Prediction Market Price Disagreement

Prediction markets can show different prices for the same event for several reasons, and more than one can apply at once: the two contracts are not the same claim under their own rules, the quotes cannot be traded at your size on both sides, fees and settlement mechanics change the net payoff, or the traders on each venue hold or receive different information. Start by separating two cases. Case A — guaranteed payoff surplus: if the complete portfolio has a strictly positive minimum net payoff across all admissible contractual resolution states, information asymmetry is not required to establish its theoretical payoff advantage. It still needs checks for executable prices, sufficient depth, explicit fees, funding costs, legging risk, settlement, and counterparty exposure. Case B — non-guaranteed opportunity: if an admissible state produces a loss, you need an independently justified probability or expected-value assessment. Information differences may be one hypothesis, but they are neither automatically established nor the only possible source of edge. A residual price difference alone establishes neither information advantage nor positive expected value; when the evidence is insufficient, retain an unresolved classification. Use the four questions below to make those distinctions before acting.

Research is a reason to diagnose before acting. A January 2026 arXiv preprint’s semantic-alignment dataset covers more than 100,000 events across ten venues from 2018 to 2025. Its price-deviation analysis covers seven of those ten venues and reports execution-aware deviations of approximately 2–4% under its stated measurement and execution-cost assumptions.¹ The authors attribute the persistence to structural frictions rather than informational disagreement. That is a historical result under the authors’ standardized cost assumptions. It is not a September 2026 market measurement, and it does not show that each reported deviation was independently exploitable by an individual trader.

This article is about the diagnosis. It does not teach arbitrage tooling, and it does not cover how to check whether your own probabilities are calibrated, which prediction market probability calibration does. It also does not measure fill quality; how execution shortfall is measured belongs to trading slippage and execution costs. What follows is the step before both: deciding what kind of gap you are looking at.

How do you tell whether a prediction-market price gap is an edge?

Save the full resolution rules, and their version, for both contracts, then compare what each pays in every possible outcome, and what the combined position pays in each, not what the titles say.
Price each leg at a quote you could execute: the ask if you buy, the bid if you sell, at your intended size and with a timestamp. Check what you hold if only one side fills.
Subtract fees and other explicit costs. The spread and price impact at your size are already in an executable price, so do not subtract them again; account separately for the cost of tying up capital, and treat settlement delay and one-sided fills as risks to record.
If an admissible state can produce a loss, ask what independently justified probability or expected-value assessment supports the opportunity. Information differences are one hypothesis, not a mandatory explanation.
Write down your classification, your reasoning, and what would prove you wrong before you trade.

How should you classify a gap?

The five labels below are the ones this article uses. Semantic, Structural, and Informational are possible causes; Non-executable is a feasibility status at the specified size, price, and time; and Unresolved is an evidence status. Multiple causes can coexist with a feasibility status and a confidence judgment. A non-executable quote is not necessarily the original cause of the disagreement, and an informational label remains a hypothesis.

Classification dimension and observed signal	Label or possible cause	Evidence required	What remains uncertain
Possible cause: titles match but rules differ in source, data version, cutoff, threshold, or void terms	Semantic: the contracts are different claims, and that difference may account for some of the gap	Saved rule text from both venues showing a payout-relevant difference, which shows the contracts are not identical but not how much of the gap it causes	How much of the gap the difference explains; which wording is worth more; how often the rules produce different outcomes
Feasibility status: the quote is a midpoint or last trade, or depth is smaller than your specified size	Non-executable: you cannot trade both sides at that price and size	Bid or ask side, depth, UTC timestamps, realized or simulated fills	Whether the depth is still there when you trade; what a one-sided fill leaves you holding
Possible cause: fees, settlement timing, capital lock-up, or venue mechanics change the net result	Structural: frictions may remove the net opportunity while the displayed price difference remains	The fee schedule for that contract and order type; settlement timing; your cost of capital	Fees at the actual fill; the true funding cost; how long settlement takes
Possible cause: a residual difference is consistent with asynchronous information or dated evidence shows one venue moved first	Informational (a hypothesis): the groups hold different beliefs or learned news at different times	Timestamped evidence that one venue moved first after a dated event, and a prediction recorded before resolution	Whether a gap closing shows it (it does not); quotes alone rarely show why prices differ
Evidence status: rules, fees, or causal evidence are ambiguous, missing, or conflicting	Unresolved: no diagnosis is supported yet	A written list of what is missing	Everything; do not treat it as an edge

For each gap you consider, record four separate judgments: the primary explanation, any secondary contributing factors, whether the trade is feasible at your size, and how confident you are in the evidence. A gap can be feasible but poorly explained, or well explained but not tradable. Different wording is not automatically a different exposure; the test is whether terminal payouts and the conditions that resolve them are equivalent. Nothing in this table says one venue is right and the other wrong; two quotes that differ do not show which is closer to any objective probability.

