A prediction market's price is only as tested as its age allows

Correction, 29 August 2026. The long-lived figures below are wrong and the method that produced them was wrong.

What is wrong. The long-lived cohort's median is 0.09%, not 0.00%, and its interquartile range is 0.02% to 0.37%, not 0.00% to 0.10%. The sentence saying these markets sit at zero was not a measurement. It was produced by counting unknown as zero.

The n is wrong too. That cohort was reported as 772 contracts. Around 45% of them had no 24-hour figure at all, so their concentration was never measured. The honest statement is a smaller measured n inside a larger cohort, and this piece gave only the larger number.

Why it happened, and why it is worse than a rounding error. Polymarket's API returns the 24-hour turnover field for only about two thirds of contracts, and the script read an absent field as 0. That would be tolerable if the absence were random. It is not. Absence tracks contract age almost perfectly: 1.2% of contracts under a fortnight old are missing the field, rising to 10.4%, then 30.1%, then 43.5% of those over six months old. So the missing data is concentrated precisely in the cohort whose "zero" was the piece's most quotable claim, and the error ran in the direction that flattered the finding. Random would have been a smaller problem.

What survives, recomputed on measured contracts only. The gradient holds and stays monotonic. Re-measured on 29 August across 1,745 open contracts, the four bands read 68.6%, 1.5%, 0.30% and 0.10% by age, against the 34.3%, 1.4%, 0.1% and 0.0% published here. The short-lived cohort is unchanged, because almost none of it was missing data. A limit on that pair of statements, added 30 August. Band one moving from 34.3% to 68.6% and the short-lived cohort being unchanged are not obviously consistent, since both populations are contracts under a fortnight old. They are different cuts, taken on different days, against different populations: band one is every contract under a fortnight, the short-lived cohort is that group narrowed to those resolving within three days, and the re-measurement ran against 1,745 open contracts where the original ran against 1,613. That may account for it entirely. We have not demonstrated that it does, and on a page about a measurement that asserted more than it had established, this correction should not do the same. The interquartile ranges of the two extreme cohorts still do not touch. Three orders of magnitude separating a fortnight-old contract from a nine-month-old one is still three orders of magnitude, and the thesis of this piece is intact.

Those recomputed figures are a fresh reading on a population that has itself moved, 1,745 contracts against the 1,613 measured here on 28 August. They are not a re-derivation of the same set.

The uncomfortable part. This piece has a methods section that names the 166 contracts it excluded and says plainly that the exclusion changed the answer by a factor of ten. It was careful about one exclusion and blind to another in the same measurement, and the one it missed was silent because absence arrived as a zero rather than as an error. A piece about how a population can be wrong, corrected for having a wrong population.

The figures in the body are left as published. The correction is here.

Two contracts quote the same number. One is three days old, the other has been open since December. Both say 3%.

They are not the same object, and the difference is measurable rather than rhetorical. We measured it.

The measurement

We pulled every open contract on Polymarket, 2,000 events, and asked one question of each: what share of everything it has ever traded arrived in the last twenty-four hours. Call it turnover concentration. A contract at 2% has been accumulating opinion steadily. A contract at 80% has done almost all of its trading today.

After exclusions (below), 1,613 contracts remain. Sorted by how long the contract has existed, the answer is a clean gradient:

Four bands, monotonic, and the ends are three orders of magnitude apart.

Taking the sharpest cut, contracts under a fortnight old that resolve within three days (n=79) against contracts over six months old that resolve no sooner than sixty days out (n=772):

The interquartile ranges do not touch. There is a gap between 0.2% and 26.1% with no middle-half contract of either kind in it. No long-lived contract sits above the short-lived median, and no short-lived contract sits below the long-lived median.

That is a strong separation and not a disjoint one, and the distinction matters. The full ranges do overlap: the quietest short-lived contract in the sample reads 0.7%, and the busiest long-lived one reads 39.3%. Anyone telling you these two populations never meet is overstating it. They meet at the edges. They do not meet in the middle.

