¢, well clear of Kalshi's one-cent floor. There were 1,076 of them, at 4.5¢ on average, and 31 paid: 2.9%, with a 95% interval of 2.0% to 4.1%. That is 64 cents back for every dollar of price. Resampling whole market days, the interval runs from 44¢ to 83¢, nowhere near a fair 100¢.
From 10¢ up, a dollar of price paid back a dollar. Below 10¢, it paid back 64 cents.
It is not one week or one kind of market. Brackets at 2¢ to 10¢ paid 65 cents per dollar of price in the week of September 21 and 69 in the week of September 28. The overnight lows paid 57¢ per dollar of price, with an interval of 34¢ to 78¢. The daily highs paid 72¢, and their interval, 35¢ to 120¢, is still wide enough to reach a fair price.
Under 2¢ the gap looks wider still, 3 paid out of 821 against an average price of 1.0¢, about 36 cents per dollar of price. Read that one carefully. A contract cannot be offered below a cent, so a bracket nobody will bid on sits at a midpoint of half a cent however unlikely it is. Piece 02 found who collects that cent. Part of what looks like overpricing at the very bottom is the floor itself.
How much each band paid back per dollar of its price, a day before the close
the numbers
Where the extra three cents goes
Because exactly one bracket in a ladder pays, a ladder's six prices should add up to about a dollar. A day out, they added up to $1.026 on average. Split the ladder at 10¢ and the extra is easy to find. The long shots carried 9.5¢ of each ladder's price and paid 5.7¢. The rest carried 93.0¢ and paid 94.3¢. The whole surplus, and a little more, sits in the cheap brackets.
Long shots do come in
None of this means the cheap brackets never pay. 34 of the 599 winners were long shots a day before the close, about one in 18 (95%: 4.1% to 7.8%), and 16 of them were priced under 5¢. The winning bracket was usually not a clear favourite a day out either: its median price was 38.5¢, and fewer than a third of the winners were priced at 50¢ or more. Upsets are real. They are just rarer than the prices on them say.
What a long shot costs
The midpoint is not a price anyone pays. Taking a contract means paying the ask, plus Kalshi's taker fee: 7% of the price times one minus the price, rounded up to the next cent. That fee is at least a cent on any contract, so on a 2¢ ask it adds half the price again. On the books that had both a bid and an ask, so the ask is known, one contract of each long shot at 2¢ to 10¢ cost 7.1¢ on average and returned 44 cents per dollar (95%: 30¢ to 57¢), a loss of 4.0¢ a contract. From 10¢ up, the same spread and fee left 86 cents per dollar (84¢ to 88¢). Every band loses the spread and the fee. The long shots lose them on top of a price that was already too high.
What a dollar spent at the ask plus the fee brought back, by band
the numbers
Closer to the close
The desk keeps the same read 12, 6 and 1 hour before the close. A day out, the long shots paid 59¢ per dollar of price. Twelve hours out it was 62¢. After that the day has mostly happened, the cheap end of the ladder is mostly brackets the weather has already ruled out, and the price runs into the floor. Six hours out, long shots paid 7 times out of 3,091. One hour out, they paid twice out of 3,121.
| Read | brackets | paid | price | per $1 | 95% |
|---|---|---|---|---|---|
| 24 h out | 1,897 | 34 | 3.0¢ | 59¢ | 42–76¢ |
| 12 h out | 2,619 | 25 | 1.5¢ | 62¢ | 37–87¢ |
| 6 h out | 3,091 | 7 | 0.9¢ | 26¢ | 12–40¢ |
| 1 h out | 3,121 | 2 | 0.8¢ | 8¢ | 0–21¢ |
long shots: brackets with a midpoint under 10¢ at the read, settled · price is their average midpoint · per $1 is cents paid per dollar of price · the 24-hour read starts with the markets of 2026-09-19, the others with 2026-09-18 · intervals resample whole market days
An old pattern, in a new place
Racetrack bettors have shown the same thing since at least 1949, when R. M. Griffith found that horses at long odds won less often than their odds implied. It is called the favourite-longshot bias, and it has turned up in betting markets ever since. On Kalshi's temperature ladders it shows up only below 10¢. Above that, the desk cannot find it.
What it means for the desk
Three things, none of them a recommendation.
- The price is honest where the ladder is uncertain. From 10¢ up, a day before the close, a bracket paid about as often as its price said, in every band the desk could test.
- The cheap end is where the price and the record disagree, by about a third of the price. The fee then takes at least a cent on every contract, which at those prices is a large share of what the contract costs.
- The record is short. Fifteen market days is enough to separate 64¢ from 100¢, but not enough to say how steady the gap is. So the desk has posted it as call C-012: on the markets of November 1 to 30, brackets priced 2¢ to 10¢ a day out will pay back less than 85¢ per dollar of price. The desk gives that eight in ten, and it is graded on December 2. Meanwhile the calibration page re-grades every settled price as it lands.
Method
Prices come from Kalshi's public market-data API, read by the desk's robot once an hour for every bracket of the daily-high and daily-low temperature ladders in 20 cities. Settlement is Kalshi's own result for each market. For each market the desk keeps its last read at or before 24, 12, 6 and 1 hours before the market's close (Kalshi's close time, the end of the measured day), taken within the ninety minutes before that moment, or none. The price is the midpoint of the best bid and the best ask, counting an empty bid as zero. The desk keeps the spread only when both sides are present, so the cost figures use only those books (2,982 of the 3,594 brackets a day out). The cost is the ask, the midpoint plus half the spread, plus Kalshi's taker fee on one contract: 7% of the price times one minus the price, rounded up to the next cent.
A long shot is a bracket whose midpoint was under 10¢ at the read. "Paid per dollar of price" is the number of brackets that paid, times 100¢, divided by the sum of their prices. Rates carry a Wilson 95% interval. The per-dollar figures carry a 95% interval from resampling whole market days, 1,000 draws, because brackets that settle on the same day share its weather and the same ladder's brackets are not independent. The weeks are ISO weeks of the market date, and the first is a partial week from September 19.
The window is every market day settled when the figures were run, 2026-09-19 to 2026-10-03 for the 24-hour read. Every market with a read in that window and a settlement is in. Every figure here comes from python agents/longshots.py --from 2026-09-19 --to 2026-10-03 (add --horizon 12, 6 or 1 for the other reads) on the desk's tables, and the calibration page grades the same reads live.