What counts as a price disagreement?

A price disagreement, as used here, is any case in which two quotes that appear to describe the same question imply different probabilities. The most common case is the same event listed on two venues, which is this article’s focus. A less obvious case is a set of linked contracts on one venue that should add up to a fixed total but do not.

Six terms need to stay apart:

Price is what a contract costs on a venue.
Market-implied probability is a price read as a probability under that contract’s payoff and fee terms. It is an interpretation, not an objectively correct probability.
Executable quote is the price at which you could actually buy or sell your intended size when you look.
Payoff is what a contract pays in each possible outcome.
Expected value is the probability-weighted payoff net of cost. It needs probabilities that a price gap does not supply.
Arbitrage is a portfolio whose net payoff, after costs, is nonnegative in every outcome the contracts’ rules allow and positive in at least one. A stricter test requires a strictly positive minimum net payoff. Both are statements about payoffs, not about whether you can execute, fund, and settle the position safely.

The claim that price gaps are informative rests on a shortcut: that the two prices are quotes on one question, so the difference must reflect what traders know. Each part of that shortcut can fail. The contracts may not ask the same question. The quotes may not be prices you can trade. And the two groups of traders may not be pricing the same thing at the same moment.

Question one: is it really the same economic exposure?

Four things get treated as one, and they are separate tests:

Same real-world event. Both contracts concern one underlying occurrence.
Compatible resolution criteria. The source, data version, threshold, and cutoff lead to the same outcome across all admissible states under the applicable contract rules. The wording need not match, but you have to check.
Identical state-contingent payoffs. In every state of the world that can occur, both contracts pay the same amount.
Equivalent timing and payment reliability. Both settle on a comparable schedule and pay reliably.

A matching headline establishes only the first. The relevant theoretical test for treating two contracts as one exposure is the third: list the possible outcomes and check what each contract pays in each. The fourth is operational, and it stays separate because two contracts can pay identically on paper and still differ in when, and how reliably, the money arrives.

The rules are the contract. Polymarket’s documentation states that every market has pre-defined resolution rules specifying a resolution source, an end date, and how edge cases are handled, and that the title describes the question while the rules define how it resolves.² Kalshi’s help center likewise says the rules for determining an outcome, the information used, and its source are included in each contract’s terms.³ Polymarket also says it may issue an “Additional context” clarification after trading has begun, which is why the version of the rules you traded under belongs in your record.²

Gebele and Matthes built an identity test into their research. They treat two listings as one event only after jointly reading the natural-language descriptions, the resolution semantics, and the temporal scope, and they call a failure of that test semantic non-fungibility.¹ Compare these items in the two rulebooks:

Resolution authority and source. The same body and the same data series? Who decides if the source is unavailable?
Release versus revision. A first published figure or a later revised one?
Threshold and cutoff. “Above” or “at or above”? Which timestamp, in which time zone?
Cancellation and invalidation. If the event is voided, does each contract refund, resolve, or stay open?
Postponement. What happens if the event moves past the end date?
Exceptional or disputed outcomes. Polymarket’s resolution page, for example, describes a rare “50/50” outcome in which each token redeems for $0.50.² A contract with a partial-payout outcome is not equivalent to one that can only pay $1.00 or $0.
Settlement currency and procedure. In what unit is the payout made, and who or what decides a disputed outcome?
Special settlement terms. Anything else in the contract’s terms that changes what is paid.

The same paper gives two concrete cases. In one, a daily-high-temperature contract on one venue referenced NOAA’s Central Park station while the matching contract on another referenced LaGuardia Airport, which the authors note frequently record different temperatures. In the other, for the 2024 U.S. presidential election, some venues resolved on media network calls and others on the inauguration outcome, which the authors treat as a subset relation rather than an equivalence.¹ Both are the paper’s historical examples; rules on any current contract may differ.