That 39.3% is worth one more sentence, because it is the tail of the distribution and the tail is where a ratio is least trustworthy. We checked it at trade level: almost all of that day's flow came from a single address. It is a real number and it is not a market. Keep it in the range, since removing an inconvenient maximum is how ranges get flattering, but do not read it as 772 long-lived contracts having a busy end.

What we excluded, and why it changes the number

Three filters, and one of them moved the result by a factor of ten.

Contracts under $20,000 of lifetime turnover were dropped. Below that a ratio is arithmetic on noise: a market with $80 of lifetime volume and $40 today reads 50% and means nothing.

Contracts less than 36 hours old were dropped, because a 24-hour share of a life shorter than two days is close to tautological.

And 166 contracts whose end date has passed but which have not been marked closed were dropped. This is the one worth stating. Expired-but-open contracts sit at zero turnover indefinitely, and they cluster in exactly the bucket a naive cut calls "resolving this week". Leaving them in dragged the median for imminent contracts from 6.6% down to 0.7%, and would have produced the opposite finding. A reader who cannot see that exclusion cannot judge the number, which is why it is here rather than in a footnote.

A limit on the instrument itself, not on the population. Every concentration ratio on this page is volume24hr divided by lifetime volume, both read from Polymarket's public API. That field is intermittent, which the correction of 30 August addressed: when it is absent it arrived as a zero and dragged ratios down. On 1 September we found a third state we had not accounted for. On one contract the field returned a twenty-four-hour figure equal to its lifetime volume while trade records showed the contract had not traded at all in the preceding forty hours, and an hour later the same contract returned no figure at all. The field can be present, plausible, and wrong.

That matters here in a specific and limited way. A wrong value of that shape produces a ratio near 100%, and for a short-lived contract a ratio near 100% is also the correct answer, because a contract four days old genuinely does most of its trading inside any given day. In the short-lived cohort the true reading and the faulty one are the same number, so nothing in the method distinguishes them.

We are not claiming this happened to the contracts measured here. Sampled on 1 September, no open contract older than a month reported that shape: zero of 472 markets and zero of 385 events. The failure is real, transient, and rare, and we have no record of what the field returned on 27 August, so we cannot rule it in or out for this population.

What we can say precisely is what the number rests on. The concentration figures on this page are ratios computed from a venue-reported field, on a cohort where that field cannot be independently checked against the same source. The one measurement that does discriminate is a distinct-wallet count reconstructed from trade records, which appears in the wallet table for the worked examples and was not performed across the full population. The gradient across four orders of magnitude is not plausibly an artefact of a rare transient fault, and the exclusion analysis above is unaffected because it turns on contracts at zero turnover rather than at full. But a reader entitled to judge the number should know that the instrument has a failure mode which is invisible precisely where the finding is strongest.

The worked examples

A short-lived contract, at the middle of its class. Polymarket's 2026 Tour Championship winner market opened on 22 August. Golf's season finale began at East Lake on 27 August. Read on the morning of the 27th, the contract had turned over $62,384 across its entire life and $47,341 of that in the preceding twenty-four hours: 75.9%, on a contract five days old.

The striking part is how unremarkable that is. Against the 79 short-lived contracts in the sample it sits on the 47th percentile, almost exactly the median of 77.9%. It is an ordinary member of an interesting population, and that is the finding. The class is unusual; this contract is not.

Note, 31 August 2026. The Tour Championship contract has resolved, and the figure above can no longer be checked.

The tournament ended on 30 August 2026 and the market settled. The 75.9% is not being revised, because nothing about it has been found to be wrong. It was a reading taken on the morning of 27 August, of a live market, and it was correct on that morning.

What has changed is that it is no longer verifiable. A rolling twenty-four hour turnover figure is a property of a market that is still trading. On a settled contract there is no trailing window to read, so the number cannot be reproduced from the venue, by us or by a reader. It has moved from checkable to uncheckable without passing through wrong, and those are different things: a wrong figure should be corrected, and an uncheckable one should be labelled.

Nothing else on this page is affected, and the reason is the whole argument for how it was built. The band medians, the interquartile ranges, the contract counts and the wallet table are statistics of a population of 1,613 contracts measured on a stated date. A population measured on a date does not decay when one contract in it resolves; it remains exactly what it was, a description of that day. The F1 comparison is unchanged for the same reason, and because that market is still open.

