Polymarket Arbitrage

Polymarket Up/Down is a bounded-price instrument on a 0-to-1 scale; the price of Shares is the probability-weighted value of a $1 payout. Exchange futures quote a dollar price. The two meet only through the conversion of the underlying level into a conditional probability, which is why the gap between the Polymarket mid and the futures price is an energy source for traders.

On 2026-09-09 the ETH 5-minute market repriced in one direction: the Binance futures price fell from 2,452.84 to 2,393.34 between 14:41:45 and 14:42:50 UTC — a $59 move over 65 seconds. The Polymarket mid flipped from 0.815 to 0.065 in the same interval, and both references re-priced to the new level at 14:42:50. That same-second cross-venue alignment is the pattern to measure with.

Not every gap is tradable. The ladder can move a 99.6-cent width in one second (documented from 0.044 to 0.996), and a stale local quote can hold for seconds or minutes while the venue-priced level has already moved. A "gap" that a snapshot captures must be separated from an executable two-sided book by replaying the order flow behind it.

• Align by UTC second: join Polymarket frames to exchange ticks on floor(utc), as the same-second studies do; sub-second drift between clocks destroys the signal before any arithmetic does.
• Use the executable boundaries, not the mid: the mid sits between ladder bounds that are often one-sided, so the relevant spread is bid-to-ask-to-conversion, not mid-to-mid.
• Require a filled flow: an arb hypothesis needs the maker/taker sides you would actually hit — the archive's one-sided frames (90%+ of the September sample) tell you when a leg has no resting book.
• Window the study over the whole ladder and the final minute, where one-sidedness concentrates and where the only price discovery is the ladder itself.

Cross-venue mean reversion is the strongest pattern: after a same-second reset, robust markets revive within a width consistent with one or two ticks (0.99/0.995 dominate, 98.6% of sampled frames), and the executable floor tracks the futures level. Replay a few hundred same-second resets and the reversion is visible in seconds, not minutes.

Intra-venue ladder arbitrage — buying the mid-adjacent level and selling the far level when the ladder is one-sided — exists only where two sides are actually resting; with ≥90% one-sided books over the sample, the honest entry is to test your definition of the book before capital sits in a crossed or empty frame.

The settlement boundary is the third kind: in the final minute, ask presence in only 3 of 18 sampled frames means a takeable exit requires hitting a one-sided ladder, which is a different risk profile than a market with two resting sides.

• Pull any ETH 5-minute market for 2026-09-09 and Binance futures for 14:41–14:44 UTC.
• Join on the second; record mid-to-mid and executable-floor-to-futures gaps per second.
• Compute the same-second re-price count: how many seconds the gap widened after the 14:42:50 reset versus before.
• Publish the count with the window and the pair named, and treat anything not replayable as a fixpoint, not a signal.

The arbitrage on Polymarket is narrower and faster than the dashboard versions of it, and that is exactly why it survives: the durable edge is a same-second reprice caught on the executable floor, measured against a reference that moved, on a book that reset in the same second.

Work the loop in the order that removes hope first: replicate the 2026-09-09 ETH window at 250ms, count the same-second resets, then replay your entry logic against those frames. Every step that fails to reproduce from raw rows is a step you were about to fund with a thesis.

When the reproduce passes, the honest conclusion is not "arb it every day" but "the gap is real in this window, in this pair, on this side." Operationalize that window. The archive exists to tell you which window it is.

• Implied probability
• Liquidity
• Order flow
• Settlement data