Why a hit rate on its own says nothing
75.9% paid
Kalshi's same-day temperature favourites, the brackets priced between 65c and 90c, paid 1,093 of 1,440 times since 18 September. On its own, that looks like a strong record. But the price had already said about as much: the average ask on them was 78.4c, a 78.4% chance as the market priced it. They paid 2.5 points less often than that.
A hit rate means something only beside the price paid for it. A record of favourites that win three times in four is what a market full of three-in-four favourites produces by itself.
As read from the scorecard's views on 2026-10-23 · 95% interval on the share paid: 73.6% to 78.0% · the weeks, one by one
-
Was every prediction dated somewhere independent, before the event?
A record kept only by the seller can be edited after the fact. A date someone else holds cannot.
Ours, partly, and we say so. Each of our calls carries its posting date and a number in the order posted, with none skipped, and a changed call fails the site's own build check. But no outside service has timestamped those calls yet, so their dates rest on our word. Each call posted from 16 October on also links a receipt of the figures it rested on, run before it went up. The calls, numbered
-
Was it scored against a price that was really on offer, fees included?
A prediction scored against a price nobody was offered, or with the fee left out, reads better than it was.
Ours: at the ask that was there. The favourites are counted against the ask resting when the robot read them, not the midpoint; the studies charge Kalshi's taker fee from its own schedule. Where a page uses the midpoint, as calibration does, it says the midpoint is nobody's trading price. How the favourites are read · the fee
-
Are the misses kept, all of them?
A record that drops its misses, or keeps only the predictions it charged for, is a sample chosen after the results.
Ours: every miss stays, first. On the calls page the misses come first (every one so far), as written, each with its receipt; a call is never edited after it is posted, and an ambiguous rule is graded against us. Nothing here is sold, so there is no free tier to leave out. The desk this site began as lost to the market's price, and that loss is published too. Misses first · the desk's record · the corrections log
-
How many predictions, and what is the margin of error?
Ten predictions prove almost nothing either way. A rate without its count and its interval cannot be read.
Ours: the count beside every rate, and its 95% interval where the page draws one. Calibration, the games and the box at the top carry their intervals; a scorecard week with under 20 settled is shown and not read. Our own calls are few: too few to tell whether the chances we state run high or low, and the calls page says so. Calibration, with its intervals · our record
-
Is the hit rate set against the price that was paid?
This is the question the first four lead to, and the one a hit rate alone cannot answer: the price already carries the market's own chance.
Ours: that is the whole measurement. Every rate on the scorecard and the calibration page sits beside its price: how often brackets at a price paid, against that price. The box at the top is one of them. The scorecard · calibration
This guide names no site and grades no one's record but ours. Mispricd sells nothing: no paid subscriptions, no ads, no affiliate or referral links. Its calls are research claims about how markets behaved, never a recommendation to trade.