Prediction-market arbitrage is usually presented as a comparison between two venues: find the same event on Kalshi and Polymarket, buy opposite outcomes for less than their combined payout, and lock in the difference.
Hongjian Shi and Daojing Zhai studied a more complex problem. They searched within Polymarket for contradictions between different contracts tied to the same event and therefore constrained by the same logic.
In a football match, for example, an “Over 3.5 goals” contract cannot win if “Over 2.5 goals” loses. Four goals satisfy both conditions, three satisfy only the second, and two satisfy neither. Under comparable settlement rules, Over 3.5 should therefore not be priced above Over 2.5.
Yet each contract trades in its own order book. Polymarket matches buyers and sellers within each book, but it does not continuously enforce consistent prices across every related market. These temporary inconsistencies create combinatorial arbitrage.
In Combinatorial Arbitrage in Prediction Markets, revised on October 25, 2026, the authors simulated such trades using historical full-depth order books. Their idealized strategy generated $2.5 million in guaranteed profit after spreads and fees.
That figure is not verified earnings from a real trader. When the researchers added only one second of execution delay, just 54% of orders filled completely and 22% of the delayed orders ultimately lost money.
How arbitrage can emerge between different contracts
The paper provides a real example from the Cumberland United–Heidelberg United Australia Cup match. The authors compared two contracts:
- Over 3.5 goals;
- Over 2.5 goals.
At one point, the fee-adjusted price available for selling Over 3.5 was $0.9267, while Over 2.5 could be bought for $0.7789.
This violated their logical relationship. The stricter condition was valued more highly than the broader one, even though any result with four or more goals makes both contracts winners.
The trade sold the more expensive Over 3.5 contract and bought the cheaper Over 2.5 contract. For every admissible score, the position guaranteed at least $0.1479 per contract pair after fees, provided both legs executed at those prices.
Another contradiction can arise between mutually exclusive outcomes. An exact score of 0–2 and Over 2.5 cannot both win because a 0–2 score contains only two goals. If the after-cost sale prices of those contracts exceed their maximum possible combined payout, they can also form an arbitrage portfolio.
One event may contain markets for the winner, exact score, spread, match and team totals, both teams to score, and other conditions. Pairwise comparisons are not always enough: a profitable portfolio can span several order books and depend on several logical relationships at once.
How the researchers found the contradictions
The analysis covers Polymarket from June 23 to August 3, 2026. The authors used:
- about 745,000 contracts with metadata and settlement rules;
- 73.22 billion order-book messages;
- 229.9 million on-chain fill events;
- full book depth rather than only the best displayed price.
They recovered relationships in three layers.
Venue-asserted relationships. Polymarket itself groups some mutually exclusive outcomes. A home win, draw, and away win, for example, form a group in which exactly one contract should settle Yes.
Relationships derived from sports data. For contracts on the same match, the researchers mapped market type and line to a common score state. This exposed dependencies between exact scores, spreads, totals, and match winners.
Relationships inferred from contract text. Software rules parsed standardized thresholds and dates, while language models proposed matches between more complicated descriptions. Additional checks were applied before those relationships entered the calculation.
The authors recovered 1.9 million logical relationships, reduced them to roughly 570,000 constraints after cleaning, and covered more than 255,000 contracts. They used mathematical programming to find optimal combinations and capped most simulated portfolios at five legs.
Where the $2.5 million came from
The system detected almost 492,000 arbitrage windows in which the best quotes implied a positive guaranteed profit. A price discrepancy alone, however, does not mean enough volume is available to trade.
After checking full order-book depth, execution was possible in roughly 125,000 windows, or 25.4% of detected cases. For contracts that had already settled, the idealized participant:
- took all displayed depth available in each opportunity;
- included spreads and fees;
- reused replenished liquidity as it appeared in the historical books;
- held positions until settlement;
- reused capital released at settlement.
Under these assumptions, guaranteed profit reached $2.5 million on $113.7 million of cumulative committed capital, or 2.2%. The strategy required no more than $11.1 million at any one time, and the paper calculates a weekly return of 3.5% on this peak capital over the six-week sample.
If the simulation takes only the first execution in each window and ignores later replenished volume, the result falls to $1.1 million.