That asymmetry is why the piece rests on the population rather than on this contract. The worked example was chosen to be ordinary, sitting on the 47th percentile of its class, and an ordinary example is illustrative rather than load-bearing. If the argument had been built on the Tour Championship, today it would need rewriting rather than annotating.

A long-lived contract, for contrast. Polymarket's F1 Drivers' Champion market has been open 261 days and has taken $201.7m. In the same twenty-four hours it took $99,050, which is 0.05% of its life. One number, and it is the whole contrast.

The limit on all of it, which nobody states

Here is what this measurement cannot see, and it is not a small thing.

A concentration ratio counts dollars, and dollars carry no record of who moved them. Seven thousand nine hundred and thirty-nine dollars from one wallet and the same sum from four hundred wallets are the same number in the numerator. They mean opposite things, and the ratio cannot tell them apart.

We found this the hard way. A cluster of contracts in this dataset showed 28% to 39% turnover concentration and looked, on the ratio alone, like a dormant market waking up. Checked at trade level, a single address accounted for 88% to 98% of the day's flow in four of the five. It was one participant, not a market. That cluster is not in this piece.

So every contract cited above carries its own wallet check:

The Tour Championship is a market: thirty-four addresses, none dominant. The F1 figure needs its caveat said out loud, and we would rather say it than leave it: 87% of that day's small flow was one address. It does not change the claim being made, which is that $99,050 against $201.7m of lifetime turnover is a rounding error either way. But it is exactly the shape that fooled us once, and a reader is entitled to see it rather than take our word that we checked.

Sampling more widely inside the short-lived group, the breadth is real: an esports contract with 1,657 trades across 644 distinct wallets, another with 1,037 across 533. Largest-wallet share in that same sample ran from 9.9% to 85.9%, which is precisely why no single contract should be quoted from a measurement like this without looking.

What follows from it

Not that old prices are wrong. A long-lived contract can be perfectly well priced, and a market that has traded $201m over eight months has been tested by a great many people.

What follows is narrower and more useful: the amount of testing behind a quote is a function of when you look, and it is not visible in the quote. A 3% on a contract in its first week has been argued over by whoever happened to be there that day. The same 3% on a contract with months of turnover behind it has survived more disagreement. Both are the market's best estimate. Only one of them has been checked.

That is worth holding whenever a price is quoted as evidence, including by us. If you want the mechanics underneath it, our explainers cover how prediction market odds work, what liquidity actually means for a contract and why prediction market crowds tend to be accurate, which is a claim that depends on there being a crowd.

All figures were read from Polymarket's public API between 08:00 and 09:30 UTC on 27 August 2026. Turnover concentration on an open contract is a rolling twenty-four hour figure and moves through the day: the Tour Championship contract read 89% at 08:00 UTC and 75.9% ninety minutes later. The gradient across 1,613 contracts does not move that way, which is the reason the piece is built on the population rather than on any single market.

Frequently asked questions

Does a high turnover concentration mean a market is unreliable?

No. It means the price has been tested less than a quiet ratio would suggest, which is a statement about how much disagreement has been priced in, not about whether the number is right. A short-lived contract concentrating 78% of its turnover into its final days is behaving normally for its class: our sample of 79 such contracts put the median at exactly that. The reason to know it is that two identical quotes can carry very different weights of evidence behind them.

Why does contract age predict this better than how soon the market resolves?

Because time to resolution mixes two different things together. The bucket of contracts resolving within a week contains genuinely new markets created for an imminent event and markets that have been open for months and happen to be finishing. Cut by time to resolution, the median 24-hour share was 6.6% against 0.0% for long-dated contracts, a real difference but with enormous spread inside it. Cut by contract age, the gradient is monotonic across four bands. Age is the cleaner variable.

Can you tell from the price whether a market has been widely traded?

Not from the price, and not reliably from the volume either. Lifetime turnover tells you how many dollars have moved, not how many participants moved them. The only way to distinguish a thousand dollars from one wallet from a thousand dollars from four hundred is to read the trade-level data and count distinct addresses, which is what we did for every contract named in this piece and what the ratio on its own cannot do.