If the terminal payouts or the conditions that trigger them differ, the contracts are close relatives, not the same asset, and part of the price gap may reflect that difference. Nothing in the gap tells you which venue’s wording is worth more. A comparison tool that lists two markets as equivalent has made this judgment for you; treat that as a starting point for the rules comparison, not a substitute for it.

Non-identical contracts do not by themselves rule out arbitrage, and identical contracts are not required for it. What matters is the combined position’s net payoff in every admissible resolution state, after all costs. Three cases need to stay apart:

Identical-event parity. Both contracts share one YES condition, so YES on one and NO on the other pay exactly $1.00 in every state.
Subset–superset coverage. If one contract’s YES condition is verified to contain the other’s, buying YES on the broader contract and NO on the narrower leaves at least one winning leg in every state. Under a simplified model with complete settlement, the pair pays at least $1.00 even though the contracts differ. Section 3.6 of the first preprint describes this construction and assumes each contract’s resolution function is known exactly.¹
Unhedged outcome exposure. If some admissible state leaves the position losing, it carries exposure to how the rules differ. The worked example below is this case: its rules diverge in both directions, no containment holds, and one state pays nothing.

If the minimum net payoff is strictly positive, the position is an arbitrage in the strict sense; if it is only nonnegative everywhere and positive somewhere, it meets the weaker theoretical definition above. Neither is a claim about execution, funding, settlement, or counterparty risk, which stay open even when the payoff condition holds.

Question two: can you trade these quotes, on both sides, together?

A gap can be real, correctly classified, and still not tradable. Start with which side of each book you are looking at. You buy at the ask and sell at the bid, so comparing midpoints or last trades compares numbers nobody may be able to trade. Some venues also display only one side of a binary book. Kalshi’s API documentation says its order book returns only bids and that a YES bid at price X is equivalent to a NO ask at $1.00 − X.⁴ That conversion follows from a complementary binary-orderbook mechanism; it is not something every platform provides, so confirm it in the venue’s own documentation, and confirm that the quantity is available at the converted price, before relying on it.

A second Gebele et al. preprint shows why mechanics matter. It separates payoff-space no-arbitrage, which follows from terminal payoffs, from protocol-executable no-arbitrage, which depends on the position transformations a venue makes available before settlement. It studies Polymarket’s negative-risk markets, where linked binary contracts represent mutually exclusive outcomes and the NegRisk Adapter supports only the NO-to-YES conversion before settlement. In its order-book sample, positive payoff-bound violations concentrated on the unsupported YES side, and the authors describe that pattern as consistent with pre-settlement conversion strengthening enforcement by reducing capital lock-up. It is also a v1 preprint with no journal reference listed, and the authors call the comparisons between market regimes descriptive rather than causal.⁵

That study concerns one mechanism on one venue, not cross-venue gaps. It is not direct evidence that an arbitrary spread between two platforms can be executed, and it does not generalize to every prediction market. The transferable point is narrow: a price relationship that follows from payoffs is tradable only if the mechanics available to you let you enforce it.

For an individual, four practical constraints matter:

Size and depth. A quote may fill only a small amount at the displayed price. Record the size available on both legs at the time you observe them. Prediction market liquidity and executable price covers turning a displayed probability into a price you could fill, and minimum edge after trading costs covers how much margin a result needs before costs consume it.
Timestamps. Two quotes are comparable only if they were observed close together. A gap made of one fresh quote and one stale one is not a gap.
Legging risk. Two trades at two venues are not simultaneous. If one fills and the other does not, or fills partly, you hold a directional position you never intended, so decide in advance what you will do with an unfilled leg.
Separate accounts and capital. Each venue needs its own funded account, and the offsetting positions cannot necessarily be netted against each other, so capital stays committed on both sides until resolution. The first preprint makes the same point about cross-platform positions.¹ Eligibility and access depend on where you are and how you fund an account, which is a personal check no article can do for you.

Question three: do fees, settlement, and capital explain it?

