Arbitrage on Polymarket is real, but it is not what the “free money” articles promise. There are three structural kinds — same-market YES+NO mispricing, negative-risk set arbitrage across multi-outcome events, and cross-platform spreads — and all three share the same catch: the profit is pennies, the windows are seconds, and the winners are whoever detects and executes fastest. Here is how each one actually works, what it pays, and what the operators who do this at scale actually look like on-chain.
The invariant everything rests on
Every Polymarket market settles at exactly $1 per winning share. That creates hard arithmetic: in a binary market, YES + NO must be worth $1 at resolution; in an event with N mutually exclusive outcomes, exactly one YES pays $1 — so the sum of all YES prices should be $1, and the sum of all NO prices should be $N−1. Whenever live prices drift from these identities, the difference is, in principle, riskless profit. Arbitrage is the business of collecting that drift.
Kind 1: same-market — when YES + NO ≠ $1
The simplest case. If YES trades at $0.55 and NO at $0.43, buying both costs $0.98 and pays $1 at resolution no matter what happens — a locked $0.02 per pair. (Polymarket even lets you merge a full YES+NO pair back into $1 of collateral immediately, no waiting.) The reverse also exists: when YES + NO trades above $1, you split $1 of collateral into a pair and sell both sides.
Why you rarely see it: this is the most-watched invariant on the platform, and bots close the gap within seconds. The drift appears during fast markets — news shocks, panic sells, thin books — exactly when execution is hardest.
Kind 2: negative-risk — the multi-outcome version
Multi-outcome events (“who will win the nomination”, Fed decision buckets, temperature ranges) are where the real, persistent inefficiency lives, because the invariant spans many order books at once. Two mirror forms:
- Sum of YES < $1: buy every YES; exactly one resolves at $1. Your profit is $1 minus what the full set cost.
- NO sets: collect a NO in every outcome of the event, and Polymarket’s negative-risk machinery (the
convertfunction) turns the complete set into collateral — locking the spread the moment the set is complete.
The more outcomes an event has, the more books have to agree with each other — and the more often, somewhere, they don’t. A 30-outcome event is 30 order books that must sum correctly; nobody’s limit orders update simultaneously.
This is not theoretical. We have spent weeks watching an 8-wallet operation run negative-risk harvesting at industrial scale — hundreds of thousands of tiny fills across ten thousand markets, consolidated off the order book into one converting hub wallet, recycling the proceeds back into more buying. A full breakdown of that machine is coming in its own article. The relevant lesson here: the strategy is real, profitable in aggregate, and already industrialized. You would be competing with that.
Kind 3: cross-platform — Polymarket vs. Kalshi
The same event often trades on both Polymarket and Kalshi at different prices. Published measurements put typical pre-cost spreads around 1.5–4.5% with windows of a few seconds on major events. After fees, slippage, and moving money between a USDC-on-Polygon platform and a regulated USD exchange, the net is far thinner — and one risk dwarfs the arithmetic: the two platforms may resolve the “same” question differently. Each side settles by its own rules and its own oracle. If the rules diverge on an edge case, your “hedged” position is two open bets. Read both rule sets before trusting any spread.
Why bots win — and what the game actually is
Notice what all three kinds have in common. The profit per unit is cents. The mispricing appears unpredictably, anywhere across tens of thousands of order books, and disappears in seconds. That shape of opportunity is not a trading problem — it is a monitoring problem. The winners are not smarter about probability; they simply see more books, more trades, and more flow at once, and they never sleep.
That is worth stating honestly even though we sell data: detecting book-level mispricing needs order book data, and Polymarket’s own websocket is the right source for that. What the order books cannot tell you is who you are competing with and how they operate — that lives in the attributed trade tape and the money flows: the fill patterns of the harvesters, the off-book share movements, the conversion cycles. We only found the 8-wallet machine because the tape has names on it.
The honest scorecard
- Is the profit real? Yes — resolution arithmetic guarantees it, and on-chain records show operators collecting it at scale.
- Is it available to you? Only if you can detect drift across thousands of books and execute in seconds, repeatedly, with capital parked until resolution or conversion. The idea is free; the infrastructure is the moat.
- What kills the “risk-free” label? Partial fills (half a set is a bet, not an arb), fees, capital lockup, and — for cross-platform — settlement-rule mismatch.
- Where should a builder start? Not by racing the incumbents. Start by watching them: replay how the professionals’ fills actually happen, study the flow, and find the corner of the platform where the competition is thin.
The machines that farm these spreads all began the same way — with someone watching the tape closely enough to notice money lying on the table. The tape is available; the noticing is up to you.