With the presidential election less than 50 days away, we have entered the high season of election forecasting. Whether you are a citizen looking for comfort or a journalist hunting for a story, you can choose from a multitude of forecasts ranging from a near-certain Trump victory to a clear Clinton win.
But what if you are interested in accuracy?
Then don’t trust any single forecast. We don’t. We average forecasts within and across six different methods to predict the election outcome.
As of today, our approach known as the PollyVote predicts Clinton to gain 52.8 percent of the national two-party vote (vs. 47.2 percent for Trump). Our latest electoral vote forecast, which follows the same principle of combining forecasts, predicts that Clinton will gain 347 electoral votes compared with 191 for Trump.
That’s consistent with the consensus all year, even if it’s sharply different from recent polls. Let us explain how we get here.
Don’t trust single models
Don’t place your bets on a single forecasting model. Every model has limitations and is subject to bias because of, for example, which information it includes (or ignores) and which method it uses.
It’s also not wise to put too much faith in models that have done well in the past. The most accurate models in one election were often among the least accurate in the next election. This is because every election is different and has its idiosyncrasies, which the models have trouble accounting for. As a result, the performance of models varies over time.
So what should you do?
Forecasting research has produced many guidelines for how to create accurate predictions. Among the two most important ones are:
- Combine forecasts from different methods that use different information.
- Keep it simple: There is no evidence that complexity improves accuracy.
We strictly adhere to two basic principles with the PollyVote project, launched in 2004 to demonstrate advances in forecasting research for predicting election outcomes.
PollyVote’s 2016 forecast
As you can see in the following table, five of PollyVote’s six forecasting components predict a Clinton win.
The only exception is the econometric models component, which includes forecasts from 15 so-called fundamentals models. That method’s combined prediction shows a marginal edge for Trump, expecting him to gain 50.4 percent of the two-party vote.
In contrast, the PollyVote does not expect a close race. As noted above, we find Clinton nearly six points ahead of Trump.
The difference in vote share numbers between the PollyVote and econometric models, and to some extent polls, is mostly due to higher forecasts for Clinton from methods that rely on people’s combined judgment, such as prediction markets and expert surveys, as well as index models, which specifically account for candidate characteristics and campaign issues.
The chart below shows how the vote share forecast has developed over time. Since it became clear that Trump will become the nominee in early February, the forecast has remained remarkably stable, even in times when Trump has been gaining ground in the polls.
So how much should we trust this forecast? Given the PollyVote’s historical track record and stability, it is highly unlikely that Trump wins — that is, that the forecast swings in his favor.
PollyVote’s track record
Since its launch in 2004, the PollyVote has always predicted the correct winner. On average, the PollyVote has more accurately forecast election outcomes than any one of its component methods. Across the last six presidential elections since 1992, the PollyVote would have missed the outcome by only one percentage point on average across the last 100 days before Election Day.
To put this in perspective, polling averages have made errors that were nearly three times larger. In fact, during that same time period, polls predicted elections less accurately than any other method.
The fact that polls perform so poorly might be surprising, since they are so intensively covered in the media. But polls dominate campaign coverage because in the news business, newsworthiness trumps accuracy. Polls allow journalists to cover the horse race, reporting on who is ahead or explaining the latest numbers by campaign gaffes or tactics. Since polls often vary wildly, even if conducted at about the same time, reporters can easily find polls that support almost any narrative.
Of the methods incorporated into PollyVote, the most accurate single approach — although not as accurate as the PollyVote — was citizen forecasts. This method generates forecasts by simply asking people who will win. Unfortunately, most pollsters still don’t regularly include this highly predictive question in their surveys. They should, and journalists should report on it.
Obviously, the PollyVote will not always outperform each of the many forecasts it aggregates. After the event, you will almost certainly find a forecast that turned out to be more accurate.
But be aware. Don’t think you are able to identify the most accurate forecast right now, 50 days before the election. Most people who try to do this fail, because they end up picking a forecast that most suits their biases.
The PollyVote helps you to avoid this by providing a structured and predefined method to combine forecasts.
Andreas Graefe is a research fellow at Columbia University’s Tow Center for Digital Journalism and LMU Munich’s department of communication studies and media research; the Sky Professor at Macromedia Hochschule in Munich; and leader of the PollyVote project. Find him on Twitter @pollyvote.