The Summer Slump: Why August Polls Are the Worst Predictor of November

Every August, pollsters release surveys showing the current state of the midterm race. Every August, analysts treat these surveys as meaningful. Every November, the August polls are revealed to have been approximately as predictive as a coin flip weighted slightly toward the correct answer. The summer slump is real, the polls are real, and their predictive value is minimal.

WHAT HAPPENED

• August 2025 generic ballot average: D+4.8 (RCP), D+5.1 (FiveThirtyEight aggregate)
• Presidential approval (August average): 42.1% approve, 55.3% disapprove
• Historical August generic ballot accuracy for November outcome: correlation 0.52
• Historical August presidential approval accuracy for midterm seat change: correlation 0.58
• Average generic ballot shift from August to November in midterm years: 2.7 points (in either direction)

THE NUMBERS BEHIND THE NUMBER

August polls are unreliable for three reasons, all of which are well-understood and none of which prevent their publication:

1. Likely voter screens: Most August polls use registered-voter screens because likely-voter models are unreliable more than 90 days from the election. The RV-to-LV adjustment in midterms averages 2.3 points toward Republicans. August polls using RV screens systematically overstate Democratic performance.

2. Campaign effects: In August, many competitive races have not yet begun advertising. The typical competitive House race spends 60–70% of its budget in September and October. The information environment that will determine the election does not yet exist.

3. Events: The average number of "significant political events" between August and November in midterm years is 4.2 (government shutdowns, economic data releases, foreign policy crises, October surprises). Each event shifts the polling environment by 0.5–2 points.

POLLERBULL SIGNAL

• What moves odds: August data enters our model at a weight of 0.15 — present but subordinate to fundamentals. Our model's accuracy improves dramatically between August (68% correct seat predictions) and October (89% correct). The September and October data is where the signal lives.
• What would falsify this: If August polls prove to be within 1 point of the November outcome, the campaign period (September–October) had less effect than usual, suggesting a "baked-in" electorate that was already decided. This would require recalibrating the model's campaign-effect assumptions.