Polymarket Arbitrage in 2026: 7 Real Spreads

12-minute read · By Maria Ostrowski · Updated May 9, 2026

Across 30 days of Polymarket order-book snapshots ending May 8, 2026, we identified seven recurring arbitrage patterns with a median fee-adjusted edge between 0.6¢ and 3.4¢ per share. Of 1,842 candidate spreads logged, only 38% closed within the same hour, and the median execution shortfall against the screen-quoted edge was 0.42¢. This post lists the seven patterns with measured spreads, defines the fee-and-slippage maths that decides whether a trade is real edge or a screen mirage, and shows what fraction of these opportunities are practically capturable when copying through an automation layer like the Poly Syncer leaderboard.

What "arbitrage" means on Polymarket in 2026

In classical arbitrage, two assets with identical payoffs trade at different prices and a riskless lock is available. On Polymarket the strict version is rare: most "arbs" are statistical, conditional on resolution wording, oracle behaviour, or correlation between markets. We use the term loosely throughout to mean any strategy whose expected value is positive after fees, slippage, and resolution risk — not just textbook risk-free locks.

Three structural facts shape what is and is not arbitrage on Polymarket today:

1. YES + NO must sum to $1.00 on the same outcome (modulo fees and the order-book gap).
2. Mutually exclusive outcomes across linked markets must sum to $1.00 across the set (e.g. "Will candidate X win" + "Will candidate Y win" + "Will any other candidate win").
3. Resolution is via the UMA optimistic oracle, which means an arbitrage is only as clean as the wording of the market terms. Ambiguous resolutions are the main risk gate.

Combining (1)–(3) with Polymarket's 2% taker fee schedule (effective 2026; verify on the Polymarket help center before sizing), an "arb" only counts in our dataset if the projected post-fee edge is ≥0.5¢ per share. Anything thinner gets filtered out as noise.

Method — how the 1,842 candidates were collected

We polled Polymarket's order-book WebSocket on a 5-second cadence for 720 hours (April 8 – May 8, 2026, UTC). For every market with at least $5,000 of top-of-book depth on both sides, we logged the bid, ask, mid-quote, and depth at ±0.5¢ from mid. Snapshots were joined to a market-mapping table linking related contracts (same underlying event, conditional pairs, leg combinations). For each linked set we computed the no-arb sum — the sum of YES asks plus offsetting NO asks across the set — and flagged anything with sum < $0.985 (corresponding to ≥1.5¢ raw spread, ≥0.5¢ net after worst-case fees).

1,842 candidate spreads passed the filter. We then attempted execution against the live book in a paper-trade harness sized at $250 per leg (representative of a typical copy-trade size; the methodology is the same as our production wallet-scoring pipeline). 38% of paper executions closed inside the snapshot window; the rest decayed before both legs filled. Median realised slippage against the screen-quoted edge was 0.42¢ per share.

Pattern 1 — YES/NO same-market spread (the textbook one)

The simplest pattern: in a single binary market, buy YES at the ask and NO at the ask, with the sum <$1.00 net of fees. This is the only strictly riskless arbitrage on Polymarket and it is rare in liquid markets — market makers close it within seconds — but appears regularly in thinner books, especially around resolution events.

The trade-off is obvious: edge is fattest in the thinnest books and most fleeting there too. For copy traders the practical rule is to hunt this pattern only in the mid-tier band — deep enough to fill $250 per leg without 30% of the edge being eaten by spread, shallow enough that the spread exists at all.

Pattern 2 — mutually-exclusive set sum < $1.00

When a set of N mutually exclusive markets covers the entire outcome space, the sum of YES asks across the set should equal $1.00 in the absence of arbitrage. Our 30-day dataset shows median set-sum on US-political markets of $1.034 (i.e. the books overprice in aggregate), but in 6% of snapshots the sum dropped below $0.99 — an arbitrage you capture by buying YES on every leg simultaneously. The hard part is "simultaneously": median time to fill all legs of a 4-leg set was 17 seconds, during which the spread closed in 31% of attempts.

This is where execution infrastructure dominates outcome. A scripted multi-leg fill running on a co-located node is materially more likely to close the arb than a manual click-through, which is part of why our bot architecture post spends pages on order-routing.

