Tracking Polls

Rolling Averages and Daily Updates

Tracking polls survey a new batch of respondents each night and report rolling averages, typically over three or four nights. If 300 people are surveyed each night, the three-night rolling average represents 900 respondents. Each day, the oldest night's data drops off and the newest night's data is added, creating a continuous stream of results.

The value of tracking polls is not their absolute accuracy on any given day but their ability to detect movement. If the tracking average shifts from 48-46 to 45-49 over a week, something has happened, even if the absolute numbers are not perfectly accurate. Campaigns use tracking polls as an early warning system: did last night's debate performance move numbers? Did an opposition attack land? Is the race tightening?

Noise vs. Signal

The challenge with tracking polls is separating real movement from random noise. With small daily samples, normal sampling variation can produce 2-3 point swings that mean nothing. A candidate who appears to surge 3 points on Monday and drop 3 points on Tuesday may have experienced no real change at all.

Experienced analysts look for sustained directional movement over several days rather than single-day changes. If a tracking poll shows the same candidate gaining ground for five consecutive days, that is likely real movement. A single dramatic shift should be treated with skepticism until confirmed by the next few days' data or by other polls.

Worked Example: How Much Movement Is Real?

Suppose a tracking poll interviews 300 voters per night and reports a three-night rolling average (n = 900). Watch how the window moves and why the margin of error matters.

First, the margin of error. For a proportion near 48% with n = 900, the standard error is sqrt(0.48 x 0.52 / 900) = sqrt(0.0002773) = 0.0167, so the 95% margin of error is 1.96 x 0.0167 = 3.3 points on A's share alone. The margin on the gap between two candidates is wider still (roughly 6-7 points), because both shares carry error.

Now read the table. Monday's A-minus-B gap is +2; Friday's is -4, a 6-point swing. Is it real? Any single day's shift of 1-2 points sits comfortably inside noise. But the direction is consistent across all five days, moving B's way every single reporting cycle. That monotonic drift, not the size of any one jump, is the signal. Compare that to a hypothetical where A read 48, 45, 49, 46, 48: same range, no direction, pure noise. The discipline is simple: trust the trend line, distrust the daily wiggle, and demand several consecutive days before you act.