Prediction market arbitrage: the complete guide
Prediction market arbitrage is buying YES on an event at one venue and NO on the same event at another, so the pair costs less than the $1.00 it must pay out. When both markets resolve on identical terms, exactly one leg pays — whichever way the event goes.
How does a YES + NO hedge work?
Every binary prediction market contract settles at $1.00 (event happens → YES pays; doesn't → NO pays). If "Will X happen?" trades at YES $0.47 on Polymarket while the same question's NO costs $0.505 on Kalshi, the pair costs $0.975 and pays $1.00 either way — a 2.5% gross edge before fees. Nothing about the outcome matters; only that both venues resolve the same question the same way.
Why do prediction markets disagree?
Because the market is fragmented. Polymarket, Kalshi, PredictIt and Limitless serve different jurisdictions, currencies and user bases — a US-regulated exchange, an offshore crypto book and a UK betting exchange each attract different flows, and capital can't move freely between them (KYC walls, geo-restrictions, on/off-ramps). Prices drift apart faster than the small arbitrage community can pull them back together.
How big are real prediction market arbitrage edges?
Smaller than the marketing you'll see elsewhere. On a live board of 40–70 detected pairs, most gross spreads sit between 0.5% and 3%; anything above ~10% is almost always a mismatched pair, a stale quote, or a market you can't actually trade. Annualized returns depend heavily on time-to-resolution — a 2% edge that locks capital for two years is a worse deal than a T-bill.
What are the risks of prediction market arbitrage?
The math is the easy part. The losses come from everything around it:
- Resolution mismatch — "similar wording" is not "same market." Win vs nomination, first-round vs runoff, different price indexes or deadlines: if the two markets can resolve differently, both legs can lose. This is the #1 killer.
- Fees — Kalshi charges a taker fee, PredictIt takes 10% of winning-side profit; a 2% gross edge can be sub-1% net.
- Depth — the displayed price may hold for 20 contracts, not 2,000. Walk the order book before sizing.
- Jurisdiction — Kalshi and PredictIt are US-only; Polymarket and Limitless restrict US persons. Many pairs cannot legally be held by one person.
- Capital lock-up — your money is parked until resolution, which can be years for election markets.
- Taxes — winnings on one venue may be taxable without netting the other venue's losing leg.
Is prediction market arbitrage risk-free?
No. It is a modelled edge that becomes real only if both markets resolve on identical rules, your orders fill at the displayed prices, and the after-fee, after-tax math still clears. Treat every "guaranteed profit" claim as a red flag — including ours, which is why we don't make one.
The honest framing is that you are not taking market risk (you don't care who wins) but you are taking resolution risk, execution risk and liquidity risk. Those are smaller than directional risk, but they are not zero — and when they bite, they tend to take the whole position rather than a slice of it.
What actually happened: our own settled record
Most writing on this topic quotes modelled edges. These are outcomes — every figure below comes from signals PROVISUM published and then followed to the venue's own resolution over the last 30 days. Losses included.
The 19 both-legs-lost results are the important number. Every one was a pair that looked identical across two venues but resolved on a different question — which is the risk this whole page is about. Hypothetical: it assumes both legs fill at the quoted price. Not financial advice.
What this excludes, and why. A further 3 settled trades (0 of which lost both legs, +$30 per $1,000) came from a matching fault we have since fixed — sports markets where one leg covered only part of a match, a specific scoreline, or the opposite team. The scanner can no longer produce those pairs, so counting them as its current failure rate would be misleading. Including them, the full 131 settled trades return −$9,630 per $1,000. We publish both rather than quietly dropping the losses.
Is prediction market arbitrage legal?
Trading on a regulated venue is legal where that venue is licensed — Kalshi is a CFTC-regulated designated contract market and is available to US persons; PredictIt operates under its own regulatory arrangement. Polymarket restricts US persons, and Limitless is on-chain and similarly restricted. Arbitrage itself is not a special legal category: it is just two ordinary trades.
The practical constraint is not legality but eligibility. Holding both legs means holding an account at each venue, and the venue's terms — not your intent — decide whether you may. Opening an account from a restricted jurisdiction, or via a VPN, breaches those terms and typically ends in frozen funds at settlement, which is the worst possible moment. Nothing here is legal advice; check your own jurisdiction.
Can a US person trade both legs?
Usually not on the pairs that look most attractive. Kalshi and PredictIt are US-accessible; Polymarket and Limitless exclude US persons. A hedge that pairs a US venue with a non-US venue therefore needs two people, or two jurisdictions — not one account.
