Pari-Mutuel and Fixed Odds: Two Ways Prices Get Set

The mechanics are simple and the consequences are not.

All stakes on a market go into a single pool. A fixed percentage, the take-out, is removed for the operator and applicable levies. What remains is divided among the tickets that selected the winning outcome, in proportion to their stakes.

That means the payout per unit is a ratio: the remaining pool divided by the amount staked on the winner. If a lot of money backed an outcome, the pool is split many ways and the payout is small. If little did, the payout is large.

Notice what is absent. Nobody set a price. No operator formed a view on probability, quoted a number, or took the other side. The participants priced each other, and the operator's revenue is fixed by the take-out percentage regardless of which outcome occurs.

That structural difference produces almost every practical distinction that follows, and it is why intuitions carried over from fixed-odds markets tend to mislead in a pool.

The take-out is not the same as an overround

The price is provisional until close. Displayed odds during the entry period are projections based on money in the pool at that moment, and they move as more arrives. The number that determines payouts is the one at close. Any analysis using pre-close prices as if they were final is measuring something different from what settled.

You move the price. Placing a stake changes the distribution of the pool, so the act of participating reduces the payout on the outcome selected. The effect is negligible when your size is small relative to the pool and becomes decisive as it grows, which places a hard ceiling on scale that does not exist in the same way with a quoted price.

There is no counterparty view to disagree with. In a fixed-odds market you can ask whether your model disagrees with the operator's price and treat that as a hypothesis. In a pool, the price is an aggregation of participant money, so disagreement means you differ from the crowd rather than from a pricing desk. Those are different claims about different populations.

Late money matters disproportionately. Because the price is determined at close, money arriving near the end can reshape it substantially, and informed participants often act late for exactly that reason.

For a dataset, the practical implication is that you should store the closing pool figures, not only the last observed projection, and that the two should be treated as distinct variables.

With fixed odds an operator posts a price, and whoever accepts it has locked that price regardless of what happens afterwards.

The operator is now exposed. If the outcome occurs they pay at the agreed rate, which means their pricing has to balance two objectives at once: estimating probability accurately, and managing how much they stand to lose on each outcome.

That second objective is what produces behavior that looks strange when you assume prices are pure probability estimates. A price can move because money arrived unevenly rather than because anything was learned, and from outside those are indistinguishable. It is also why margin is embedded in the quoted numbers rather than stated separately, and why removing it requires an assumption about how it was distributed.

The compensating advantage for a participant is certainty. You know your price at the moment of acceptance, your own size does not change it retroactively, and analysis can compare a locked price against a model output cleanly.

So the structures trade off in a specific way. Pools give you a known take-out and an unknown final price. Fixed odds give you a known price and an unknown margin distribution. Neither is simpler, and the analysis code for one genuinely does not work for the other.

If your data covers markets of both types, a few conventions prevent a lot of confusion later.

Label the market structure explicitly on every record, because a payout ratio from a pool and a quoted price from a fixed-odds market look identical in a numeric column and mean different things.

For pools, store the closing figures, the take-out percentage, and the pool totals if available, and keep pre-close projections as separate timestamped observations rather than overwriting them.

For fixed odds, store the source and the timestamp, since prices differ across operators and a price without a time cannot be compared against anything.

Do not compute a devigged probability from a pool payout, since there is no overround to remove and the take-out is a different quantity. Applying a devigging method there produces a number that looks like a probability and is not.

And keep derived fields separate from raw ones throughout, so that when a convention changes you can recompute rather than discovering the original values are gone.

Parlay Ledger works at this layer, ingesting odds across sports and normalizing them across sources so that downstream analysis is comparing like with like. Most of the difficulty in multi-source sports data is exactly this kind of reconciliation rather than the modeling that follows it.