Markets vs. Polls

Polls and prediction markets both attach a number to an uncertain event, and they are often quoted side by side, a poll showing a race near even next to a market pricing one candidate well ahead. The two numbers come from opposite methods. A poll draws a sample of people and records what they say. A market opens a contract and records the price at which participants trade it. Because the methods differ, the numbers measure different things, and a poll and a market can both be accurate while appearing to disagree. This guide works through the comparison one dimension at a time, from how each instrument gathers information to what the accuracy record shows, and it ends on how to decide which to trust for a given question. It assumes some comfort reading a contract price as a probability, the conversion covered in Prices and Probabilities.

The differences line up along a handful of dimensions:

Each row is a section of this guide. Taken together they explain why the two instruments so often print different numbers, and why that difference is usually information rather than error.

How do opinion polls measure public opinion?

A poll is an exercise in sampling. Rather than contact an entire population, a pollster interviews a smaller group chosen so that its composition resembles the whole, then treats the group's answers as a stand-in for everyone's. Random selection is what makes the stand-in valid, because it gives every member of the population a known chance of being included and keeps any one segment from speaking louder than its share (Pew Research Center). Sampling alone rarely lands on a perfectly representative group, so pollsters apply weighting. Some groups answer surveys less often than others, and weighting corrects for that by counting an underrepresented respondent's answer for more and an overrepresented respondent's for less, against population benchmarks such as census figures. That adjustment is standard practice, and it is where much of a pollster's judgment enters.

Every poll carries a margin of sampling error, the range within which the true population value most likely sits. A survey of about a thousand people typically reports a margin near three percentage points at a 95% confidence level, so a candidate measured at 50% is really being placed somewhere around 47% to 53%. The margin describes only the error from surveying a subset instead of everyone. It says nothing about the larger errors that come from reaching the wrong people or modeling turnout incorrectly, which is why the professional guidance is that a poll's real uncertainty can run to roughly twice the reported margin (AAPOR).

Timing and question wording further shape what a poll's number means. A poll is a snapshot, fielded over a few days and accurate as of those days, so it captures opinion at a moment rather than a trajectory. Wording decides what the snapshot depicts. A standard horse-race poll asks respondents whom they intend to vote for, which measures current preference. Asking instead whom they expect to win tends to produce more accurate forecasts, because a respondent reporting an expectation effectively summarizes the leanings of everyone they know rather than reporting a single vote (Rothschild and Wolfers). That distinction matters for the comparison ahead, since the expectation question quietly turns a poll into a small forecasting exercise of the kind the Wisdom of Crowds guide examines. For most of what a poll is built to do, though, the intention reading is the point. A campaign wants to know its current support and a journalist wants to describe where opinion stands, and both are asking about the present state of a population. That is not the quantity a market prices.

How do prediction markets aggregate expectations into a price?

A prediction market starts from a contract rather than a questionnaire. A venue lists an event contract tied to a yes-or-no question. The contract settles at $1 if the event occurs and at $0 if it does not, and participants trade it at any price in between. The price is the aggregate. When a contract trades at 65¢, the market is collectively pricing the event at roughly a 65% probability, since a claim that pays a dollar with a coin-flip's worth of chance is worth about fifty cents and the same logic scales across the range.

No one is sampled and no one is asked to represent anyone. Participation is voluntary and it costs money, so the mechanism selects for people who believe they know something and are willing to back the belief. Conviction enters through position size rather than a checkbox, which weights a confident, well-funded judgment more heavily than an idle opinion. Economists have studied these prices for decades and generally find that they behave as probability forecasts, drawing on information scattered across many participants that no single forecaster holds, and that they have outperformed moderately sophisticated benchmarks (Wolfers and Zitzewitz). The reach extends well beyond elections. Federal Reserve staff evaluating a regulated exchange's macroeconomic contracts found that their forecasts of inflation readings carried significantly smaller errors than the consensus of professional economists surveyed by Bloomberg, and were never significantly worse across the series studied (Diercks, Katz and Wright). A prediction market can price questions a poll cannot practically ask, from a data release to a court ruling, as long as the outcome can be defined and settled.