That leaves two genuinely single-jurisdiction combinations: Kalshi + PredictIt for a US trader, and Polymarket + Limitless for a non-US trader. On a typical PROVISUM board, fewer than one signal in twenty is takeable from a single account — which is exactly why we label every row with the jurisdictions it needs rather than quoting a headline number you cannot act on.
How much capital do you need?
More than people expect, because the edge is a percentage of the capital you lock, not of the profit you collect. A 2% net edge on a $1,000 hedge is $20 — and that $1,000 is unavailable until the market resolves.
Three things set the floor. Depth: many books hold their quoted price for only tens of contracts, so a large stake walks the price against you and eats the edge before you finish filling. Venue caps: PredictIt limits each contract position to roughly $850, which caps any PredictIt-legged trade regardless of your bankroll. Fixed frictions: withdrawal fees and on-chain gas do not scale down, so very small stakes are eaten by costs. In practice a few thousand dollars per venue is the point where the arithmetic starts to work, and the binding constraint is almost always order-book depth rather than your balance.
What fees should you subtract?
Fees decide whether a spread is an opportunity or a rounding error:
- Kalshi charges a taker fee of roughly
0.07 × price × (1 − price)per contract, rounded up to the cent — largest for contracts priced near 50c, negligible at the extremes. - PredictIt takes 10% of profit on the winning side, plus 5% on withdrawals. The withdrawal fee is charged when money leaves the venue, not per trade, so it matters far more to an occasional trader than to one recycling capital.
- Polymarket currently charges no trading fee, but it is on-chain: gas and stablecoin on/off-ramp spreads are real costs.
- Limitless is on-chain with the same gas consideration.
A 2% gross spread routinely lands under 1% net. Any tool that shows you gross spreads without subtracting venue fees is showing you a number you cannot earn.
What does a realistic annual return look like?
Judge every edge on an annualised basis, because a hedge locks capital until resolution. A 1.4% edge on a market that settles in three weeks is roughly 24% a year. The same 1.4% on a 2028 election market is about 0.6% a year — worse than a Treasury bill that carries no resolution risk and no lock-up.
This single test removes most of what looks attractive on a raw spread board. It is why PROVISUM refuses to badge a signal as takeable unless its annualised return beats the risk-free rate: an edge that is real but worse than doing nothing is not an opportunity.
Can you automate prediction market arbitrage with a bot?
Partly, and the hard part is not the part people automate. Detection genuinely benefits from automation: comparing tens of thousands of markets across venues every few minutes is not something a human can do. Public bots exist that watch Polymarket and Kalshi and flag price gaps.
Execution is where automation gets dangerous. A bot that fires on a title match will happily buy both legs of two markets that are not the same question — and that loses the entire stake rather than the edge. Venues also rate-limit, require separate KYC, settle in different currencies, and cannot be filled atomically: there is always a window where you hold one leg naked. Our position is that matching and pricing should be automated and the resolution-rule check should not be. That is a judgement call on the meaning of two sentences of English, and it is the exact place where money is lost.
How are the profits taxed?
Treatment varies by country and by venue, and this is not tax advice — but one asymmetry catches people out. Your winning leg is a gain on one venue while your losing leg is a loss on a different venue, often under a different reporting regime. If you cannot net them against each other, you can owe tax on the winning leg while getting no relief for the losing one, turning a thin positive edge negative after tax. Confirm your own treatment before sizing up.
How do you find prediction market arbitrage opportunities?
Manually: watch the same events across 4 venues and compare order books — practically impossible at scale. A workable process has five steps, and only the first four can be automated:
- Crawl every venue's live prices continuously, not on demand.
- Match the same real-world event across venues, with guards strict enough to reject "similar" pairs — different stages, ordinal placements, thresholds, deadlines and chambers all look alike to a naive text match.
- Price both hedge directions net of each venue's fees, and keep the worse of the two outcomes as the guaranteed figure.
- Verify depth by walking the real order books for the size you intend to trade, not the top-of-book quote.
- Read both sets of resolution rules yourself before committing money. Nothing above substitutes for this.
PROVISUM does steps one to four every 10 minutes and shows you what survives — including the signals whose edge disappears once the books are walked, because those are the ones that would have cost you.
See the live gaps
PROVISUM scans Polymarket, Kalshi, PredictIt and Limitless every 10 minutes and surfaces the cross-venue hedges that are actually on the board right now — fee-adjusted, depth-checkable, honestly labeled.
Keep reading: Crypto arbitrage: real cross-exchange spreads, measured · Polymarket vs Kalshi: fees, access & price gaps · Polymarket arbitrage: finding mispriced markets · Oracle risk: when both legs of a hedge lose · Liquidity: why the quoted edge is not the real one