Even with matching rules and tradable quotes, two quotes are not quotes on the same net payoff. Two screening quantities and four cost and risk categories are easy to blend together or count twice, so keep them separate:

Quantity	What it is	How to treat it
Displayed difference	The gap between the quotes on screen, often midpoints or last trades	A screening signal only; you cannot trade a midpoint
Payoff coverage	Whether the combined position covers every admissible resolution state (identical contracts, or verified subset–superset coverage)	Separate coverage from directional risk: a portfolio may eliminate event-direction exposure while its acquisition cost is too high and produces a guaranteed loss; use the minimum state-contingent net payoff to test for a guaranteed non-loss or strictly profitable payoff
1. Executable acquisition cost	What you would pay on both legs at the executable side and depth for your size; its shortfall against the payoff in the covered states is the executable gross difference	Already includes the spread and price impact at that depth, so do not subtract them again
2. Explicit fees and other measurable cash expenses	Fees and any other charge you can price before trading; subtracting them from the gross difference gives the net difference	Cash costs; check whether each venue charges them in cash or through the shares you receive
3. Capital commitment and opportunity cost	Money tied up on each leg until it can be reused, and what it could otherwise earn	Estimate from your own hold-time and cost-of-capital assumptions and keep it on its own line, separate from cash fees
4. Execution, settlement, venue, and counterparty risks	Price movement between observation and execution, incomplete fills and legging exposure, settlement or payout delay and disputes, and venue or counterparty failure	Not priced here, and not zero: no number is assigned, and identical payoffs do not remove them

Two displayed contract prices that sum to less than $1 therefore do not establish an arbitrage, and they are not a reason to call a position risk-free. Quotes cannot show whether the combined payoff covers every admissible state, or whether the costs and risks above leave anything, so price quotes alone cannot establish an arbitrage opportunity. Payoff coverage can remove event-direction exposure without making the position profitable: compare the minimum state-contingent net payoff with zero. A nonnegative minimum with a positive payoff somewhere supports the weaker theoretical arbitrage condition; a strictly positive minimum supports the strict condition. Theoretical payoff arbitrage is separate from execution, settlement, venue, and counterparty risk, which remain even when coverage holds.

Fee treatment differs by venue and market. Polymarket’s documentation states that only takers pay fees, that the fee is C × feeRate × p × (1 − p) with C the number of shares and p the price, that the rate varies by market category, that the amount is calculated in USDC and peaks at 50% probability, and that geopolitical and world-events markets are fee-free.⁶ The same page describes a maker rebate program funded by collected taker fees and redistributed daily, with rates that vary by category. A rebate is a program payout, not a price on your order, so do not net one against your cost in a pre-trade calculation. The page does not say how the fee is deducted from your balance; confirm that at order time, because a fee taken from the shares you receive changes the payoff per dollar rather than adding a separate cash cost.

Kalshi’s help center says it charges a transaction fee on the expected earnings on the contract, that some markets have additional maker fees for resting orders that later execute, and that some markets have fees that differ from others. It points to a separate fee schedule for the calculation.⁷ Actual fees depend on the venue, the contract, the order type, and the schedule in force when you trade, so check the current schedule instead of relying on a remembered rate. Because taker and maker treatment differ, the order type you use changes the net gap. Keep three items on separate lines: the explicit fee, the spread already built into the executable price, and the opportunity cost of funding. They are different things, and folding them together hides which one does the damage. The spread is not deducted a second time after you have priced at the executable side; only price movement after your observation adds something further, and that belongs with the unquantified execution risks.

Settlement mechanics add a second layer, and they are not one event. Trading close, outcome determination, final settlement, redemption, and the point at which capital can be reused can each happen at a different time, and venues document them differently. Kalshi says determination can take from one hour to more than twelve hours after a market closes and usually depends on when the source agency’s data arrives; the page does not state payout timing, so that belongs to the contract terms and the venue’s payout process.³ On Polymarket, an outcome is proposed with a bond and then faces a two-hour challenge period. A dispute triggers a new proposal round, and a second dispute escalates to a token-holder vote. The page gives roughly two hours after proposal when undisputed and four to six days when disputed. It describes trading stopping and winning tokens becoming redeemable for $1.00 each after resolution, but it does not say how soon redemption is available or when the funds can be reused.² These are each platform’s own documented processes, not universal rules, and neither says one venue is better. They mean the capital behind each leg can be tied up for different lengths of time, and a payoff that arrives later is worth less, all else equal.