The $2.5 million therefore cannot be described as money earned by the researchers or a particular bot. It is the output of a historical simulation under specified execution rules. It measures the scale of inconsistent pricing in the dataset, not a statement from a live trading account.
Why one second was enough
In the baseline simulation, orders were submitted against the book state observed when an opportunity was detected. In a real system, time passes while it:
- receives the book update;
- validates the logical conditions;
- optimizes contracts and quantities;
- builds and submits the orders;
- waits for the matching engine to process them.
Another participant may take the required liquidity or change a price during that interval. Multi-leg positions are especially exposed: one leg may fill while the others fill only partly or not at all.
The authors reran the same orders after delays of 100, 250, 500, and 750 milliseconds, and one second.
| Delay | Orders filled in full | Quantity filled | Orders losing at settlement |
|---|---|---|---|
| 100 ms | 71.6% | 92.4% | 14.2% |
| 250 ms | 65.1% | 89.5% | 17.3% |
| 500 ms | 59.9% | 86.9% | 19.6% |
| 750 ms | 57.0% | 85.3% | 20.9% |
| 1 second | 54.0% | 84.0% | 22.0% |
With a one-second delay, the filled orders retained 79% of their baseline profit at settlement. The aggregate result remained positive, but individual portfolios were no longer risk-free.
The key is uneven execution. If every leg is reduced in the same proportion, the portfolio’s logic remains intact. But when one contract fills completely, another only partly, and a third not at all, the original guarantee disappears. The trader is left with directional exposure they never intended to take.
More complex relationships lasted longer
The study found another pattern: the harder a logical relationship was to discover, the longer its pricing contradiction tended to survive.
Among windows lasting at least 0.3 seconds, median duration was:
- 2.9 seconds for relationships explicitly identified by Polymarket;
- 6.3 seconds for relationships reconstructed from sports fields;
- about 31 seconds for relationships inferred from contract text.
This does not mean complex opportunities are necessarily more profitable or easier to execute. Their settlement rules, deadlines, and admissible states must be matched correctly. An error in the relationship itself is more dangerous than an ordinary price move: a portfolio may appear guaranteed only because the algorithm omitted a possible outcome.
The result nevertheless illustrates the value of market-structure analysis. Many programs can monitor the obvious sum of home-win, draw, and away-win prices. Fewer participants will detect a contradiction spread across contracts with different descriptions, allowing it to persist longer.
How this differs from Kalshi–Polymarket arbitrage
Our earlier article, “7,518 NFL market checks—and not one Kalshi–Polymarket arbitrage”, examined identical or opposite outcomes traded on two different venues.
There, the main challenge was obtaining executable prices at the same time, confirming that settlement rules matched, and accounting for both platforms’ fees.
This paper solves a different problem. Every simulated trade occurs within Polymarket, but the system must reconstruct the full set of possible outcomes rather than match identical contracts. Over 3.5 is not the same contract as Over 2.5, but the former logically implies the latter. An exact score is simultaneously related to the match result, totals, and other markets.
This arbitrage is harder to identify, but it reveals a defining feature of prediction-market architecture: many separate order books describe the same reality without being required to maintain coherent prices at every moment.
Conclusion
The research shows that a mathematically guaranteed payoff and a practically risk-free trade are not the same thing.
Once every leg has executed at the calculated prices, logical relationships can genuinely guarantee a payout under every admissible outcome. Until then, execution risk remains: prices move, depth disappears, and the portfolio’s legs fill unevenly.
The paper’s $2.5 million is an important estimate of inconsistent pricing on Polymarket, but it comes from an idealized replay of historical books. The delay experiment marks the boundary of that estimate: one second did not erase the model’s aggregate profit, yet it turned 22% of delayed orders into losing positions at settlement.
A real arbitrage system therefore needs more than correct mathematics. It must process data quickly, coordinate leg sizes, control partial fills, and decide in advance what to do when the guaranteed portfolio cannot be assembled in full.
Source: Hongjian Shi, Daojing Zhai — Combinatorial Arbitrage in Prediction Markets. All results refer to the authors’ historical simulation and are not verified returns from a live trading account.



