How We Handle Uncertainty

Why we don't publish win probabilities

A statement like “72% chance to win” is not a measurement — it is the output of a forecasting model, and it depends on assumptions you cannot see: how large polling errors tend to be, how they correlate across states, how much opinion will move before election day. Different defensible assumptions yield meaningfully different probabilities, and a single-digit-precision number suggests a confidence the underlying evidence doesn't carry. Probabilities for one-off events are also hard to hold accountable: when a “72%” favorite loses, the model can always say that was the other 28%.

Our founding principle is verify-don't-trust: everything we publish should be checkable against the published method and source data. Polling averages and competitiveness ratings meet that bar — every weight and threshold is on the methodology page. A win probability would not, so we don't publish one. We aggregate; we don't predict.

What the simulator is for

Race pages include a turnout simulator. It is an exploration tool, not a forecast. It starts from the demographic patterns in the available polls, then lets you change the assumptions yourself: raise or lower turnout among demographic groups, choose how undecided voters break, or remove a candidate from the field. The output is the arithmetic consequence of your scenario — “if turnout looked like this and undecideds broke like that, the polls imply this result.”

That is the honest version of what forecasting models do internally, with one difference: instead of choosing the assumptions for you and reporting a single authoritative-looking number, the simulator puts the assumptions in your hands and shows you how much the answer depends on them. If small slider movements swing the result, that is the lesson — the race is sensitive to things nobody knows yet.

How to think about a two-point lead

Suppose the average shows one candidate up by two points. Individual polls carry sampling error of a few points on the gap between candidates, and whole polling fields have missed by more than two points in past cycles when they shared a blind spot. A two-point edge is therefore real information — it is better to lead than to trail — but it is comfortably within the range where ordinary error, late movement, or turnout differences decide the outcome.

The right reading is “a close race in which one side holds a slight edge.” The wrong readings are “the leader will win” and equally “it's all noise, polls mean nothing.” Polls measure something real, imprecisely. Both denial and false certainty waste the information.

Where uncertainty shows up on this site

• Ratings widen their uncertainty band when polling is sparse, stale, or scattered, shrinking toward Toss-Up — we under-claim rather than overstate a lean.
• Races rated without usable polling are tagged “fundamentals only” on the ratings page.
• Ungraded pollsters are labeled “unrated” and weighted neutrally — we never invent a grade to fill a gap.
• We never call races ourselves; we report named organizations' calls, and vote totals are labeled unofficial until states certify (see our editorial standards).

Go deeper

For the mechanics of the average itself, see How to Read a Polling Average; for what the labels mean, see What Race Ratings Mean. The full formulas live on the methodology page.