That also separates two ways of realizing a gap. Holding both contracts to resolution depends on the combined position’s terminal payoff and ties up capital until settlement; Gebele and Matthes note that such offsetting positions are capital-intensive and ill-suited to short-horizon price alignment.¹ Exiting before resolution, if the gap narrows, depends on exit-side depth and on the gap actually converging, and it carries mark-to-market risk in the meantime. A gap that converges early does not show that the payoffs were equivalent.

When the executable gross difference is smaller than explicit fees plus your cost of tying up capital, the price difference still exists, but its net economic value may be zero or negative, and the unquantified risks then sit on top of a figure that is already non-positive.

Question four: if the payoff is not guaranteed, what supports the expected value?

If the complete portfolio has a strictly positive minimum net payoff across all admissible states, information asymmetry is not needed to establish its theoretical payoff advantage; continue checking execution, funding, settlement, and counterparty exposure. If an admissible state can produce a loss, this is a non-guaranteed opportunity: it needs an independently justified probability or expected-value assessment. A difference in information or belief is one possible hypothesis, not an automatic explanation or the only source of edge. The Gebele and Matthes preprint offers a useful benchmark for how much weight to give that hypothesis. The authors argue that in their data the persistent deviations reflect structural frictions, not informational disagreement, and they support that partly with the 2024 U.S. presidential election, where they describe outcome-relevant information as effectively common knowledge after the election-night network calls and prices still did not converge.¹ That is the authors’ interpretation of a large historical sample. It does not show that any particular gap you are looking at is structural rather than informational.

Read the paper’s limits before leaning on it. Its price-deviation analysis covers seven of its ten venues, and it uses standardized fee and spread assumptions and reported mid-quotes for order-book venues, not observed fills. Events were matched with a language-model-assisted pipeline that the authors checked with human annotation on samples. It is also a v1 arXiv preprint with no journal reference listed.¹

Two groups of traders can hold different views, or receive news at different moments, and prices can differ until capital moves between venues. Whether a gap on a given day is that or one of the mismatches above is usually not observable from the quotes alone. For a non-guaranteed opportunity, test the information hypothesis only as one possible explanation: which side would you expect to move, why, and how would you know within a stated period that you were wrong? Evidence that supports it is specific and dated, such as one venue moving first after a public event and the other following. A gap simply closing is not that evidence; gaps close for many reasons, including changes in fees or liquidity and capital that has nothing to do with belief. If you cannot justify the probability or expected value independently, retain the unresolved classification.

A worked example with assumed numbers

All figures below are invented, synchronized illustrative quotes. They are not observed market data, and they are not tied to any real venue or contract.

Quotes. Venue A shows YES at an ask of $0.62. Venue B shows YES at a bid of $0.67. To pair with a YES bought at Venue A, buy NO at Venue B. On a venue whose order book works on the complementary binary basis, a YES bid of $0.67 is equivalent to a NO ask of $1.00 − $0.67 = $0.33, so a buyer of NO pays $0.33 by matching that bid, provided that quantity is available. Assume Venue B works this way. Not every platform does; on some venues a YES bid and a NO ask are the same resting order shown two ways, and on others they are not, so that venue’s own matching mechanics and documentation decide it.⁴

Assumptions. One YES contract at Venue A and one NO contract at Venue B. Each pays $1.00 if it wins. Both legs fill completely at these prices, which are executable at the intended size, so they already include the spread and price impact at that depth. Combined fees are $0.02 per paired position, purely illustrative and not any venue’s schedule. The model adds no other charges, funding cost, or payout adjustments, and it does not price the risks in category 4 of the cost table.

Gross acquisition cost: $0.62 + $0.33 = $0.95.
Total assumed cost: $0.95 + $0.02 = $0.97.
If the contracts resolve consistently, exactly one leg pays $1.00, so the net result is $1.00 − $0.97 = +$0.03. The executable gross difference is $1.00 − $0.95 = $0.05, and the net difference after the assumed fees is $0.03.

