How parimutuel prediction markets work

If you have used a prediction market before, it was probably one where outcomes trade as shares. YES costs 63 cents, NO costs 37 cents, you buy the one you believe, and if you are right your share redeems for a dollar. That is an order book market, and the price is a real price: someone sold you that share at that number because they disagreed with you.

A parimutuel market has none of that. There is no price, no share, no counterparty, and nothing to buy. There are two pots. You put money in one of them. That is the entire user interface, and almost everything surprising about parimutuel markets follows from how little is going on.

Two pots and one division

A binary parimutuel market holds two totals: everything staked on YES, and everything staked on NO. When the market settles, one side is declared the winner and the whole pot — both sides — is redistributed to the people on the winning side, in proportion to what each of them put in.

The contract behind this app expresses it in four lines:

total      = total_yes + total_no
fee        = total × fee_bps / 10_000
payout_pool = total − fee
payout(u)  = stake(u, winning_side) × payout_pool / winning_pool

Read that last line slowly, because it is the whole product. Your payout is your share of the winning side, applied to the entire pot. Not to your side. To everything, including all the money staked on the side that lost.

A worked example. Suppose a market closes with 6 ETH on YES and 2 ETH on NO, and a 2% protocol fee.

You staked 1 ETH on YES. Your share of the winning side is 1 / 6. Your payout is 1 × 7.84 / 6 = 1.3067 ETH. You risked 1 and came back with about 1.31.

Now run it the other way: NO wins. The winning pool is 2 ETH, and someone who staked 1 ETH on NO owns half of it. Their payout is 1 × 7.84 / 2 = 3.92 ETH. Same market, same fee, same stake — a very different result, because they were on the smaller side.

Odds are pool share, and pool share is not a price

That asymmetry is what people mean when they say a parimutuel market has odds. The "odds" on YES in the example above are just payout_pool / total_yes, or about 1.31x. The odds on NO are about 3.92x. Neither is a price anyone quoted. Both are ratios that fall out of how the money happens to be distributed at that instant.

Which leads to the single most important property of this market type, and the one that costs people money when they miss it:

Your payout is determined by the pool split at settlement, not by the split at the moment you staked.

There is no such thing as locking in odds here. If you stake on YES when YES holds 1 ETH and NO holds 5, the board is showing you a very attractive ratio. But if forty more people then agree with you and YES climbs to 20 ETH, your slice of the winning side has shrunk from most of it to a twentieth of it. The pot grew, your share shrank, and the number you originally saw was never a commitment — it was a snapshot of a distribution that was always going to keep moving until the deadline.

This cuts both ways, and it is the reason the mechanism is interesting rather than merely simple. Being early on a side that stays unpopular is precisely what pays. Being early on a side that becomes the consensus pays much less than the board suggested when you joined, because you have been diluted by everyone who arrived after you and agreed.

If you are used to fixed-odds betting, the mental adjustment is: you are not taking a price, you are buying a fraction of a pot whose size and composition are both still being decided.

Nobody can move against you, and nobody has to take the other side

The flip side of having no counterparty is that a parimutuel market cannot fail to fill. In an order book, a market on an obscure question simply has no liquidity — you want to bet, and nobody will quote you. Here, staking is always possible, because you are not asking anyone to agree to anything. You are adding to a pot.

It also means there is nothing to front-run in the usual sense. There is no order to jump ahead of and no price to push around, because there is no price. What another participant can do is change the distribution, which changes what your winning stake is eventually worth. That is a real effect, but it is not adversarial in the way a book is — it is arithmetic, visible to everyone, in public, the whole time.

The cases where nobody wins anything, and everyone gets their money back

Two situations make the payout formula meaningless, and both are handled the same way: full refund, zero fee.

A pool is empty. If nobody staked NO, then everybody who staked YES was right and there is nothing to win — there was never a counterparty. The contract refunds every stake in full and takes no fee. This matters more than it sounds: charging a percentage on a bet that had no opposite side would simply be taking user money for having provided nothing, and a market that does this is one to avoid.

The market was voided. If the outcome cannot be established — including the case where the settlement window closes with nobody having settled it at all — the market voids and every stake is returned in full, again with no fee.

In both cases the fee is not reduced, it is zero. A protocol fee is compensation for running a functioning market; where no market functioned, there is nothing to compensate.

The small print that is actually worth reading

A few implementation properties determine whether a market is safe to put money into. These are the ones to look for in any parimutuel contract, and what this one does:

• Division truncates, so payouts round down. The sum of everyone's payouts is always slightly less than the payout pool. That remainder — less than one base unit per claimant — stays in the contract permanently. It is never paid twice and never swept. That rounding direction is what keeps the contract solvent: rounding up would eventually leave the last claimant unable to withdraw.
• Pausing can block staking. It must never block claiming. If an emergency pause can stop you withdrawing money you have already won, the pause is a custody risk wearing a safety label.
• Payments are pull, not push. You call claim and the contract sends you your money. It never iterates a list of winners, which is what makes a market with many participants settle at all.
• No claim expiry, and no sweep of unclaimed funds. Ever. Your winnings do not become the protocol's after ninety days.
• ERC20 stakes are credited by measured balance delta, not by the amount requested. A token that takes a fee on transfer, or rebases, therefore cannot desync the contract's internal accounting from the balance it actually holds.
• The logic is immutable. The markets here are clones with fixed implementation code — not upgradeable proxies. Nobody can change the payout rule under money that is already staked.

Where the outcome comes from

None of the above says anything about how a market learns who won, and that is genuinely the hardest part of any prediction market. A perfectly implemented payout formula fed by a settlement process someone can influence is not a prediction market, it is a donation. That question is big enough to deserve its own treatment, and the settlement page covers where an outcome is read from, what makes a market void, and who is able to act at each stage.

For the mechanics in more depth than this article goes — what a market is allowed to ask, the stake limits, and how claiming works step by step — see how it works. To look at markets that are actually open, start here. And read the risk disclosure first: everything above describes how the mechanism behaves when it works, which is a different subject from whether you should be staking money on it.