Prediction Market Mispricing Scanner — Polymarket

Finds contradictions in Polymarket pricing. Not forecasts — contradictions.

Nobody knows whether a price will rise. But if a higher strike is priced above a lower one, or if buying every outcome of an event costs less than the $1 it pays, that is wrong under any view. Those signals need no argument.

What it returns

Implied probability distributions. Polymarket runs strike ladders — "Bitcoin above $62,000 on August 14?" at many strikes for one expiry. Read together they form a cumulative distribution, which yields an implied median, an 80% range, and a term structure that widens with time. This is the information an options surface carries, extracted from a prediction market.

0.89 days left   median $63,263   80% range $62,151 – $65,182
1.89 days left   median $63,411   80% range $61,896 – $65,494
2.89 days left   median $63,420   80% range $61,316 – $65,700

Monotonicity violations. In that same ladder the curve must fall as strikes rise. Where it does not, the two strikes contradict each other:

$70,000 → 0.0015   vs   $72,000 → 0.0020   (excess 0.0005)

Complete-set costs. Binary markets where YES + NO costs under $1, and multi-outcome (negRisk) events where every mutually exclusive outcome together costs under $1 — both priced at live asks, net of fees, with the size actually executable at that level.

Three mistakes it avoids

Midpoints cannot show a gap. outcomePrices from the Gamma API are midpoints and sum to 1 by construction. Scanners built on them report edges that vanish the moment you try to buy. Every price here comes from the CLOB order book.

Fees are not a constant. Polymarket's taker fee is shares × rate × p × (1-p) — a parabola, largest at even odds and vanishing in the tails. Subtracting a flat buffer is simultaneously too small near $0.50 and too large at the extremes. Markets whose fee rate cannot be determined are excluded rather than assumed free.

Partial outcome sets are not arbitrage. Buying 13 of an event's 45 outcomes for $0.28 looks like a 72% edge, but the winner may be among the other 32. Worse, the fewer outcomes counted the better it scores, so a naive ranking fills its top slots with the most broken candidates. Events missing any outcome are disqualified and reported with the reason.

negRisk sets are ranked by annualised return, not absolute profit. Their legs belong to different markets, form no market's complete set, and cannot be merged — capital is locked until resolution, so a two-year lock-up with a real gap is worth less than a two-week one.

Input

Field	Default	Meaning
assets	["bitcoin","ethereum"]	Assets whose strike ladders to read
includeDistributions	true	Implied distributions and violations
includeBinary	true	YES + NO complete sets
includeNegRisk	true	Multi-outcome complete sets
onlyProfitable	false	Keep only rows with an edge or a violation
limit	20	Rows per section
minLiquidity	1000	Skip thin books
maxOutcomes	60	Skip events too wide to quote fully

Output

One dataset, one row per finding, each tagged with type: implied_distribution, binary_complete_set, negrisk_complete_set.

Honest limits

• Prices move. A gap visible when the actor runs may be gone seconds later.
• Legs are quoted at the best ask only; filling more than executable_shares walks the book and costs more.
• Gas and any platform fees beyond the taker fee are not included.
• Most of the time nothing is profitable, and the run says so. A scanner that always finds opportunities is measuring midpoints.

Data comes from Polymarket's public Gamma and CLOB APIs. No key required.