Election polls vs probability of winning: interpreting the numbers

2 numbers that look alike and are not

A poll gives you a vote-share estimate. If a candidate sits at 52%, that means 52% of the respondents, after the pollster's own weighting, said they would vote for them. A probability of victory is a different quantity: it is the frequency with which your candidate wins across the scenarios whatever model, or order book, generates. An 80% win chance is not the claim that 80% of voters back the candidate. A 52% support reading is not the claim that they win 52% of the time. Best practice in survey research treats these as separate outputs and warns against dressing a hypothetical scenario up as a validated election model.

Because the 2 numbers answer different questions, lining them up side by side is a category error. You can hold a 52% vote-share estimate and an 80% model chance of victory at the same time with no contradiction: one came out of respondents, the other out of assumptions about how those respondents turn into seats or offices. The useful work is keeping them apart, and knowing which one your decision actually depends on.

10 scenarios, 8 wins: the arithmetic behind a model probability

The example below is authored for illustration. It is not a validated election model and should not be treated as one. Suppose you write down 10 possible election scenarios and give each one the same assumed weight. Your candidate wins 8 of them and loses 2. The frequency of winning under this model is 8 / 10 = 0.8, which you can write as 80%. Nothing in those 10 scenarios says the candidate's vote share is 80%. You can assign the same candidate 52% support inside each scenario and still land 8 wins, provided the electoral rules and turnout assumptions produce that split. The 0.8 is an output of the model, not of the poll.

The value of writing this out is seeing where the number comes from. Change the weights, the rules, or how you treat undecided voters, and the win frequency moves even though the vote-share estimate stays at 52%. A model probability is a property of your model, not a fact about the election.

What a poll cannot tell you on its own

A poll is a measurement with uncertainty attached, some of it stated and some of it not. Sampling error shrinks as the sample grows, but non-sampling error, which covers question wording, interview mode, nonresponse and people misreporting their intentions, does not shrink the same way. Disclosure matters for the same reason: a headline number published without fieldwork dates, question wording or weighting method leaves you unable to judge the estimate, and survey best practices push for transparent methods and nonresponse reporting.

Turnout is the next loss point. A poll can measure stated intention well and still miss the result because the people who actually vote are not the people who answered. Undecided respondents have to be allocated somehow: dropped, split or modelled, and each choice moves the number. Shared poll errors add a further problem. If several pollsters use similar methods at the same time, their mistakes correlate, so agreement between polls is weaker evidence than it first appears. Fieldwork dates tell you what was true when the questions were asked; a poll from last month does not describe an electorate that has since changed.

Reading an order book next to a poll

On a prediction market the displayed price is usually the midpoint between the best bid and the best ask. When the spread exceeds 0.10, the last trade is shown instead. That midpoint is not an executable guarantee: a market order eats through several price levels and incurs slippage, and depth matters because a thin book can move a long way on small size. Record the observation date, because yesterday's price is not today's tradable price. The order-book documentation describes the mechanics in full.

So a candidate trading near 0.80 while a poll puts them at 52% support is not a contradiction. The 0.80 is the market's implied price for the event the contract defines, often winning the seat or the office, while 52% is a vote-share estimate. The question worth asking is not which number is right but which event each number points at, and what assumptions bridge them. If the contract defines a different event from the one your poll informs, there is no direct comparison to make at all.

A worked decision: when a quote is worth crossing

A private trader with a written, unvalidated model gives the candidate 0.80 to win the seat. The order book has a best bid of 0.74 and a best ask of 0.86, so the midpoint is 0.80 and the spread is 0.12. Since the spread exceeds 0.10, the venue displays the last trade instead, say 0.79. The trader's question is whether any executable quote disagrees with the model by more than the cost of crossing the spread.

Paying up at 0.86 means buying 0.06 above the model's 0.80. Selling at 0.74 means receiving 0.06 below it. On a symmetric basis, neither side clears the model by enough to pay for the crossing unless the trader's own number is materially tighter than the model's uncertainty allows. The losing case is straightforward: the model is wrong, the poll is stale, the contract defines a different event than the one the model prices, or the book is too thin to fill at the quoted level. In any of those, the apparent edge is an artefact of the comparison, not an opportunity.

Splitting the 2 quantities side by side

The table below separates the quantities by what they measure, what moves them, and what they cannot support. It is descriptive, not a ranking of methods.

Before you compare any poll number to any model or market output, check the 2 rows refer to the same event. If they do not, redefine the event first and start again.

Limits of this walkthrough, and where to go next

No validated election model, real poll sample, margin of confidence or country-specific authorisation is provided here. The 52% figure and the 10 scenarios are authored illustrations, not empirical findings, and the order-book notes describe mechanics rather than any particular current contract. Nothing in this piece predicts an election or promises a gain.

The next concrete step is to write down the event your model or market price refers to, then write down the event your poll data refers to, and compare the 2 definitions directly. If they differ, fix the definition before you convert anything into a probability. If they match, check fieldwork dates, how undecided voters were treated, turnout assumptions, and whether you have more than one genuinely independent source. When those checks fail, the sensible action is usually to do nothing and wait for better data.

Related reading

3 adjacent topics deserve attention before you rely on any single number. Keeping your forecast honest over time is at least as much a matter of the calibration guide as of modelling, since past predictions have to be scored against what actually happened. The broader politics page gathers the electoral and policy context that decides what a contract is really about. And for tracing how a figure travels from raw data to a decision, the AI-workflow guide lays out a structured review process.