Get more FiveThirtyEight
Throughout this evening, we’ll be updating election night forecasts as states are called for presidential and senate candidates. To clear up any misinterpretations, we’re not trying to project states based on partial returns. So if (for example) Trump is leading Missouri by 5 percentage points with 40 percent of precincts reporting, that won’t matter to the model.
Instead, our election night model is much simpler than that. It relies upon only these three things:
1. Our pre-election forecasts.
2. States that are “called” by our partners at ABC News.
3. The amount of time that has passed since the polls closed in a state, if it hasn’t been called yet.
To repeat, these forecasts do not use votes counted so far. They also do not use exit polls. They do not look at margin of victory. The only input is a single designation for every state: “D” (called for the Democrat), “R” (called for the Republican”) or blank (not called yet), based on calls made by the ABC News Decision Desk. We can also call states for independent candidates or project that the Georgia or Louisiana Senate races will go to a runoff.
Having a state called for you helps in two ways in the model.
• It gives you electoral votes.
• It helps you in our forecast for the other states. For example, if Wisconsin has been called for Clinton, the model can infer that she’s more likely to win Minnesota. And it really helps candidates if they win in an upset, since that’s a sign that they may be beating their polls everywhere.
Our election models have been running tens of thousands of simulations each day in order to account for the relationship between different states in the Electoral College. The most important takeaway is that state outcomes are correlated: If Trump (unexpectedly) wins Virginia, for example, he’s also extremely likely to win North Carolina. So each simulation creates a plausible map based on a state’s region and demographics. In one simulation, perhaps, Clinton might outperform among Hispanics, therefore winning Florida and Arizona despite losing Ohio and Iowa.
These simulations are useful for our election night forecast also. Once we know the results in some states, we can make better inferences about the results in the remaining ones. So as states are called, we update the forecasts accordingly based on a series of regression analyses that relate every state to every other one.
The model also considers how long it’s been without a call in a state. If it originally expected Clinton to win New Jersey by 10 percentage points, for instance, but New Jersey still hasn’t been called five hours after polls have closed there, it will discount that lead significantly, assuming the state is closer than pre-election polls had it.
As a final word of caution, I wouldn’t read too much into the first few called states unless there are substantial upsets (Clinton winning Indiana, for example). But once swing states start to be called and the model has more data to work off, its projections will be more meaningful.