The longshot gap, and the spread that swallows it

This study was pre-registered. Its question, its dataset boundary, its eligibility rules, its price buckets, its hypothesis and the exact test a bucket has to pass were written down and committed before any price was compared to any outcome, together with what would be published if the result came out null or pointed the other way. Three dated amendments record every change since, each one written before the changed collection ran.

That matters here more than it would on a descriptive piece. This page publishes a hypothesis and a finding in the same document at the same moment, and from the outside that is indistinguishable from running the numbers several ways and keeping the framing that worked. The registration is what makes the difference checkable rather than asserted: the plan has an earlier commit date than the analysis, and the amendments say what was changed, when, and why.

The amendments are not hidden in a footnote because two of them cost the study precision. The collection window was narrowed to 90 days, because pricing an older market against the exchange's current fee schedule would compute a cost nobody ever paid. The eligible set was reduced to a seeded uniform sample, because the design owes one history request per market and the eligible set ran past 300,000, which is more load than this site should put on somebody else's free public endpoint. The quote budget was then cut further, and that third amendment carries an obligation with it: every bucket on this page publishes the smallest mispricing it could have detected, because a smaller sample widens every interval, and wider intervals make this study's own hypothesis easier to confirm.

Every settled event on Kalshi was walked by cursor through the documented public REST API at api.elections.kalshi.com, with nested markets, from 2026-10-01T03:07:33.383Z. No account, no authentication, and no page of the consumer website: kalshi.com returns a bot challenge to automated access including on its own robots.txt, which was not worked around.

Enumeration goes through events rather than the flat markets listing, and that is a correctness choice rather than a convenience. A probe of the flat settled-markets endpoint returned 200 of 200 rows as auto-generated multivariate combination shards. A study built on it would have measured the exchange's parlay generator instead of the exchange.

The walk saw 5,011,948 settled binary markets closing inside the 90 day window that opens 2026-07-03. 304,714 of them met every eligibility rule except the quote test: a real binary market rather than a combination shard, a settled yes-or-no result, at least 48 hours between opening and closing, and some traded volume. A seeded uniform draw of 60,000 was taken from those, and reference quotes were fetched for a 15,000-market prefix of it.

12,755 markets across 9,970 events carried a genuine two-sided quote at the reference horizon and form this study's universe. Every market in the quote budget was answered by the exchange, so nothing in this universe is missing because a request was refused.

A settled market's last traded price has already converged toward its own outcome, so it cannot answer a question about what the market thought before it knew. Using it would measure settlement against itself.

Instead, for every eligible market, hourly candles were read over the window from 48 hours before close to 24 hours before close, and the last candle in that window carrying a two-sided quote supplies the price. In one sentence: the last quoted price at least a day before the market closed. The median market on this page was priced 24.5 hours before it closed.

Twenty-four hours is the horizon because it sits before the late convergence that makes a settled price uninformative. The 24 hour lookback behind it exists so that a thinly traded market is not dropped for having no quote in one particular hour. A quote counts only where the bid is above zero, the offer is below a dollar and the offer is above the bid, which is the identical test the sibling cost study applies, so a market excluded there for having no two-way price is excluded here for the same reason.

Per market, from the reference candle only: the price, the midpoint of the closing bid and offer in cents, read as the market's own estimate of the probability; the half-spread, half the distance between bid and offer, which is what a taker gives up crossing the book to enter; and the outcome, one if the market settled YES and zero if NO.

Per price bucket: the share of markets settling YES, expressed in cents per contract so it reads against the price on one scale; a 95% Wilson score interval on that share; the gap, settled share less mean price; the cost, mean half-spread plus the exchange's published taker fee at the bucket's mean price; and the net edge, the size of the gap less the cost. Settlement is free on this exchange, so a position held to settlement pays no exit leg.

The fee is deliberately understated. Kalshi's filed schedule rounds its fee up to the next cent per order, so the size of that term depends on order count and roughly doubles the fee near 50 cents on a one-contract order. It is left out here, exactly as the sibling cost study leaves it out, and the direction is the point: leaving it out makes costs look lower, which makes a gap look more likely to survive them, which is the opposite of what this study expected to report. An unmodeled term should err against the study's own finding.

Pre-registered, and it is the one methodological choice that decides what this piece says.

A bucket's mispricing survives its cost only if the Wilson bound nearer the price is still further from the price than the cost is. For an overpriced bucket the upper bound has to sit below the price less the cost; for an underpriced bucket the lower bound has to sit above the price plus the cost.

An interval that merely excludes the price is not enough, and the distinction is the whole study. Clearing zero says the gap is probably real. Clearing the cost says the gap is bigger than what it takes to reach it, and only the second is a statement a reader can do anything with. A page that reported the first and called it an edge would be making a claim its own numbers do not support.

In a ten-candidate election event exactly one market settles YES by construction. Those markets are not independent observations, and a Wilson interval computed over all of them is narrower than the data supports, which would make a gap look better established than it is.

So every figure here is computed twice: once over every priced market, and once over a cut taking the single highest-volume market in each event, ties broken by ticker so the cut is deterministic. Per the pre-registration, a claim reaches the findings, the meta description or the outreach brief only if it holds in both. Where the two disagree the one-market-per-event figure is the headline and the disagreement is stated rather than resolved quietly.

The cut is a real cost, not a formality: it reduces 12,755 markets to 9,970, widening every interval. That it is the headline cut rather than the fuller one is the conservative choice, and it is the choice the registration fixed before anyone knew which way the numbers would go.

The model is pure arithmetic over the committed capture, with no network call, no clock and no randomness beyond the fixed sample seed, so it reproduces exactly. Its own suite holds the bucket edges, the Wilson bounds, the gap, the fee formula against the three points the exchange's filed schedule quotes, the decision rule, and specifically a case that clears zero and fails the cost, which is the one distinction the whole piece turns on.

The page's figures are generated from the run's output rather than typed, and a second suite re-reads that output, asserts the generated copy against it cell for cell, and then re-derives every bucket, every interval and the decision rule from the capture itself with arithmetic written independently of the study's model. A figure that survives regeneration while meaning something different still fails.

The power figure beside each bucket shares the decision rule rather than restating it: it walks every settlement count the bucket could have produced, applies the identical test, and reports the distance from the price of the nearest count that would have passed. A power figure computed from a looser rule than the one that gates the finding would be a second claim nothing checks.

The capture is committed, so every figure on this page recomputes offline from the published data with no dependence on where the exchange has moved since. The dataset under this page carries each bucket on both cuts.

Re-running the collection live will not reproduce these numbers and is not meant to. Settled markets age out of a 90 day window and candlestick history belongs to the exchange rather than to us. That is exactly why the capture is committed rather than re-fetched, and why the figures are pinned to a stated instant rather than to whatever the exchange looked like when a reader checked.