Prediction Market Scanner: Polymarket + Manifold Spreads

Prediction Market Scanner — Polymarket + Manifold, with cross-platform spreads

Two of the largest prediction markets, in one schema, with the price disagreement between them attached to every market that appears on both.

Polymarket and Manifold both publish free APIs. Neither tells you that they are pricing the same event differently. That number — the spread — is what this Actor exists to produce, and it is the one field you cannot get from either venue directly.

What you get

One row per market. Every row carries the venue's own data plus, when the same event is priced on the other platform, the match and the spread:

Field	Meaning
title, platform, url	the market and where it lives
probability_percent	the Yes price, as a percentage
volume_usd, volume_24h_usd	total and 24h volume (24h is Manifold-only)
liquidity_tier	A >$100k · B $1k–$100k · C <$1k
opportunity_score	0–100: volume-weighted closeness to a coinflip
end_date	ISO-8601 resolution date, normalised across venues
matched_platform, matched_title, matched_url	the same event on the other venue
matched_probability_percent	what the other venue thinks
spread_percent, spread_signed_percent	the disagreement, in percentage points
match_confidence	0–1 strength of the title match. Read this before trusting a spread.

Sorted by opportunity_score descending. Unmatched rows carry the match fields as null rather than omitting them, so the dataset has one stable shape.

Why probability_percent is the Yes price specifically

Polymarket returns outcome prices as an unordered pair. Reading the larger of the two gives you whichever side is more likely — which silently inverts the meaning of about half the rows. On "Xi Jinping out before 2027?" the real Yes price was 2.8%; the naive read reports 97.2%, the exact opposite. This Actor resolves the price at the index of the literal Yes outcome and skips any market without an unambiguous Yes/No pair rather than guessing. A row that looks valid and means the opposite is worse than no row, and it makes every cross-venue comparison meaningless, because Manifold's probability is always the Yes price.

Why the mid-tail (B) is the interesting band

A markets are deep and already efficiently priced. C markets are too thin to act on. B — $1k to $100k — is liquid enough to trade and thin enough that the crowd has not finished pricing it. If you only want one tier, that is the one.

Input

Everything is optional; the defaults scan both venues and return everything.

{
  "platforms": ["polymarket", "manifold"],
  "maxMarketsPerPlatform": 200,
  "liquidityTiers": ["B"],
  "crossPlatformMatchesOnly": true,
  "minSpreadPercent": 5,
  "maxResults": 100
}

That example asks the practical question: which mid-tail events do the two venues disagree about by 5 points or more?

maxResults caps your bill — you are charged per row returned, so set it if you want a predictable cost per run.

Pricing

Pay-per-event. You pay for output, not for runtime:

Event	Charged when
market-scanned	one per market row returned to you
cross-platform-match	one per market that carried a spread

Filters are applied before charging, so a run that matches nothing costs nothing beyond the platform's own start event. Each row is charged as it is written, so a run you stop early is billed only for the rows you received. Set Max results to cap a run's cost.

Limits, stated plainly

• Binary markets only. Multi-outcome markets are skipped; they have no single Yes price and so no comparable number.

• Matching is lexical, weighted by how rare each word is across the markets in that run. The rare words — names, places — carry the match; the genre's boilerplate ("will", "2028", "presidential", "win") carries almost none, and a pair must share at least one genuinely rare word to match at all.

  That test matters more than it sounds. Plain token overlap, measured on a live 400-market scan, paired "Will MrBeast win the 2028 Democratic presidential nomination?" with "Will JD Vance win the 2028 US Presidential Election?" and reported a 19.9-point "spread" between them — 104 matches, nearly every one of that shape. The rare-word requirement rejects it, and the same scan then produced 5 matches of which 4 were exact.

• The fifth was a near-miss, and it is the failure mode to know about. "Will JD Vance be nominated for President" matched "Will JD Vance win the Presidential Election" at match_confidence 0.61 — same person, different stage of the same race, so the 22-point "spread" means nothing. Lexical matching cannot see that distinction. A high match_confidence means the titles agree, not that the events are identical — read matched_title before acting on any spread. Word inflection is not normalised either ("cut" ≠ "cuts"), so some true pairs are missed.

• Manifold's volume24Hours has no Polymarket equivalent on the Gamma markets endpoint, so volume_24h_usd is 0.0 for Polymarket rows rather than guessed.

• One venue failing does not fail the run. You get the other venue's rows and a warning in the log; a short result is always explained rather than left to inference.

• Not financial advice. A spread is a disagreement between two crowds, not an arbitrage: the two venues settle on different rules, at different times, with different fees and different resolution sources.

Data sources

Public, documented, unauthenticated endpoints: Polymarket Gamma API · Manifold Markets API v0. No credentials needed, no login, no scraping of a logged-in surface.