Prediction Market Arbitrage: How It Works, With Worked Examples

Prediction market arbitrage is buying both sides of the same question for a combined price under $1 — Yes and No in a single market, or Yes on one platform and No on another — so that the guaranteed $1 settlement pays more than the position cost. The gaps are real: a 2025 study of Polymarket’s on-chain order data (Saguillo et al.) put realized arbitrage profit on that venue alone at roughly $40 million. The catch is everything after the screenshot: at the fee schedules of August 2026, a 2-cent gap nets under a penny per hundred contracts in Polymarket’s politics markets and a guaranteed loss on Kalshi’s standard markets — and the cross-platform version carries the least appreciated risk in this industry, two contracts on the “same” event resolving under different rules.

Here is the machine, with current fee formulas and worked numbers. This article documents how it works; it is not a recommendation to attempt it.

How does same-market Yes/No arbitrage work?

An event contract pair — Yes plus No — always pays exactly $1 at resolution, so when both sides can be bought for under $1 combined, holding them to settlement locks in the difference. A politics market quoting the Yes ask at 47¢ and the No ask at 51¢ offers a 98¢ package guaranteed to pay $1: a 2¢ gross profit per pair, about 2% on capital, no forecast required.

Then the fee formulas arrive. Polymarket charges takers shares × rate × price × (1 − price), with the rate set by category — 0.04 for politics as of August 2026, per its fee schedule (full table in our Polymarket fees explainer). On 100 pairs: the Yes leg costs $47 plus a fee of 100 × 0.04 × 0.47 × 0.53 = $1.00, and the No leg costs $51 plus 100 × 0.04 × 0.51 × 0.49 = $1.00. Total outlay: $99.996 for a package worth $100.00. The visible 2¢ gap nets four tenths of a cent — on a $100 position.

Run the same quotes through Kalshi’s formula — taker fee = 0.07 × contracts × price × (1 − price), rounded up to the cent, per the exchange’s fee schedule (July 2026 revision) — and the trade goes negative: $1.75 per leg, $3.50 against $2.00 of gross edge, a guaranteed $1.50 loss. Near mid-range prices, a Yes/No gap must exceed roughly 3.5¢ on Kalshi’s standard markets, or 2¢ in Polymarket’s politics category, before a taker keeps anything; only in Polymarket’s geopolitics markets, fee-free as of August 2026, does the whole 2¢ survive.

These gaps barely exist on one book anyway: buying No at 51¢ is the same trade as selling Yes at 49¢, so the quotes above form a crossed book, and matching engines pair Yes-buyers with No-buyers directly. What the Polymarket study calls market rebalancing arbitrage, mispricing within a single market, appears in bursts of fast news and disappears in seconds. Its second form, combinatorial arbitrage, lives across linked markets: a five-candidate event whose Yes prices sum to 96¢, a “by March” contract priced above the “by June” contract that contains it. Those persist longer, because closing them takes capital in several markets at once.

How does cross-platform arbitrage work between Kalshi and Polymarket?

The cross-platform version needs no crossed book — just two venues disagreeing. Suppose a September Fed rate-cut market trades at 44¢ Yes on Kalshi while the equivalent Polymarket market trades at 48¢ Yes, pricing its No side at 52¢. The quotes are illustrative; the fee math is current as of August 2026, with the Polymarket leg in its economics category (rate 0.05).

Leg	Order	Cost	Taker fee	Total
Kalshi	Buy 100 Yes at 44¢	$44.00	0.07 × 100 × 0.44 × 0.56 = $1.73	$45.73
Polymarket	Buy 100 No at 52¢	$52.00	0.05 × 100 × 0.52 × 0.48 = $1.25	$53.25
Combined		$96.00	$2.98	$98.98

A cut pays the Kalshi leg $100; no cut redeems the Polymarket No shares for $100. Either way, $100 arrives against $98.98 deployed: $1.02 locked in, about 1% over the life of the trade, roughly 9% annualized on a six-week resolution — respectable, until the risk section below.

The trade also demands funded accounts on both platforms, each leg fully collateralized — no clearinghouse nets one against the other — and the $100 lands on whichever venue won, so half the bankroll must travel home before the next trade. Kalshi’s ACH withdrawals are free but can take up to five business days; Polymarket’s are instant and free in USDC on Polygon, but reaching bank dollars takes an off-ramp with its own fees. Capital in transit earns nothing, and that drag compounds faster than a 1% edge.

