Why do pari-mutuel odds change before a race?

Because the odds are calculated from how much money is in the pool for each runner. As new stakes arrive, each runner's share changes, and the displayed odds are recalculated. The final odds are fixed only when the pool closes, which is why early displayed odds can differ substantially from final payouts.

Does late money predict race outcomes?

It can carry information, and it is commonly studied, but whether it adds value beyond the final odds is an empirical question for each market. Final pari-mutuel odds already incorporate late money and tend to be well calibrated, so any model using flow must show it improves on the pool's own implied probabilities.

How should you evaluate an AI odds analysis claim?

Ask what exactly is being predicted, whether accuracy was measured out of sample, whether probabilities are calibrated, and whether the model beats the market's final implied probabilities under a proper scoring rule such as log loss. A single headline accuracy percentage without those answers does not establish that a model adds information.

What data do you need to analyse pool money flow?

Timestamped snapshots of pool totals and each runner's share, captured at regular intervals until the pool closes, for every pool type offered on the race. Include scratchings and whether pools were combined across jurisdictions. Final odds alone cannot show flow, because they record only where the money ended up.

Reading Money Flow in Pari-Mutuel Pools

In a pari-mutuel pool, displayed odds reflect how money has been distributed so far, so a change in odds is a change in money flow. Models can use the size, timing, and direction of flows, including late money and gaps between separate pools on the same race. The signal is real but noisy, and extraordinary accuracy claims need calibration testing.

Why money flow is the price

In a pari-mutuel pool, all stakes on an outcome type go into one pool, a fixed takeout is removed, and the remainder is shared among winning tickets. The displayed odds for each runner are computed from its share of the pool. There is no operator setting a price; the price is the crowd's money.

That makes pool data unusual among sports data. Every update to the displayed odds is a direct record of how money moved, which is why analysts study not just final odds but how they evolved as the pool filled.

Signals analysts derive from pool data

Several features can be built from pool snapshots. Each is a hypothesis to test, not an established edge.

Late money

Money that arrives close to the start, after most information is public, is often studied separately from early money. The hypothesis is that some well-informed participants act late to reveal as little as possible. Whether that holds in a given market is an empirical question, not an assumption.

Flow size relative to pool

A given amount matters more in a small pool than a large one. Normalising flows by pool size avoids mistaking noise in thin pools for signal.

Cross-pool differences

When the same race is offered in separate pools, such as different bet types or pools in different jurisdictions that are not combined, implied probabilities can differ. Consistent differences between pools, after accounting for takeout, are a feature a model can use.

Consistency across bet types

Win, place, and exotic pools on the same race imply related probabilities. Disagreements among them reveal where participants in different pools see the race differently.

The limits of the signal

The final price is not known when you act. Odds keep moving until the pool closes, and late flows change them after any analysis is done. A model that uses near-final odds as an input is using information that may not be available at the moment a decision would be made.

Thin pools are noisy. Small pools move sharply on single large entries, which may be informed or may not be.

Takeout is large. Pari-mutuel takeout rates are typically substantial, so a flow signal has to overcome a meaningful cost before it carries any practical value.

The crowd is often right. Final pari-mutuel odds tend to be well calibrated overall, which means much of the information in money flow is already in the price.

How to judge accuracy claims

Content about AI odds analysis sometimes advertises detection or prediction accuracy of well above 90 percent. Headline percentages like that should prompt specific questions.

Accuracy at what? Predicting that a heavy favourite finishes near the front is easy; the relevant question is whether probabilities are better than the market's own.
Out of sample? Accuracy measured on the data used to build the model says little.
Calibrated? Do outcomes predicted at a given probability occur at that rate over many races?
Compared with the final odds? A model must beat the pool's own implied probabilities, measured by log loss or a similar proper scoring rule, to add information.

A claim that cannot answer these questions is a marketing number, not a measurement.

Storing pool data for analysis

Flow analysis requires time series, not just final odds: pool totals and per-runner shares captured at regular intervals up to the close, with timestamps and the pool type. Store raw snapshots and derive probabilities from them, so takeout assumptions can be changed later. Record which pools were combined and which were separate, because that determines whether cross-pool comparisons are meaningful. Keep scratched runners and late changes in the record too, since pools are recalculated when a runner is withdrawn and naive analysis will read that recalculation as a flow of money.

This page describes what pool data does and does not show and is not betting advice.

Frequently asked questions

Why do pari-mutuel odds change before a race?: Because the odds are calculated from how much money is in the pool for each runner. As new stakes arrive, each runner's share changes, and the displayed odds are recalculated. The final odds are fixed only when the pool closes, which is why early displayed odds can differ substantially from final payouts.
Does late money predict race outcomes?: It can carry information, and it is commonly studied, but whether it adds value beyond the final odds is an empirical question for each market. Final pari-mutuel odds already incorporate late money and tend to be well calibrated, so any model using flow must show it improves on the pool's own implied probabilities.
How should you evaluate an AI odds analysis claim?: Ask what exactly is being predicted, whether accuracy was measured out of sample, whether probabilities are calibrated, and whether the model beats the market's final implied probabilities under a proper scoring rule such as log loss. A single headline accuracy percentage without those answers does not establish that a model adds information.
What data do you need to analyse pool money flow?: Timestamped snapshots of pool totals and each runner's share, captured at regular intervals until the pool closes, for every pool type offered on the race. Include scratchings and whether pools were combined across jurisdictions. Final odds alone cannot show flow, because they record only where the money ended up.