Does Prediction Market Arbitrage Work? An Honest 2026 Answer

Does prediction market arbitrage work as advertised? The screenshot version is easy: Kalshi at 62 cents, Polymarket at 66, look at the gap. Reading past the screenshot means going through two fee formulas, the size resting on the thinner leg, both rulebooks, and the question of which exchange you can open an account on in the first place.

We have already covered how the trade is constructed. This is the other question, checked against live exchange APIs and venue documentation on 8 August 2026.

Does Prediction Market Arbitrage Work in Principle? The Gaps Are Real

The gaps themselves are genuine. The largest study so far, a January 2026 preprint covering over 100,000 events across ten venues from 2018 to 2025 (Gebele and Matthes, arXiv:2601.01706), found cross-venue deviations on equivalent markets averaging 2 to 4 percent on an execution-aware basis, persisting even in liquid settings. The authors blame structural frictions rather than disagreement about the world. Gaps survive because closing them is expensive.

The same paper has the number nobody quotes: roughly 6 percent of events are listed concurrently across platforms. The other 94 percent have no second leg at all. For the mechanics, see our guide to how cross-venue arbitrage is constructed. What follows is what happens after you construct it.

Which Venue You Can Reach, and What It Charges

Venue access decides this before fees do. Kalshi is a CFTC-designated contract market open to US persons. Polymarket is two exchanges, not one. Polymarket International (Adventure One QSS Inc.) is not CFTC-regulated and lists the United States as close-only on both frontend and API at the time of writing, meaning an existing position can be closed but a new one cannot be opened. Polymarket US (QCX LLC) is a separate CFTC-designated contract market with its own market list and its own fee schedule. Check the geoblock reference on docs.polymarket.com, and your own jurisdiction, before pricing anything below.

Both venues price event contract fees the same way: a coefficient times contracts times p times (1 minus p), peaking at a contract price of 50 cents. On Polymarket International takers pay 0.07 in crypto, 0.05 in sports, economics, culture, weather and other, 0.04 in politics, finance, tech and mentions, and nothing in geopolitics, while makers are never charged. On Polymarket US there is one taker coefficient, 0.06, across every category, and makers earn a rebate rather than paying. Kalshi uses the same quadratic form with a per-series multiplier, rounded up per order. Rates as published at the time of writing: docs.polymarket.com/trading/fees, docs.polymarket.us/fees, Kalshi's published fee schedule, and our own breakdowns of Kalshi fees and Polymarket fees.

That is breakeven gross spread per contract pair, taker on both legs, at a Kalshi multiplier of 1. The four middle columns add one Polymarket International category coefficient each; the last adds the flat 0.06 Polymarket US coefficient. Two notes for reading it: p times (1 minus p) is symmetric, so the $0.60 row also prices the $0.40 leg opposite it, and a Kalshi series running at a 0.5 multiplier halves the Kalshi column and everything built on it.

Set that against the 2 to 4 percent deviation and the comparison is looser than it looks. The paper reports an average, not a median, and does not say whether the deviation is 2 to 4 cents on a dollar contract or 2 to 4 percent of price. If it is cents, the typical dual-listed gap sits at or near a 3.5 cent crypto breakeven and below it in several categories, which is an order-of-magnitude check rather than a settled result. The coefficients move too: nineteen Kalshi MLB series shifted to a 0.5 multiplier on 7 August 2026. Before sizing anything, pull GET /trade-api/v2/series/fee_changes, which needs no authentication and returns the fee type and multiplier per series.

One Live Pair, Priced All the Way Through

On 8 August 2026 we pulled two markets that look like the same question. Kalshi's KXIMPEACH-29-JAN20 showed a YES ask of $0.62 against a 61 cent bid; Polymarket International's Trump impeachment market showed a best bid of 0.66. YES and NO in the same market sum to $1, so that bid implies a NO ask of 0.34. Commit $0.96 for $1.00 of payout: 4.00 cents gross, 4.17 percent on capital.

Kalshi's taker fee on 100 contracts at $0.62 is 0.07 x 100 x 0.62 x 0.38, or $1.6492, rounded up to $1.65, so 1.650 cents per pair. Polymarket International's politics rate adds 0.898 cents. Drag of 2.548 cents leaves 1.452 cents net, 1.513 percent on capital, with fees taking 64 percent of the gross. The money then sits until 20 January 2029, which annualises to about 0.615 percent. Price the same pair against Polymarket US, where the taker coefficient is 0.06 everywhere, and that leg costs 1.346 cents instead: drag 2.996 cents, net 1.004 cents, 1.046 percent on capital, roughly 0.43 percent a year. That second version is the one a US trader is looking at, assuming the market is listed on the US exchange at all.

Size decides the rest. That Kalshi ask was quoted for 0.96 contracts, which is what the book returned rather than a rounding artefact: the orderbook endpoint reports resting size in fixed-point units, so a price level can genuinely rest under a full contract. The whole trade therefore nets about one cent, in total, not per contract. The impeachment pair was at least tight, 61 bid against 62 ask. Long-dated books elsewhere are routinely several cents wide, and crossing a spread wider than the edge you are harvesting is the least discussed way this loses money.

The Trap: Both Legs Can Lose

Here is what the fee math hides. Kalshi's rules resolve YES if the President of the United States has been impeached before 20 January 2029, with a secondary condition in the series rules: the President must be in office at the time an impeachment resolution passes. Polymarket's counterpart is titled "Will Trump be impeached before his term ends?" and resolves YES if the House approves articles of impeachment of President Trump by simple majority before 20 January 2029, with neither trial, conviction, nor removal required. One contract is about an office. The other is about a person. (Kalshi series KXIMPEACH via api.elections.kalshi.com; Polymarket via gamma-api.polymarket.com, slug will-trump-be-impeached-before-his-term-ends.)

