How accurate is Polymarket? The honest answer is that it depends on how — and when — you measure. By one widely cited analysis, the platform's prices are roughly 90% accurate a month before a market resolves and north of 95% in the final hours. By a competing academic study of the 2024 US election, Polymarket finished last among the major prediction venues, behind both Kalshi and PredictIt. Both findings can be true at once, because they are measuring different things. This piece walks through what the data actually says, who said it, and where the caveats live.

FrenFlow is an aggregator, not a market operator. We have no stake in Polymarket looking good or bad, so what follows is the unflattering parts included alongside the flattering ones.

The Case That Polymarket Is Highly Accurate

The strongest case for Polymarket's accuracy comes from calibration data: when its prices say an outcome is 70% likely, that outcome tends to happen about 70% of the time. The most referenced source is a public Dune dashboard built by New York–based data scientist Alex McCullough, which Polymarket itself surfaces on its accuracy page.

According to that analysis, as of mid-2026 Polymarket's prices were correct over 90% of the time a full month before a market resolved, climbing to roughly 95–96% accuracy in the four hours before resolution. The headline statistical figure is a Brier score reported around 0.0843 across resolved markets — and a Brier score for a 12-hours-ahead prediction of 0.0581 across roughly 90,000 predictions, per McCullough's data cited by economist Scott Lincicome. For reference, McCullough's own rule of thumb is that a Brier score below 0.125 is good and below 0.1 is great, so those numbers are genuinely strong.

A Brier score measures the gap between forecasted probabilities and actual outcomes; lower is better, zero is perfect. The mechanism behind the numbers is straightforward: a price on Polymarket is a probability. A contract trading at 63 cents implies the market thinks the event is 63% likely. (If that framing is new, see our guide to how to read Polymarket odds.) When thousands of people are putting real money behind those estimates, the aggregate tends to be well-calibrated — that is the "wisdom of crowds" argument, and on Polymarket's own resolved-market data it holds up.

One important qualifier: these figures are calibration metrics over a broad universe of resolved markets, many of them short-dated and near-certain by the time they close. That is exactly why the more adversarial studies, which isolate genuinely uncertain political races, paint a different picture.

The Studies That Push Back

The most damaging finding for Polymarket comes from a December 2025 study by Vanderbilt University researchers Joshua Clinton and TzuFeng Huang, reported by DL News. They analyzed roughly 2,500 political prediction markets carrying about $2.5 billion in volume across the Iowa Electronic Markets, Kalshi, PredictIt, and Polymarket during the final five weeks of the 2024 US presidential campaign.

Their accuracy ranking, by the share of markets that called outcomes better than chance: PredictIt at 93%, Kalshi at 78%, and Polymarket at 67%. In other words, the platform with the largest handle — Polymarket recorded a historic $2.4 billion in 2024 election volume — scored lowest on this particular measure. We include that because it is true and inconvenient, not because we expect you to like it. The same researchers also flagged broader efficiency problems across all the venues: prices for identical contracts diverged between exchanges, daily price changes were weakly or negatively correlated, and arbitrage gaps widened in the final two weeks before the vote.

Kalshi disputes the methodology. Spokesperson Jack Such told DL News the study "completely misunderstands prediction markets, and shouldn't be taken seriously," arguing that accuracy should be judged by calibration — whether 20%-priced contracts resolve "yes" 20% of the time — and that on that basis the venues look far better. (The fact that Kalshi outscored Polymarket here is part of a wider comparison we cover in Polymarket vs Kalshi.)

There is a second, narrower rebuttal worth citing. Good Judgment — the forecasting outfit built on Philip Tetlock's "superforecaster" research — ran Polymarket head-to-head against its own forecasters across 25 recent central-bank rate decisions. Per Good Judgment's published results, its Superforecasters posted a better daily Brier score on 76% of days, averaging 0.135 against Polymarket's 0.159 — making the market roughly 18% worse than expert forecasters on that specific set of questions. None of this means Polymarket is "bad." It means a single accuracy number for a whole platform is close to meaningless. The honest takeaway is that the platform is well-calibrated in aggregate, weaker on contested political races measured by hit-rate, and beatable by trained forecasters on some structured questions.

