Three analyses · one archive · 612,661 settled markets
The final hour is where the market breaks.
We came at Kalshi’s prices three separate ways — calibration over time, the spread on two-sided markets, and violations of basic no-arbitrage bounds. All three point at the same thing: markets that trade continuously degrade in the final hour before close, while markets that resolve on a schedule don’t.
Finding 1 — calibration goes U-shaped in time
Finding 2 — the measured spread is ~101¢
Finding 3 — 28.9% of linked prices break no-arbitrage bounds
FINDING 01 · CALIBRATION DECAY
Sports sharpens, then snaps. Economics just sharpens.
Measured how far the market’s price sat from reality at 7 days, 24 hours, and 1 hour before close. Sports markets get sharper as the game nears — then the final hour breaks calibration, error jumping 6×. Economics markets, which resolve on a scheduled release with no in-play trading, improve all the way down. The shape of each line is the finding.
Distance from reality (lower = sharper) · by time before close
SportsEcon
Event-clustered bootstrap (2,000 resamples), n = 8k–352k snapshots per point. The final-hour worsening splits into a buying premium (trade prices worse than midpoints) and a genuine belief shift. Weather is excluded: an hour out, 97% of weather contracts are near-certain, so its aggregate error reflects near-determined outcomes, not sharp pricing.
FINDING 02 · THE MEASURED SPREAD
Kalshi’s real cut is a quarterof a sportsbook’s.
On two-outcome markets — where both sides must sum to a guaranteed 100¢ — quoted midpoints sum to almost exactly 100¢, which validates the whole method. Last-trade prices sum to ~101¢: that 1¢ gap is the real cost of crossing the book. A standard sportsbook line sums to about 105¢. The catch: that premium concentrates in the final hour, where over-100 pricing jumps to 60%.
~101¢
Last-trade sum on two-sided markets — the true round-trip cost
~100¢
Midpoint sum — validates the method to within a cent
60.5%
Share of pairs over 100¢ in the final hour, vs 47% a day out
43,704 structurally two-outcome events, archive-wide. Midpoints summing to ~100¢ is the internal control: an efficient two-sided quote must sum to the guaranteed payout, and it does.
FINDING 03 · NO-ARBITRAGE VIOLATIONS
Nearly one in three linked prices broke a basic bound.
In a strike ladder, a higher threshold must be less likely than a lower one — its price cannot be higher. Across 198,314 adjacent-rung comparisons, 28.9% violated that bound, and it wasn’t noise: 15,675 inversions exceeded 51¢. The violation rate nearly triples into the final hour, and it lives almost entirely in sports — the same continuously-traded signature as the other two findings.
Ladder monotonicity violation rate · by time before close
7 days out
11.5%
24 hours out
16.5%
1 hour out
31.8%
28.9%
Overall ladder violation rate, archive-wide
12.4%
Of two-sided pairs priced under 100¢ — the free-money direction, mean margin 12.9¢
6.1%
Economics violation rate — the calm, scheduled-release control
612,661 settled markets, computed server-side over the full archive. These are bound violations, a statement about pricing efficiency — not takeable profit. Executing requires simultaneous fills on every leg, and fees erase the smallest edges. What they measure is how often, and by how much, linked prices stopped being internally consistent.
WHY IT MATTERS
Three methods. One story.
Calibration decay, the measured spread, and no-arbitrage violations are independent analyses that share no code and no assumptions. All three land on the same divide: continuously-traded markets break down in the final hour; scheduled-release markets stay clean. When three separate methods agree, it stops being a quirk of one measurement and starts being a property of the market. That’s the finding — and it’s a map of exactly where, and when, prices stop being trustworthy.