The confirmed divergence runs one way. Impeach a successor while that successor is in office and Kalshi resolves YES, while Polymarket, which names Trump, resolves NO. Anyone holding Kalshi NO against Polymarket YES loses both legs on that branch. The reverse branch, a post-term impeachment of Trump himself, is not settled by the published wording: the title says "before his term ends" and the description says "President Trump", so whether it would resolve YES after he left office is a question the rulebook does not answer. Treat that one as wording risk rather than a known outcome. We could not find a named trader caught by either branch, so this is a documented mechanism, not a documented casualty.

The divergences run deeper. Kalshi expires at 15:00 UTC on 20 January 2029 while Polymarket counts through 12:00 PM ET, leaving two hours where one venue is shut and the other is still counting. Polymarket's API also returns an end date that contradicts its own prose rules, so a scanner keyed to structured data reads a deadline that does not govern. Settlement then splits three ways. Kalshi settles KXIMPEACH from congress.gov, named in the series settlement sources field, a public record that either shows articles passed or does not. Polymarket International names the same class of source, information from the federal government of the United States plus consensus of credible reporting, but attests and disputes it through a UMA optimistic oracle with a $500 bond, which resolves by token-holder vote inside a dispute window rather than by exchange determination. Polymarket US settles under its own CFTC-regulated rulebook and does not use UMA. Two legs can therefore settle on different answers, or days apart. Part of that 4 cent gap is payment for exactly that risk rather than mispricing.

Bots, Scanners, and Calculators: What Software Can Check

A prediction market arbitrage calculator is short enough to paste into a spreadsheet. Let Pk and Pp be the two ask prices, n the number of contract pairs, m the Kalshi series multiplier, and cp the Polymarket coefficient (0.07 crypto, 0.05 sports and economics, 0.04 politics and tech on International; a flat 0.06 on Polymarket US):

• gross = 1 - (Pk + Pp)
• kalshi_fee = roundup_to_cent(m * 0.07 * Pk * (1 - Pk) * n) / n
• poly_fee = cp * Pp * (1 - Pp)
• net = gross - kalshi_fee - poly_fee

Substitute the live pair, with Pk 0.62, Pp 0.34, m 1, n 100 and cp 0.04, and it returns 4.00 cents gross, 1.650 cents of Kalshi fee, 0.898 cents of Polymarket fee, 1.452 cents net. Swap cp to 0.06 for Polymarket US and the net falls to 1.004 cents. A scanner is harder but still tractable: quote deltas, fee-adjusted spreads, top-of-book size (the best price only, with no view of the depth resting behind it), and late-resolution gaps where one venue has settled and the other still trades.

What no scanner checks is the previous section. The public scanners we reviewed do not attempt cross-venue market matching, and none of them reconcile resolution text, settlement source or settlement time. The venue APIs make the same point without appealing to anyone's code: Kalshi's order book endpoint requires signed authentication, so a free scanner sees top of book and overstates what you can fill, and a June 2026 paper, Shen and co-authors, The Ghosts of Polymarket (arXiv:2606.16852), counted 980,133 filled Polymarket orders reverted between matching and on-chain settlement. That is a count over the paper's sample and not a rate, so read it as evidence that un-fills happen at scale rather than a probability you can price; a leg that un-fills leaves you naked on the other venue. We build prediction market monitoring tools at Predictefy, so treat this as an interested party's view: price and size are the automatable half, and rule equivalence still ends with a human reading two rulebooks. Our rundown of prediction market APIs covers the build path.

Is Prediction Market Arbitrage Profitable, and For Whom?

Makers face different arithmetic. On Kalshi's plain quadratic series, which includes that impeachment market, makers pay nothing; maker fees apply only to designated series, overwhelmingly sports. Polymarket makers are not charged on either exchange, and on Polymarket US they are paid a rebate. Both-maker breakeven is close to zero, which is the honest reason desks do this and you probably cannot: resting orders on two venues gives no assurance both fill, and one filled leg is a directional position you did not choose. Taker costs are tiered on both Polymarket exchanges, on different ladders. International runs six tiers off 30-day weighted volume, from 3 percent back at $2,000 to 50 percent above $10 million; Polymarket US starts at 10 percent on $250,000 of prior-month volume and reaches 50 percent above $10 million. Weighted volume is not raw notional, so the bottom tier is further away than the headline threshold suggests. Tiers at the time of writing, per docs.polymarket.com/programs/taker-rebates and docs.polymarket.us/fees.

Capital is the other constraint. Nothing nets across venues: $0.62 sits on Kalshi and $0.34 on Polymarket, neither offsetting the other, until resolution. Depth then caps deployment long before your balance does. A 2026 study of Polymarket NBA books (Cheng, Yang and Zou, arXiv:2605.00864) reconstructed over 75 million order book snapshots across 173 games and found single-market anomalies, meaning YES and NO priced under a dollar together inside one book, exceedingly rare: seven executable in-game episodes, median duration 3.6 seconds. Of the more frequent combinatorial opportunities, 76.9 percent were capped at an average executable size of 14.8 shares. Neither figure measures a cross-venue Kalshi and Polymarket gap, which nobody has published. None of this is financial, tax, or legal advice.

Frequently Asked Questions

Conclusion

Prediction market arbitrage clears at institutional scale, on maker flow and rebate tiers, and is close to structurally unavailable to a retail trader crossing the spread on both legs. The pair above is Kalshi against Polymarket International; a US trader's version runs against Polymarket US, a different exchange with a different market list and a worse taker coefficient. If you want an edge rather than an arb, the slower strategies in our guide to making money on Polymarket are a better use of the same capital. Reprice every number here against the live APIs before you act on it.