The result is a number of a different kind from a poll's. A poll reports how a population currently leans. A market reports how likely an outcome is. On a single election the two can sit far apart without either being wrong, because a narrow lead in stated support can still correspond to a lopsided chance of winning.

Why does staking money make a prediction market different from a poll?

The deepest difference between the two instruments lies in what a participant has at stake. Answering a poll is free and anonymous. A respondent can report a preference they hold loosely, one they have never examined, or one they would not act on, and nothing rides on whether it turns out to be right. A market position is neither free nor consequence-free. Opening it costs capital, holding it carries risk, and the account settles against reality when the event resolves. That single feature changes who participates and how their views are counted.

The first effect is selection. Because a position costs money, people with no view and people with no confidence tend not to take one, while people who believe they have insight are drawn to act on it. The second is weighting by conviction. A trader who is nearly certain can commit more capital and move the price further than one who is merely guessing, so the price leans toward the judgments held with the most conviction and capital behind them. A market also pays for disagreement in a way a poll never does. A participant who breaks from the consensus and turns out to be right is paid by those who were wrong, which rewards independent judgment rather than punishing it.

The same mechanism carries a cost. Weighting by capital means weighting by wealth and risk appetite, not by correctness, so a market can be tilted by a small number of large, confident participants who happen to be wrong. A poll gives a fund manager and a retiree the same single voice, and a market does not. Whether that is a strength or a weakness depends on whether the people with capital are also the people with information, which varies from one market to the next. The incentive to be accurate is real, and it is not the same as a guarantee of accuracy.

Do prediction markets update faster than polls?

A poll and a market operate on different clocks. Fielding a poll takes days and costs money, so even heavily polled races are measured in periodic releases with gaps between them, and a poll taken before a debate cannot reflect the debate. A market runs continuously. Its price moves whenever a participant places, cancels, or reprices an order, which can happen within seconds of a headline and long before the next poll enters the field. The mechanics of how orders move a price are covered in Order Books.

The speed is measurable. A transaction-level study of the 2024 US presidential market on one venue found that the time the market took to absorb new information and correct its price fell over the cycle from hours to under a minute (Tsang and Yang). For a fast-moving question, that responsiveness is the market's clearest advantage, since a poll cannot report what it has not yet fielded, and by the time a survey confirms a shift the market has usually priced it already.

Continuous updating cuts both ways. A price that can move on real news can also move on noise, on thin trading, or on a single large order, and a market will sometimes lurch on a rumor that a poll, fielded slowly and deliberately, would never have registered. The two instruments sit at opposite ends of a trade-off between responsiveness and steadiness, and neither end is the right one in every situation. A market gives up stability for speed, and a poll gives up speed for a more deliberate reading.

Are prediction markets more accurate than polls?

The question that gets asked most is also the one with the most tangled answer. Prediction markets are older than scientific polling, and in their first era they forecast well. Large markets trading presidential outcomes operated on Wall Street from 1868 to 1940, at times turning over heavier volume than the stock exchanges, and they called most elections correctly in the decades before the modern poll existed (Rhode and Strumpf). Scientific polling displaced them, and the head-to-head lay mostly dormant until real-money markets returned.

The longest clean comparison comes from the Iowa Electronic Markets, an academic exchange that opened in 1988. Across five presidential elections, its election-eve forecasts landed closer to the outcome than 964 contemporaneous polls about 74% of the time, with a mean absolute error near 1.3 percentage points, and the market's edge over polls was largest months before the vote rather than on its eve (Berg, Nelson and Rietz). A bias-adjusted comparison of the 2008 cycle reached a similar conclusion, finding that debiased market forecasts beat debiased poll-based forecasts early in the cycle and in the most uncertain races (Rothschild). The recurring theme in the academic record is that a market tends to hold its largest advantage well before an election, exactly when polls are least predictive.