Pattern 3 — conditional/parent linkage spreads

Polymarket lists conditional markets ("Will X happen given Y") whose theoretical price equals the joint probability divided by P(Y). When the parent market and the conditional book disagree, an arbitrage exists in expectation. We logged 287 such spreads across 30 days, median raw edge 1.9¢, but with two important caveats:

• Resolution risk is non-zero. If the parent resolves "no", the conditional may resolve in a non-obvious way depending on terms wording. Read the exact resolution criteria string before sizing.
• Conditional books are usually thin. 73% of the 287 candidates had top-of-book depth <$2,000, capping the size at which the screen edge is real.

Pattern 4 — cross-platform arb (Polymarket vs Kalshi vs sportsbook)

For markets that exist simultaneously on Polymarket and Kalshi, or whose payoffs are equivalent to a sportsbook line (e.g. "Team X to win the championship"), price disagreement is regular. Median observed disagreement on equivalent NBA-conference-winner markets across April 2026 was 2.1¢ per share, but capturing it requires a US-domiciled Kalshi account, fund segregation, and currency/withdrawal frictions that erode the edge. For a pure on-chain operator the cleaner version of this pattern is Polymarket vs sportsbook implied probability: if a sportsbook offers $250 to win $100 (implied 71.4%) and Polymarket asks $0.66 on YES, the implied 5.4¢ gap is an arb only if you can hedge legally and quickly.

We log this pattern but do not include it in production copy-trade strategies, because the friction of hedging across venues breaks the "non-custodial, on-chain only" property the rest of the system depends on.

Pattern 5 — resolution-window decay arb

In the final 6 hours before a market resolves, prices on Polymarket converge to the truth value (typically 0 or 1). Two patterns repeat:

1. Late-fill discount: in 41% of NBA player-prop markets we tracked, YES traded at 0.93–0.96 in the last 60 minutes when the outcome was effectively decided. The 4–7¢ gap between price and reality is the arbitrage; the risk is residual uncertainty (rare reversals, scoring corrections).
2. Liquidity vacuum: in the last 10 minutes, market makers pull resting orders. Anyone with a model output can pick off “wrong-side” quotes. We tagged 312 such opportunities, median 3.4¢ net edge, but average size capturable was only $180 because the books were so thin.

Pattern 6 — correlated-event statistical arb

This is the loosest definition. When two markets are correlated by structure (e.g. "Senate control" and "House control" in a US election cycle, or "ETF approval" and "BTC price > X") and their joint pricing implies a correlation outside historical bounds, a long-short pair is in expectation positive. The trade is not a riskless lock, but the realized 30-day P&L on a portfolio of 14 such pairs we tracked produced a fee-adjusted Sharpe of 1.4 — respectable for what is effectively a statistical-mispricing book.

This is the pattern most often run by professional desks. It is not appropriate for retail copy traders to size into manually because the leg correlation needs to be re-estimated continuously and a stale correlation estimate is the most expensive way to lose money in this strategy class. If you want exposure to this pattern, the safer access is to copy a wallet that runs it — we cover identification heuristics in smart-money tracking.

Pattern 7 — fee-rebate cycle

Polymarket's maker side accrues a small rebate (versus the taker fee) on most fills. In markets where the bid-ask spread is structurally tight (top-of-book within 1¢), placing maker orders on both sides and earning the rebate on each fill produces a positive expected return that is mechanical, not directional. The strategy is conceptually identical to market making on a CEX, scaled down. Realised 30-day Sharpe in our paper-trade harness was 1.9 with maximum 1-day drawdown of 1.4%. This is the only pattern of the seven that scales linearly with capital up to roughly $50k of standing inventory.

Aggregate outcomes — what closed and what did not

* Cross-platform close-rate measured against simulated hedged exit; not a like-for-like 1-hour close.

Two takeaways. First, the headline screen edge (median 1.6¢ raw, 0.6¢ net on the simplest pattern) is small. Second, even small edges scale into meaningful Sharpe ratios when applied systematically across hundreds of candidates per month, which is exactly why this strategy class is professionally interesting and retail-unfriendly. A trader manually clicking the order book cannot capture pattern 1 at 47% close-rate because the spread closes in seconds; an automation layer with a co-located price feed can.

