Election Forecasting: Polls, Models and Prediction Markets

A forecast turns a pile of polls into a single, confident-sounding number: a 70% chance of winning, a projected seat count, a range. For a decade that number came mostly from a few statistical models run by newsrooms and independent analysts. They combined polling with what political scientists call fundamentals, such as the state of the economy and whether an incumbent is running, and they published probabilities that millions of people checked daily.

That arrangement has broken up. Some of the best-known models closed or changed hands. Betting markets, once a curiosity, now handle billions in wagers on elections and present their prices as probabilities. Language models are being asked to forecast too, and some researchers now combine them with human forecasters. At the same time, the public often cannot tell a forecast from a poll from a market price, and treats all three as a prediction of the result. This chapter separates them and asks what each is worth.

Why this matters to you

Forecasts shape how campaigns spend money, how journalists frame a race, and possibly whether some people bother to vote. If you use them, you need to know what goes into them, how they fail, and how to read the probability they give you. The most important failure is not random noise but shared error: when every poll misses in the same direction, averaging more polls makes the forecast look more certain without making it more accurate. If you are weighing a betting price against a poll, you need to know that they measure different things, expectations about who will win against intentions about how to vote. And if you are tempted to publish or trust a machine's forecast, you need a way to tell whether it is forecasting or simply remembering.

What you'll find inside

The chapter opens with how a forecasting model is built, combining polls, fundamentals and prior beliefs into a probability. It then asks whether fundamentals still explain much, and explains correlated error, the reason more polls can tighten a forecast without improving it. Two pieces take on the betting markets: what a price on a prediction market actually represents, and how markets and polls compare when they disagree. You will see how language models perform as forecasters, and why only questions settled after a model's training can show real skill. The chapter covers election-night models that call races from partial returns, and the evidence that publishing a confident probability can change whether people turn out. It closes by leaving the United States behind, to show how forecasting works, and fails, in multi-party systems with proportional representation.

The honest note

The characteristic caveat of this chapter is that confidence is cheap and calibration is hard. Betting markets are reasonably well calibrated on average, yet they can be pushed around at the edges, and the evidence so far does not show that they know something the polls do not. Statistical models are only as good as the polls they start from, and they cannot correct a miss that every poll shares. Language models can look impressive on questions whose answers were already in their training data. None of this makes forecasting useless. It makes it a tool that needs reading with care, and this chapter shows you how.

The nine pieces in this chapter

• P2-01 · The Forecasting Model: Polls plus fundamentals plus priors
• P2-02 · Fundamentals vs Polls: The economy, incumbency and what they still explain
• P2-03 · Correlated Error: When every poll is wrong the same way
• P2-04 · Prediction Markets: A bet is not a poll
• P2-05 · Markets vs Polls, Head to Head: Expectations vs intentions, again
• P2-06 · Machines That Forecast: LLMs and ensembles
• P2-07 · Election-Night Models: Calling races from partial returns
• P2-08 · The Forecast Changes the Election: Publishing a probability moves turnout
• P2-09 · Beyond the US: Forecasting in multi-party and PR systems