Prediction Markets vs Polls

Do prediction markets predict better than polls? The historical record leans yes, but with real exceptions and a recurring caveat about timing. This is the angle most people actually care about, because it is the one Polymarket's marketing leans on hardest.

The pro-market evidence is older but solid. A frequently cited 2008 analysis of the Iowa Electronic Markets found the markets beat traditional polls in roughly 74% of comparisons across five US presidential elections. Separate examinations of the 2008 and 2012 races found market forecasts outperformed the polls in those cycles as well. The intuition is that a market price absorbs polling data plus everything else — money flows, news, expert opinion — and updates continuously rather than in discrete survey waves.

The counter-evidence is just as real. A 2012 historical assessment (Rhode and Strumpf, published in the International Journal of Forecasting) concluded that after scientific polling arrived in the 1930s, election betting markets performed no better than polls as predictors — markets were far superior in the pre-poll era (1880–1932) but lost their edge once good surveys existed. Other work has found that last-minute market prices add essentially nothing once you control for final-week trial-heat polls. And as recent coverage in Undark argued in 2026, prediction markets are unlikely to replace polling outright, because the two answer different questions.

The fair conclusion: prediction markets are a strong forecasting signal and historically competitive with or better than polls, but the "markets always beat polls" claim is an overstatement. They are best treated as a complement that aggregates information, not a replacement oracle.

Why Accuracy Depends on When You Look

A prediction market's price converges toward the truth as the event approaches — so accuracy measured near resolution will almost always look far better than accuracy measured early. This timing effect explains most of the apparent contradiction between the optimistic and pessimistic studies.

McCullough's dashboard makes the pattern explicit: Polymarket's reported accuracy climbs from around 90% a month out to roughly 95–96% in the final four hours. That is not unique to Polymarket; it is how all markets behave. Early on, genuine uncertainty is priced in, so a market sitting at 55/45 will be "wrong" about 45% of the time by construction — that is the market correctly expressing a coin-flip, not a failure. Closer to resolution, the unknowns collapse and prices snap toward 0 or 100.

Liquidity compounds the effect. On the same dashboard, markets with more than $1 million in total volume post dramatically better Brier scores — on the order of 0.025 twelve hours before resolution — than thin markets, because deep order books make manipulation expensive and reward sharp traders for correcting mispricings. The practical implication for anyone reading the odds: a liquid market hours before it closes is a far more reliable signal than a thin market weeks out. Quoting a single accuracy figure without stating the timing or the volume is the most common way these numbers get abused.

Where Prediction Markets Get It Wrong

Prediction markets fail most predictably in three places: thin markets, tail events, and situations where a large trader's flow gets mistaken for information. Each is a structural weakness, not a one-off.

Thin markets are the biggest one. With little volume, a single sizeable order can move the price far from any reasonable probability, and there is not enough opposing capital to correct it quickly. That is also the manipulation surface. A 2026 agent-based modeling study on arXiv examined how strategic actors can distort prediction-market prices, and the concern is most acute precisely where liquidity is lowest.

The 2024 election produced the canonical real-world case. A French trader known as "Théo" deployed an estimated $30 million across four accounts on Trump-victory positions, with total exposure reported above $85 million, and ultimately netted roughly $79 million by Chainalysis's estimate. Critics argued the size of the bets was inflating Trump's odds beyond what polls implied. Polymarket investigated, brought in outside experts, and concluded there was no manipulation — that Théo was taking a directional position based on a genuine read of the race (he commissioned his own "neighbor" polls and bet the shy-Trump-voter thesis). He happened to be right. But the episode shows how hard it is, in real time, to distinguish a smart whale from a manipulator — and that ambiguity is itself a weakness.

Tail events and long-shot bias round out the list. Research going back to 2008 documented that markets tend to overprice underdogs in winner-take-all contests, and genuinely surprising, low-probability outcomes are exactly where crowds — and prices — are least reliable.