The picture is not one-sided, and the 2024 US presidential election became the case study for both readings. Markets favored the eventual winner while several poll aggregators showed a coin flip, which produced a wave of coverage crediting the markets with seeing what polls missed (CoinDesk). Careful post-mortems complicated that story. A study of more than two thousand political markets across four venues found accuracy varied sharply by platform, with a far higher share of PredictIt markets beating chance than Polymarket markets, and documented weak efficiency and persistent price gaps between venues for identical contracts (Clinton and Huang). A separate analysis found that estimates implied by financial options were more accurate and more stable than the prediction markets, which had overstated the winner's probability with high variance during the campaign (Saiegh).

Part of the disagreement dissolves once the two instruments are read as measuring different things. A synthesis of the post-2024 evidence concluded that markets and polls are better understood as complements than rivals, with markets reliable when they are confident and liquid but doing little better than statistical models on the popular vote and the Electoral College, and faring poorly in down-ballot races that attract little trading (Undark). The honest reading of the record is that markets tend to beat polls at forecasting a defined outcome, especially early and in liquid markets, while remaining noisy where trading is thin. The Market Accuracy guide takes up the calibration record in detail.

When do polls and prediction markets fail?

Each instrument fails in its own characteristic way, and recognizing the pattern helps decide how much to discount a given number.

A poll's failures cluster in the part of its error the margin does not cover. When many pollsters make similar methodological choices, their errors point in the same direction, so a systematic miss shows up across an entire field of polls at once rather than in a single outlier, and averaging many polls does not remove an error they all share. Late in a campaign, polls can also cluster more tightly than the underlying uncertainty warrants, a herding tendency in which a pollster hesitates to publish a result far from the consensus (Silver Bulletin). Nonresponse compounds both problems, since the people who decline to answer may differ systematically from those who agree, and no weighting scheme can fully correct for a group it cannot observe. These are the reasons a poll's real uncertainty tends to exceed its stated margin (AAPOR).

A market's failures cluster around thin trading. When little money stands behind a contract, its price is a handful of positions rather than a genuine consensus, and a single large order can move it several points with no new information behind the move. Low-volume political markets in particular tend to compress toward 50%, pricing lopsided outcomes closer to even than they should be, a pattern measured across hundreds of millions of trades on the major venues (Le) and echoed in reporting on lightly traded contracts that swung on small sums (Popular Information). At the extremes of the price range, the long-documented favorite-longshot bias has cheap contracts trading rich relative to how often they resolve yes (Snowberg and Wolfers), though the direction of that bias on newer venues is unsettled, with at least one large study finding no market-level bias on Polymarket (Reichenbach and Walther). Committed capital adds a further wedge on long-dated markets, where money locked in a contract until a distant resolution can hold the price away from the probability, an effect examined in Prices and Probabilities.

The failure modes rhyme with the conditions that make any crowd wise, and diagnosing which one has broken is a skill the Wisdom of Crowds guide sets out.

When should you trust a poll or a prediction market?

The comparison resolves into a practical habit. Match the instrument to the question, because the two measure different quantities and the right choice depends on what you actually want to know:

A few of these rows deserve emphasis. When a question is a well-defined future outcome with active trading and a clear resolution rule, a market is usually the sharper instrument, and its edge is largest well ahead of the event. When the question is about the present state of a population, its margin, or its reasons, a poll measures that directly and a market does not measure it at all. A market cannot tell you that a race is close, only that one side is favored, and a poll cannot tell you who will win, only who currently leads.

Divergence between the two is a signal worth reading rather than a contradiction to resolve. A near-even poll sitting under a lopsided market price is the ordinary case of the two measuring different things, and it often means a close contest has a clear favorite once the structure of the race is priced in. A market that has swung far from a stable poll on no obvious news is a prompt to check whether real information arrived or whether thin trading moved the price. The fullest read holds both numbers at once, alongside the human-judgment forecasts that neither instrument produces, a craft the Superforecasting guide covers.

Both instruments also sit downstream of something earlier. A poll measures opinion after it has formed, and a market prices it after money has moved, while the public conversation that precedes both can be measured on its own through Sentiment Analysis. Read alongside a poll and a price, the tone and volume of that conversation can flag a shift before either instrument has registered it.

A poll and a market are not two answers to one question. They are one answer each to two different questions, and reading them well starts with knowing which question you are asking.

Related guides