Polymarket Real-Arb Scanner — executable, fee-adjusted, depth-checked

Finds guaranteed arbitrage on Polymarket prediction markets — and only the arbitrage you can actually execute. Every opportunity is priced off the real CLOB order book walked to your stake size, fee-adjusted for the exact market category, with the depth available at each leg reported. No phantom edges off last-trade or mid prices — the thing that makes every generic "arbitrage scanner" useless the moment you try to fill.

• Polymarket arbitrage scanner with real order-book depth, not top-of-book teasers
• Prediction market odds API output: clean JSON per opportunity
• Handles Polymarket's negRisk multi-outcome markets (elections, futures, "how many" markets) correctly
• Per-category taker-fee model (crypto 7%, sports 5%, politics/finance 4%, geopolitics free…)
• Pay-per-event pricing — you pay for the run plus each real opportunity found, nothing for phantom noise

---

Why most "arb scanners" lie, and this one doesn't

A prediction-market arb is only real if you can fill every leg, in size, right now, after fees. Almost every scanner computes edges from last-trade or Gamma mid prices — numbers you can never trade at. They surface dozens of "arbs" that vanish on contact.

This actor does the opposite:

1. Reconstructs the complete outcome set. A "buy every outcome" arb is only guaranteed over the full mutually-exclusive set. We pull every leg of each negRisk group from the event API — never a volume-filtered subset — and disqualify any set with a residual "Other / None" outcome, even an inactive one. (Summing three of thirty World Series teams and calling it an arb is the #1 way these tools lie.)
2. Walks the real CLOB book to your stake. The reported price is the VWAP you'd actually pay to fill stakeUsd, across as many levels as it takes — not the untradeable best-ask tick.
3. Deducts the real fee. fee = shares × rate × p × (1−p), with rate set per Polymarket category. Geopolitics is free; crypto is 7%.
4. Filters bad data. Crossed books, sub-cent ask walls, and dead books are treated as bad data, not free money.

The honest consequence: on efficient, high-volume markets this actor usually returns few or zero arbs — because there genuinely aren't any. That is the point. Fewer, real opportunities beat a screen full of fiction.

What it finds

Type	Structure	Guarantee
negrisk-buy-all-no	Buy NO on every leg of a mutually-exclusive set	N−1 legs always pay $1 — guaranteed by mutual exclusivity (safe even if the set is incomplete)
negrisk-buy-all-yes	Buy YES on every leg of a complete, no-Other set	Exactly one leg pays $1 — guaranteed only over an exhaustive set
merge-yes-no	Buy YES + NO of one market for < $1	They merge to $1 — rare and transient (market makers enforce ≥ $1)

Input

{
  "stakeUsd": 100,            // capital per opportunity; prices are VWAPs to fill THIS much
  "minNetEdgePct": 1,         // only return arbs ≥ this % profit AFTER fees
  "categories": [],           // e.g. ["sports","politics"]; empty = all
  "minVolume": 1000,          // skip markets below this lifetime volume
  "maxResults": 200,
  "includeWeatherEvents": false, // daily-temperature markets, off by default
  "maxMarketsScanned": 2000,  // cost guard
  "maxBooksFetched": 400      // cost guard (order-book requests)
}

Output (one dataset item per opportunity)

{
  "type": "negrisk-buy-all-no",
  "guarantee": "mutual-exclusivity",
  "eventTitle": "EPL: 2027 Champion",
  "legCount": 20,
  "sizeShares": 312.5,
  "capitalUsd": 18.42,
  "feesUsd": 0.31,
  "payoutUsd": 19.00,
  "netProfitUsd": 0.27,
  "grossEdgePct": 3.15,
  "netEdgePct": 1.47,          // profit after fees, as % of capital
  "staleness_ms": 640,         // age of the oldest order book used
  "legs": [
    {
      "market": "Will Arsenal win the 2026-27 EPL?",
      "side": "BUY_NO",
      "executablePrice": 0.61, // VWAP to fill your stake, not top-of-book
      "shares": 312.5,
      "depthWorstPrice": 0.63, // deepest level you'd consume
      "levelsConsumed": 4,
      "feeUsd": 0.02
    }
    // …one leg per outcome
  ],
  "scannedAt": "2026-07-22T23:40:00.000Z"
}

Pricing (pay-per-event)

Event	Price
Actor start	$0.02 — covers discovery + book screening, even on zero-arb runs
Arb found (per result)	$0.02 per executable opportunity returned

You are never charged for phantom opportunities, because none are emitted — the Actor only ever writes a dataset item when it finds a real, executable, fee-cleared arb.

Limitations (read these)

• Executable arbs are rare on liquid markets. Efficient books are arbitraged away by bots in seconds. Expect most runs on top-volume markets to return 0. Value comes from scheduling frequent runs to catch transient windows, and from scanning mid-liquidity markets.
• Snapshot, not a guarantee of fill. staleness_ms tells you how fresh each book was; by the time you act, the book may have moved. Treat every result as a lead to verify, not a filled trade.
• Depth is top-of-run. Books are fetched once per run. For sub-second opportunities you need your own live connection; this is a scanner, not an execution bot.
• Single-venue. Cross-venue (Polymarket ↔ Kalshi) arbitrage is a separate actor.
• Weather/temperature markets are excluded by default — thin books, specialized resolution; enable with includeWeatherEvents if you know what you're doing.

How it works (pipeline)

Gamma market discovery (prices) → negRisk event reconstruction (complete sets) → cheap price screen → CLOB order-book walk (executable VWAP + depth) → per-category fee deduction → guarantee & sanity checks → dataset

Built by a developer who runs a live prediction-market trading system and has paid the tuition on exactly which prices are real and which are fiction.