Election coverage quotes three different kinds of number in the same breath: a poll, a forecast probability and a betting price. They answer three different questions. Below are six sentences of the kind you meet in the news. They are written as examples, not real results. For each one, decide which kind of number it is reporting.
Sentence 1 of 6
The labels matter less than the third question in each answer. It is the one that sorts these numbers fastest, because it asks what would have to happen for anyone to know whether the number was any good.
A poll is graded within weeks. The 2024 US polls missed the two-party margin by 3.3 points on average across 611 contest-level estimates (AAPOR, 2025).
A price is graded as a price, not as a probability. On Kalshi and Polymarket in 2024, the cheapest contracts on each candidate could be bought together for less than a dollar on nearly every day of the campaign, after fees. Two prices for one event that do not sum to a dollar cannot both be that event's probability (Clinton and Huang, 2026, preprint).
A forecast probability is barely graded at all. Presidential elections are rare and one year's errors are correlated across states, so on the best current argument we are decades to millennia away from being able to grade this kind of number (Grimmer, Knox and Westwood, 2024, preprint).