Was the 2016 forecast wrong when Trump won?

Not by itself. FiveThirtyEight's final forecast gave Clinton 71% and Trump 29%: an outcome expected about 3 times in 10. One election can't prove a probability wrong. A forecaster is judged over many races, by whether its 70% calls come true about 70% of the time.

Is a 70% chance the same as winning 70% of the vote?

No. The chance is how often a candidate wins across simulated elections; the vote share is how much of the vote they get in one. Race E's Gold is up 2 points in the polls, about 51% to 49%, and wins 69% of the simulations.

What happened to FiveThirtyEight?

ABC News shut FiveThirtyEight down on March 5, 2025, in a round of layoffs, and its site went offline. The Economist's presidential model is open source: its R and Stan code is on GitHub under the MIT license, and Heidemanns, Gelman and Morris describe it in the Harvard Data Science Review (2020).

How many simulations does a forecast need?

Enough that the numbers stop moving. On the stage, Gold's chance of a majority reads 52% after 100 elections, 57% after 1,000 and 56% after 10,000. Each extra digit of precision takes about a hundred times more runs.

Live model · Election forecasts

How Election Forecasts Actually Work

An election forecast runs the election thousands of times on a computer. Each run starts from the polls, adds a random polling miss, and counts who wins. A 70% chance means the candidate won about 7 of every 10 runs. The other 3 are real outcomes: polls miss, and in any one year they tend to miss together.

10,000 simulated elections

10 races. The polls have Gold ahead in 6 and Magenta in 4; Gold needs 6 for a majority. The forecast runs the election 10,000 times, each with polling misses drawn at random: a shared national miss plus each race's own.

StopForecast

Polling errorsGold's poll lead per race, A to JSeed

CH1 Gold wins
CH2 Magenta wins

Watch the Python behind this model run, line by line

How to read the stage

Each bar is a race, labeled with its poll: Gold leads six of ten races, by 1 to 8 points, and needs six for a majority. Each simulated election draws one miss every race shares plus each race's own, drawn as arrows from poll to result. In election 2 every poll misses toward Gold by 5.2 points, and Gold wins all ten. After 10,000 elections, Gold wins a majority in 56%.

What 69% means

Race E has Gold up 2. Gold wins it in 6,916 of 10,000 simulated elections and loses 3,084: a 69% chance. When a 69% favorite loses, that alone doesn't make the forecast wrong. It would be wrong if, over many races, its 69% favorites won far more or far less often than 69% of the time. FiveThirtyEight's final 2016 forecast gave Hillary Clinton 71%.

Why sweeps are common

Polling errors tend to run the same way across states in a given year, AAPOR found after 2016, and The Economist's model treats state errors as correlated. Here the five close races all go to Gold in 13% of elections and all to Magenta in 7%. Draw each race's miss on its own instead, as below: every race keeps about the same odds, but sweeps fall to 4% and 1%.

The same races, each missing on its own

10 races. The polls have Gold ahead in 6 and Magenta in 4; Gold needs 6 for a majority. The forecast runs the election 10,000 times, each with polling misses drawn at random: each race missed on its own.

StopForecast

Polling errorsGold's poll lead per race, A to JSeed

CH1 Gold wins
CH2 Magenta wins

Watch the Python behind this model run, line by line

What the model leaves out

Real forecasts build the polling average themselves, adjusting each poll for its pollster's lean and method, and blend in fundamentals such as the economy. They let similar states miss together more than distant ones, and let opinion drift day to day, so a forecast months out is less sure. Here every race shares one national miss of about 2.5 points plus its own of about 3, sizes chosen for illustration.

Questions

Was the 2016 forecast wrong when Trump won?: Not by itself. FiveThirtyEight's final forecast gave Clinton 71% and Trump 29%: an outcome expected about 3 times in 10. One election can't prove a probability wrong. A forecaster is judged over many races, by whether its 70% calls come true about 70% of the time.
Is a 70% chance the same as winning 70% of the vote?: No. The chance is how often a candidate wins across simulated elections; the vote share is how much of the vote they get in one. Race E's Gold is up 2 points in the polls, about 51% to 49%, and wins 69% of the simulations.
What happened to FiveThirtyEight?: ABC News shut FiveThirtyEight down on March 5, 2025, in a round of layoffs, and its site went offline. The Economist's presidential model is open source: its R and Stan code is on GitHub under the MIT license, and Heidemanns, Gelman and Morris describe it in the Harvard Data Science Review (2020).
How many simulations does a forecast need?: Enough that the numbers stop moving. On the stage, Gold's chance of a majority reads 52% after 100 elections, 57% after 1,000 and 56% after 10,000. Each extra digit of precision takes about a hundred times more runs.