Prediction market arbitrage is mostly a mirage

The pitch is irresistible, and the arithmetic behind it is real. A set of mutually exclusive outcomes — one of which must happen — should cost exactly $1 in total, because $1 is all that will ever be paid out. Find a set priced at 96c and you have bought a guaranteed dollar for 96 cents. No opinion required, no forecast, no risk.

So we built a scanner for it, pointed it at live Polymarket data, and found plenty. Then we checked whether any of it could actually be traded. None of it could.

The measurement

The scan looked at every multi-outcome set on the venue and summed the prices. Two of the results, verbatim:

Sets that survived contact with the real price: zero. Not a small number — zero. Every apparent opportunity was an artefact of which price we were reading.

Why the midpoint lies

An order book has two prices. The bid is the most anyone is currently willing to pay; the ask is the least anyone will accept. The number displayed on almost every screen — including ours, at the time — is the midpoint between them, and it is a price at which nobody is offering to sell you anything.

That distinction is invisible on a liquid two-sided market where the spread is a cent. It is fatal on a set with eight legs, because you pay the spread once per leg. The Arsenal set had thin quotes on several outcomes; the midpoint said 96c, and buying it actually cost $1.21.

Even a real one can be too thin to survive

Suppose you sum the asks properly and find a set that genuinely pays. Here is one we found doing exactly that: eight legs, costing $0.987 at the ask to win $1. A 1.3% return, guaranteed, on paper.

Now price in reality. Prices move in one-cent ticks, and a quoted ask is only good for the size resting at that level. If one of those eight legs fills a single tick worse than quoted, the set costs $0.997 — a 0.3% return. If two do, it loses money.

You are placing eight orders in sequence against a book other people are also trading. Getting all eight at the quoted price is the *best* case, not the expected one — and the expected case is a loss.

The rule we ended up with

A flat threshold — "show anything above 1%" — treats a two-leg set and an eight-leg set as equally safe. They are not: every additional leg is another independent chance to be filled worse than quoted. So the requirement scales with the number of legs.

A set must be able to absorb one tick of slippage for every four legs, and still pay.

Two legs at $0.987 clears it. Eight legs at $0.987 does not, and no longer appears anywhere in the product. The result is that genuine free money is shown far less often, which is the correct outcome, because it exists far less often than a midpoint scan suggests.

What to do with this

• Sum the asks, never the mids. If your tool does not tell you which it is using, assume mids and assume the number is wrong.
• Count the legs. Required margin should scale with them, because execution risk does.
• Treat any two-leg set summing to 99c as noise — that is the spread, not an opportunity.
• Be suspicious of a screener that finds arbitrage every day. Real ones are rare, small, and gone in seconds.

That is why Sharply prices every entry at the real ask, and why a claimed edge that does not survive the spread is reported as no edge rather than a small one. Most of the time the honest answer is that there is nothing here — and a tool that never says so is not measuring anything.