Why do the same events trade at different prices?

Because the two books belong to populations that mostly cannot trade against each other. As of August 2026, Polymarket’s global exchange remains closed to US persons while its regulated US exchange lists a narrower catalog, and Kalshi operates inside the US regulatory perimeter on dollar rails. Different user bases bring different information and reflexes to the same headline — and the traders who would grind the gap to zero mostly can’t be on both books at once.

Fee structures widen the wedge — a fee-free category on one venue against a 7% expected-profit formula on the other moves each break-even differently — and so does collateral: dollars in a CFTC-regulated clearing system on one side, stablecoins in smart contracts on the other, no shared margin. And some of every gap is not mispricing but basis: two contracts on “the same” event are only the same if their resolution criteria are.

What eats the “riskless” profit?

Resolution mismatch — the killer. Consider a pair of markets on whether the federal government shuts down by a date. One venue’s contract defines the event as a lapse in appropriations in effect at a specific hour; the other asks in plain language whether a “shutdown” happens, as read by its resolution process. Congress blows through the midnight deadline and passes the funding bill at 9 a.m.; no agency ever closes. The lapse-definition contract resolves Yes; the plain-reading market can credibly resolve No — and a trader holding Yes on the plain-reading venue and No on the lapse venue watches both legs settle at zero. The mirror allocation would have paid $2, but the payoff shape is the point: risking roughly $99 to win $1 means a mismatch that bites once in a hundred trades erases every win.

This is not hypothetical — Polymarket’s March 2025 Ukraine-minerals market resolved Yes with no agreement in existence — and the two venues use different referees entirely: Kalshi resolves under a rulebook filed with the CFTC, while Polymarket’s global exchange delegates to the UMA optimistic oracle, where whitelisted proposers post outcomes and disputes go to a token-holder vote; our Polymarket review covers the failure cases.

A cross-platform arbitrage is a bet that two rulebooks describe the same world — at 99-to-1 odds against the trader when they don’t.

Legging risk. The two orders can’t fill simultaneously; a move after the first fill turns the arbitrage into a plain directional bet at a worse price — precisely when fast news is moving both books.

Depth and size limits. A quoted gap is only good for the contracts resting at the quote — a 2¢ edge dies on 1¢ of slippage, which makes reading the book’s depth the core skill — and Kalshi’s rulebook adds hard ceilings: position limits set contract by contract, a $25,000 default with far higher caps on flagship series, per its CFTC-filed rules.

Locked capital and timing drift. Both legs generally ride to settlement — exiting early means new taker fees and two more spread crossings — the venues may settle the same event days apart, and every extra day the “riskless” return races the Treasury yield it needs to beat.

Quick answers

What is prediction market arbitrage? Buying both sides of the same event — Yes and No in one market, or opposite sides on two platforms — for less than $1 combined, so the guaranteed $1 settlement exceeds the cost. Profit comes from the price inconsistency, not a view on the outcome.

How big does a gap need to be to clear fees? As of August 2026, with both legs as taker orders near mid-range prices: roughly 3.5¢ on Kalshi’s standard markets, 2¢ in Polymarket’s politics category, 2.5¢ in its 0.05-rate categories, and anything above zero in fee-free geopolitics — before slippage and spreads.

Can gaps between Kalshi and Polymarket actually be traded? Mechanically yes — the worked example above nets about 1% at current fee schedules. Practically, the venues serve legally separated user bases (Polymarket’s global book excludes US persons as of August 2026) and each leg needs separate funding, so the trade is harder to assemble than the screenshot suggests.

Is prediction market arbitrage risk-free? No. Resolution-criteria mismatches can zero both legs at once, legs can fill at worse prices than quoted, depth and position limits cap size, and capital sits locked on two venues until settlement. “Riskless” describes the payoff diagram, not the trade.

How much arbitrage profit has actually been extracted? A 2025 academic study of Polymarket’s on-chain order books (arXiv:2508.03474) estimated about $40 million in realized arbitrage profit, split between rebalancing within single markets and combinatorial trades across logically linked ones.

Why don’t the platforms’ prices just converge? No common clearing, no shared user base, different fees, different collateral rails, and subtly different contracts. Some of every gap is the market correctly charging for the possibility that the two versions of the “same” question resolve differently.