The complete payoff matrix, with each contract’s resolution as the state:

Venue A resolves	Venue B resolves	Payout	Net result after $0.97 cost
YES	YES	$1 (A’s YES pays, B’s NO does not)	+$0.03
NO	NO	$1 (B’s NO pays, A’s YES does not)	+$0.03
NO	YES	$0 (both legs lose)	−$0.97
YES	NO	$2 (both legs pay)	+$1.03

The first two rows are consistent resolutions. The last two can occur only when the contracts’ effective resolution conditions differ. Suppose Venue A resolves on the first-published figure and Venue B on the figure as later revised. If the first release falls below the threshold and the revision lands above it, Venue A resolves NO and Venue B resolves YES: both legs lose and the result is −$0.97. If the first release is above and the revision is below, Venue A resolves YES and Venue B resolves NO: both legs pay and the result is +$1.03. Because one admissible state loses the full $0.97, the pair has no positive minimum payoff and fails even the weaker arbitrage definition; it carries unhedged exposure to how the two rulebooks diverge.

Run the four questions.

Same exposure? As described, the two rulebooks resolve on different releases of the figure, so terminal payoff equivalence is not established and both divergent rows are possible.
Tradable quotes? The model assumes complete fills. If the $0.33 is available for only part of your size, the unmatched remainder leaves a single directional leg.
Net costs and settlement? The spread and price impact at this size are already in the $0.95, so they are not deducted again. The assumed $0.02 of fees turns the $0.05 executable gross difference into a $0.03 net difference. The model does not price capital commitment or opportunity cost, price movement before execution, incomplete fills, or legging. The revised figure may be published later than the first release, but when each contract actually settles and releases capital depends on that platform’s and contract’s own processes, so the timing of the two legs cannot be inferred from the order of the data releases.
If the payoff is not guaranteed, what supports expected value? No independent probability or expected-value evidence is recorded, so this remains unresolved rather than an informational diagnosis.

On these assumptions the contracts are not identical, so a semantic explanation is possible but not shown to account for the whole gap; feasibility depends on depth; and the information question stays unresolved.

Where the 3% comes from. Let p be the probability of the adverse combination (A resolves NO, B resolves YES) and q the probability of the favorable one (A YES, B NO). The consistent combinations carry the remaining 1 − p − q and each net +$0.03. The expected net result per paired position is:

(1 − p − q) × $0.03 + p × (−$0.97) + q × $1.03 = $0.03 − p + q

Setting q = 0, the expected result is zero when p = 0.03, or 3%. This rests on simplifying assumptions: a risk-neutral view, complete fills, a four-state matrix with no void or partial-payout outcomes, and no costs beyond the assumed $0.02. It ignores favorable divergence, which would raise the break-even to p = 0.03 + q, and it ignores further fixed costs such as funding cost, which would lower the break-even. Incomplete fills or one-sided execution can change the portfolio and its payoff distribution entirely, so they must be modeled by recalculating expected value rather than treated as a fixed cost that mechanically lowers the original 3% threshold. The probability that real contracts diverge has not been estimated here, and a price gap does not supply it. The 3% is a property of these invented numbers: a break-even condition of a simplified model, not an empirical finding, a trade-entry threshold, or a recommendation.

If the rules were genuinely identical, neither divergent row could occur under the contracts’ terms, and the pair would have a strictly positive minimum net payoff of $0.03 in this model. That is about 3.1% on $0.97 committed, before the cost of tying up that capital and before the unquantified execution, settlement, venue, and counterparty risks. It is still not risk-free.

How do you record a gap so it can be reviewed?

Because the diagnosis is a judgment made before the outcome, it needs a record made before the outcome. For each gap you consider acting on, capture:

both contract identifiers and URLs, and the outcome definition each states;
a saved snapshot of each rules section, with its version or capture time;
a UTC timestamp for every quote;
the side you priced (bid or ask) and the quantity available at that price;
your intended quantity and, afterward, the realized fill on each leg;
the fee assumption for each leg, the source you took it from, and the order type;
your gross and net payoff estimates, with the arithmetic;
your settlement and funding assumptions, including how long capital stays tied up;
your classification (more than one may apply) with its confidence, and the competing explanations you considered;
explicit invalidation conditions: what you would observe that shows the classification was wrong;
what you will do if only one leg fills; and
after settlement, the final outcome of each contract, whether and how the gap closed, and whether your classification held up.