Why On-Chain Verifiability Is the Real Edge

Here is the part where the accuracy debate actually resolves into something useful: a market's accuracy is arguable, but a trader's track record on Polymarket is auditable and cannot be faked. That distinction is the real informational edge, and it is where an aggregator's view matters more than the operator's.

Every trade on Polymarket settles on Polygon, on a public blockchain. There is no self-reported screenshot, no survivorship-biased "track record," no marketing spin — just an address, its positions, and its realized profit and loss, visible to anyone. You can argue forever about whether a market was "67% accurate" or "90% accurate" depending on which study and which window you pick. You cannot argue about whether a specific wallet turned $10,000 into $40,000 over the past year, because the chain says so.

That is what FrenFlow surfaces: verified all-time on-chain PnL for the traders worth watching, reconstructed from settlement data rather than claims. Instead of betting on whether the crowd is calibrated this quarter, you can identify the specific people who have been right with their own money, and — if you choose — copy a verified trader automatically. For a broader landscape view, our roundup of the best prediction markets compared covers how the venues stack up, and our guide to Polymarket copy trading goes deeper on the mechanics.

Frequently Asked Questions

How accurate is Polymarket?

It depends on how and when you measure. A Dune analysis by data scientist Alex McCullough, surfaced on Polymarket's own accuracy page, reports prices that are over 90% accurate a month before resolution and roughly 95–96% in the final hours, with a Brier score around 0.0843 across resolved markets. But a December 2025 Vanderbilt University study of the 2024 election found Polymarket called only 67% of political markets better than chance — behind Kalshi (78%) and PredictIt (93%). Both can be true: calibration over many markets is strong, while hit-rate on contested races was weaker.

Is Polymarket more accurate than polls?

Often, but not always. Historical research, including a 2008 study of the Iowa Electronic Markets, found markets beat polls in about 74% of comparisons across five presidential elections, and markets outperformed polls in 2008 and 2012. However, a 2012 historical assessment concluded markets performed no better than polls once scientific polling existed, and other work found last-minute prices add little beyond final-week polls. Markets are best seen as a complement to polling, not a replacement.

Was Polymarket right about the 2024 election?

On the headline outcome, yes — its odds favored Donald Trump's victory, against a polling consensus that called the race a toss-up. But the December 2025 Vanderbilt study (Clinton and Huang) still ranked Polymarket last among major venues at 67% across roughly 2,500 political markets, and the cycle featured a French whale, "Théo," who bet over $85 million on Trump and profited an estimated $79 million. Polymarket investigated and found no manipulation. Being right on the top line and scoring lower on aggregate accuracy are not contradictory.

What is a Brier score?

A Brier score measures the accuracy of probabilistic forecasts by averaging the squared difference between predicted probabilities and actual outcomes. It runs from 0 (perfect) to 1 (worst), and lower is better. As a rough guide cited by analyst Alex McCullough, below 0.125 is good and below 0.1 is great. Polymarket's reported scores around 0.0843 (resolved markets) and 0.0581 (12 hours out) fall in that strong range — though Good Judgment's Superforecasters posted a better daily Brier score on 76% of days in one head-to-head test.

Can Polymarket odds be manipulated?

In thin, low-volume markets, yes — a single large order can move a price far from a fair probability, and a 2026 agent-based study on arXiv modeled exactly how strategic actors can distort prices. In deep, high-volume markets the cost of manipulation rises sharply, which is why liquid markets post far better calibration. The 2024 "Théo" episode showed how hard it is to distinguish a large directional bettor from a manipulator in real time, though Polymarket's review found no manipulation in that case.

The Bottom Line

So, how accurate is Polymarket? Strong on calibration, genuinely impressive near resolution in liquid markets, and historically competitive with polls — but not the infallible oracle the marketing implies, and beatable on contested races and structured forecasts by other venues and trained experts. The most accurate way to use it is to read liquid markets close to resolution, treat thin and early markets with skepticism, and remember that no single accuracy number captures the whole platform.

And when you want a signal you cannot argue with, skip the platform-level debate and look at the chain. A market's accuracy is contestable. A trader's verified all-time on-chain PnL is not.