The Limits of Prediction

Why We Cannot Know the Future

Election prediction faces irreducible uncertainty for several reasons. Human behavior is not deterministic: voters change their minds, respond to late-breaking events, and sometimes act unpredictably. Turnout is partly random: rain, transportation problems, and personal emergencies affect who votes. Small polling errors in close races produce incorrect outcome predictions even when the polls are within their normal accuracy range.

The political scientist Philip Tetlock, who has spent decades studying prediction, found that experts are only marginally better than chance at forecasting political events more than a few months out. The world is simply too complex and contingent for precise prediction. A candidate's heart attack, a financial crisis, a pandemic, or a foreign policy crisis can reshape an election overnight in ways no model can anticipate.

Why Uncertainty Is Democratic

Paradoxically, the inability to predict elections perfectly is a feature of democracy, not a bug. If election outcomes were perfectly predictable, campaigns would be pointless, voter turnout would collapse (why vote if you know the outcome?), and the democratic promise that citizens' choices matter would be hollow.

Uncertainty means that campaigns matter, that voter mobilization matters, that a late scandal or a brilliant debate performance can change outcomes. It means that voters are genuinely choosing, not merely following a predetermined script. The appropriate response to polling uncertainty is not frustration but appreciation: we cannot predict elections perfectly because people are free.

A Worked Example: What '29%' Actually Meant

On the morning of the 2016 US presidential election, forecasters published very different probabilities that Hillary Clinton would win:

Trump won. Notice which forecasts look worst in hindsight: not the one that gave Trump 29%, but the ones that gave him 1-2%. A 29% chance is roughly the odds of rolling a 1 or a 2 on a single die — something that happens about one time in three. Treating a 71% favorite as a sure thing is the reader's error, not the model's.

The polls themselves were not wildly off nationally: they showed Clinton about +3, and she won the popular vote by roughly 2.1 points, within normal error. The miss was concentrated in a few state polls — Wisconsin, Michigan, Pennsylvania — that under-sampled non-college voters, and small correlated errors across those states flipped the Electoral College. In 2020 the polls again pointed the right way but overstated Biden's margin. The recurring lesson: read forecasts as probabilities, judge a model by its calibration over MANY elections rather than one, and treat a 20-30% outcome as entirely ordinary.