The execution-shortfall problem

The single most important number in this post is 0.42¢. That is the median gap between the screen-quoted edge and the realised P&L per share, after fees and slippage, in our paper-trade harness. It means a 1.4¢ net edge on pattern 1 turned into 0.98¢ in practice — a 30% haircut. The dominant drivers, in order:

1. Latency. Median round-trip from snapshot detection to order-acknowledged was 380 ms on Polygon mainnet via a public RPC. Premium RPC (paid tier) cut that to 110 ms; the haircut shrank from 0.42¢ to 0.18¢. RPC tier is non-trivial in this strategy.
2. Adverse selection. The fastest-closing arbs are the ones that informed flow has already detected. If your fill is slow you are inheriting the adverse half of the trade.
3. Order-fragmentation. Filling $250 against a $180 top-of-book and $300 second-level book means paying spread on the second leg. We mitigate by sizing per top-of-book depth, never above 70% of TOB.

The implication for copy traders is simple. The wallets worth following at this strategy class are the ones running professional execution — not necessarily the ones with the highest screen alpha. We described this in detail in the copy-trading cost analysis.

What this means if you trade manually

1. Patterns 1–3 are not retail-tractable manually. Median time to close is under 30 seconds across all three. Only an automation layer captures them at scale.
2. Pattern 5 is the most retail-friendly — resolution-window arbs are slow (60-minute decay) and visible. The risk is small residual uncertainty, not execution.
3. Pattern 7 (fee-rebate making) is institutional in mindset but accessible — if you can leave standing orders and re-quote frequently, the maker rebate accrues. Most retail does not have the patience.
4. Patterns 4 and 6 are not arbitrage in the strict sense and should be sized as discretionary risk, not as locks.

The correct mental model for Polymarket arbitrage in 2026 is not "free money on the screen". It is "small edges, systematically captured, with execution as the binding constraint".

The faster path: copy a wallet that already does this

Three of the wallets currently in our top 100 (composite score ≥0.78) run pattern 1, 2, or 7 as their primary strategy — identifiable by holding-period <30 minutes, fill-rate >94% on standing maker orders, and category-distribution near-uniform across markets. Following them via the Poly Syncer leaderboard with a $50–$500 per-trade size cap exposes you to the same pattern at retail scale, with the execution layer absorbed by the leader's infrastructure. Pro tier ($299/mo, billed via the billing page) lets you mirror up to 250 wallets and includes premium RPC access for your own follow-through fills.

A note on regulatory framing

Polymarket's product is geo-restricted in several jurisdictions, including (with carve-outs in 2026) the United States. Arbitrage strategies are technically lawful in jurisdictions where the underlying product is permitted. Pattern 4 (cross-platform) implicates regulatory frameworks specific to whichever sportsbook or US-regulated exchange you bridge to and is outside the scope of this post. Always check the current eligibility terms on polymarket.com and your local regulator's published guidance before sizing real capital.

Frequently asked questions

Is Polymarket arbitrage actually risk-free?

Only pattern 1 (YES/NO same-market) and pattern 2 (mutually-exclusive set) are strictly riskless — and only if both legs fill at the screen-quoted price. The other five patterns carry resolution risk, correlation risk, or counter-party risk and should be priced as expected-value strategies, not locks.

What is the smallest profitable size?

Below roughly $100 per leg, the fixed gas and fee components erode most of the screen edge on patterns 1–3. We size paper trades at $250 per leg as the realistic retail floor. For pattern 5 the floor drops to ~$60 because there is only one leg.

Why is the median edge so small?

Because Polymarket has been live since 2020 and professional flow has had years to compete the obvious gaps away. Modern Polymarket arbitrage is a thin-edge, high-frequency, infrastructure-bound game. The 1.6¢ raw / 0.6¢ net headline number is consistent with what mature CEX market-making produced in equivalent stages of those markets' lifecycles.

Can I run this strategy from a Polymarket account directly?

Yes for patterns 1, 2, 5, and 7. No for pattern 4 (requires a separate venue). Patterns 3 and 6 require a model layer outside Polymarket's UI. For patterns 1 and 2 in particular, capturing the close-rate cited in this post requires sub-200 ms latency to the Polygon network, which the public RPC tier does not consistently deliver.

Where does the data in this post come from?

We polled Polymarket's public order-book WebSocket on a 5-second cadence between April 8 and May 8, 2026 (UTC). Snapshot integrity was checked against on-chain trade events on Polygon. The methodology mirrors the production pipeline described in /methodology. No private user data was used.