Three cautions keep the review honest. A realized profit or a closed gap does not show the original classification was correct; gaps close for reasons unrelated to the one you assumed, and a profit may come from a rule difference or from luck. A correct forecast that a gap would close is not a correct causal diagnosis. And a single resolved example says nothing about repeatable expectancy; that takes many gaps, with the same sample-size caution that applies to any trading record. Do not claim to know how a gap closed unless you kept intermediate price observations. The second preprint above makes a similar point about its own data: a violation disappearing does not establish that an arbitrage trade caused it.⁵

Where Costante fits

Costante’s relevance here is limited. It is a behavioral execution and review product for discretionary intraday traders, not a prediction-market tool. The parts that transfer are the habits: planning and writing down rules before acting, keeping self-defined guardrails visible, logging the decision, and reviewing what happened against what you intended. A prediction-market trader could apply that discipline to the record above, but would need to keep the rules text, both quotes, fills, and settlement outcomes in their own files or in a journal that supports those fields (the trading journal app guide covers what to look for). Costante does not connect to prediction-market venues, monitor prices, compare contract rules or capture market rules, detect price gaps or arbitrage, execute trades, recommend contracts, or provide prediction-market-specific fields or analytics. Whether a gap is worth acting on remains the trader’s own decision.

Frequently asked questions

Why do the same event’s prices differ across prediction markets?

Often for one or more of four reasons: the contracts are not identical under their own resolution rules, the quoted price cannot be traded at your size on both sides, fees and settlement timing make the net payoffs different, or traders on each venue hold or receive different information, which needs independent evidence before it counts as the explanation. A January 2026 preprint’s semantic-alignment dataset covers more than 100,000 events across ten venues from 2018–2025; its execution-aware price-deviation analysis covers seven of those venues and reported deviations of approximately 2–4% under its stated measurement and execution-cost assumptions. That is a historical finding, not a live measurement or observed individual trading profit.¹

Is a price difference between two prediction markets arbitrage?

Not by itself. Quotes alone cannot establish an arbitrage opportunity, because arbitrage is a property of the whole position, not of a gap. Its net payoff after costs must be nonnegative in every outcome the rules allow and positive in at least one; a stricter test asks for a strictly positive minimum. Two identical contracts, with YES on one and NO on the other, are one way to get there. A verified subset–superset relation, with YES on the broader contract and NO on the narrower, is another, even though the contracts differ. A position that loses in some admissible state, like the worked example above, is exposed to how the rules differ. You also need to trade both sides at your intended size and have something left after fees and other explicit costs. Even then, a cross-platform position carries execution, settlement, funding, venue, and counterparty risk, and it ties up capital until resolution, so it is not unconditionally risk-free.

How can I tell whether two contracts are the same event?

Compare the full resolution rules, not the titles. Check the resolution authority and source, whether a first release or a revision counts, the threshold and cutoff time and time zone, and what happens if the event is postponed, voided, or disputed. Then check what each contract pays across all admissible resolution states. Different wording is not automatically different exposure, but if payouts or the conditions that trigger them differ, treat the contracts as related rather than identical. That does not rule out a hedged combination, but it has to be verified state by state.

Does a price gap mean one venue is wrong?

Not necessarily. The gap may reflect different contract terms, different net costs, or different information, and the quotes alone rarely show which. A gap closing later does not by itself show that the informational explanation was right. A gap is a prompt to diagnose, not a verdict on either venue’s accuracy, and neither price is an objectively correct probability.

Sources

Costante provides educational workflow tools, not financial advice. Trading involves risk.

Gebele, J., & Matthes, F. (2026). Semantic Non-Fungibility and Violations of the Law of One Price in Prediction Markets. arXiv:2601.01706v1, submitted January 5, 2026 — preprint; no journal reference is listed on its arXiv page. Full text read September 19, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹
Polymarket Documentation. Resolution. Accessed September 19, 2026; the page carries no date. ↩ ↩² ↩³ ↩⁴
Kalshi Help Center. Market Outcomes. Page dated March 17, 2026; accessed September 19, 2026. ↩ ↩²
Kalshi API Documentation. Orderbook Responses. Accessed September 19, 2026. ↩ ↩²
Gebele, J., Mutzel, T., & Matthes, F. (2026). Executable Arbitrage and Market Efficiency in Prediction Markets. arXiv:2608.00666v1, submitted August 1, 2026 — preprint; no journal reference is listed on its arXiv page. Full text read September 19, 2026. ↩ ↩²
Polymarket Documentation. Fees. Accessed September 19, 2026; the page carries no date. ↩
Kalshi Help Center. Fees. Page dated April 19, 2026; accessed September 19, 2026. ↩