One of the best election models predicts a Trump victory. Its creator doesn't believe it.

One of the most respected and accurate forecasting models in political sciences says that Donald Trump will win the 2016 presidential election, and by a fairly comfortable margin at that.

There's just one problem: Its creator doesn't believe his own forecast.

Emory political scientist Alan Abramowitz's "Time for Change" makes predictions by considering how the economy is doing (measured by the GDP growth rate the second quarter of the election year), how popular the incumbent is (measured by his Gallup approval rating at the end of June), and whether the incumbent is running for reelection. It has correctly predicted every presidential election since 1992.

If the tepid first quarter GDP growth rate of 0.8 percent keeps up, as does Gallup's current +9 net approval rating for Obama, then the model suggests a Trump victory, 51.3 percent to Hillary Clinton's 48.7 percent. Even if GDP growth shot up dramatically to 3 percent, the model would still project a Trump win.

Alarmed? Well, Abramowitz isn't — in fact, he thinks Clinton will win easily.

Abramowitz just thinks his model is wrong this year. "The model is based on the assumption that the parties are going to nominate mainstream candidates who will be able to unite the party, and that the outcome will be similar to a generic vote, a generic presidential vote for a generic Democrat versus a generic Republican," he told me. "That's usually a pretty reasonable assumption and produces pretty accurate predictions."

"Usually a pretty reasonable assumption" doesn't mean it's always a reasonable assumption, though, and Abramowitz thinks this is a year when it clearly, obviously, isn't. Since Trump isn't a mainstream candidate, he breaks the models. "It would not shock me if he ends up losing," Abramowitz said, and "if Clinton wins the election by a very comfortable margin."

That was the consistent message I got from talking to experienced election forecasters. They each have their own models and believe in them under normal circumstances. The models that already have estimates — or that can be used for preliminary estimates, like Abramowitz's — are all over the place:

But as the election nears, forecasters are emphasizing that 2016 could be a year when their forecasts fare poorly.

Election models are valuable so long as the election is broadly similar in terms of candidate quality, campaign tactics, and party coalitions to the elections that have happened before.

That is, election models are good at predicting elections that are like past elections. They are bad at predicting elections that are not like past elections.

So here's the question: Is 2016 genuinely different from the elections that came before it? Or do the modelers just want to believe it is?

What we talk about when we talk about election models

By "model," I don't mean poll averaging and weighting of the kind that Nate Silver and FiveThirtyEight, or the team at the Upshot, do. I mean the kind of formal academic forecasting that political scientists and economists have been doing since around the late 1970s, which uses factors like economic variables, presidential approval ratings, and horse race polls to predict general elections months in advance.

Which factors, exactly, vary considerably. There are ones that almost exclusively rely on economic data, which provide a tidy explanation for why particular candidates win but don't have a great predictive track record. More reliable are ones like Abramowitz's that use at least some polling — in Abramowitz's case, presidential approval ratings. Still more straightforward models, like the well-regarded one from Columbia's Robert Erikson and UT Austin's Christopher Wlezien, actually incorporate horse race polling, albeit well in advance of the general election.

In 2012, the journal PS: Political Science & Politics organized a symposium of forecasts for the election. Each was attempting to guess what percentage of the two-party vote Obama would get; ultimately, he got 52 percent to Romney's 48.

The convener of the symposium, SUNY Buffalo professor and Polarized: Making Sense of a Divided America author James Campbell, shared this chart summarizing how close various models got last cycle. This isn't a perfect gauge of their accuracy; all models have error, and it's possible a model that's generally good flubbed it last time. But it's a decent way to get a shortlist of models to consider this time around:

James Campbell

Why the models might not hold in 2016

Few of these forecasters have issued predictions for 2016 yet; most of them require more information than is presently available. And those that have come out, or that can be used for preliminary predictions, are all over the place. There's Abramowitz's model, predicting a Trump win. Wlezien and Erikson passed along an estimate to me, based on current horse race polling and leading economic indicators, that Clinton would get 51.8 percent of the two-party vote, a nearly 4-point win over Trump.

"This is a black swan election," Erikson says. "The economy is middling, the president is around 50 percent approval, so that suggests a close race. But there is variance due to the candidates, and both Clinton and Trump can have big effects."

In other words, these models are built for generic, typical candidates like John McCain or Mitt Romney, who are well within the Republican mainstream on immigration, not someone who thinks a guy with Mexican ancestry who was born in Indiana is unfit to judge him in federal court and who wants to ban all Muslims from entering the United States.

More generally, some forecasters argue that open-seat races are harder to predict than incumbent reelection years. "Presidential approval, quite understandably, is a weaker predictor when the incumbent is not running," Campbell notes. Campbell's research with collaborators Bryan Dettrey and Hongxing Yin confirmed that both presidential approval ratings and economic factors are less influential in open-seat contests. "In open-seat contests," they write, "economic effects on the vote were consistently weaker and never achieved statistical significance."

This makes sense. Presidential approval ratings and the state of the economy seem, intuitively, relevant to whether voters will want to reelect an incumbent. They're not likely to reelect someone they disapprove of, whom they hold responsible for a poor economy, etc. But it's less clear that a president's popularity would rub off on another person his party nominated, or that that person would get credit for a good economy.

"Since they basically ignore information about the candidates themselves, econometric models are based on the theory of retrospective voting (which assumes that elections are a referendum on the incumbent party’s past performance)," political scientist Andreas Graefe of PollyVote says. "This is a useful theory, but if may be less useful if the candidates and their campaigns 'break the rules,' as is the case in 2016."

Models are basically based on 17 elections. That makes predicting tough.

Herb Snitzer/Michael Ochs Archive/Getty Images

Abramowitz says he's agnostic about whether forecasts perform worse in open-seat races; there are just too few data points. That's fair — the brute fact of the matter is that there have only been 17 presidential elections since World War II upon which political scientists can build predictive models. Even if you were to include every presidential race (and you shouldn't, given how little the process that selected, say, John Adams, has in common with modern elections) that'd only be 57.

That's a small sample size, which makes nailing down model specifications tricky. For example, the major factor in Abramowitz's model that gives Trump a strong edge is the advantage he gives to incumbent presidents (and implicit disadvantage he gives to open-seat contenders of the incumbent's party).

This variable is based on comparing 10 races (1948, 1956, 1964, 1972, 1980, 1984, 1992, 1996, 2004, 2008) with the other seven. But that's really not a lot to go on. If you were to read a medical study, say, where the control group had a sample size of 10 and the treatment group had only seven people, you'd be cautious in interpreting the findings.

Or consider what happened in 2012, when Abramowitz adjusted his model to attempt to account for the fact that general elections have been getting closer since 1980. He added a new variable that takes away some of the incumbency advantage and that hurts members of the outgoing incumbent's party in open races when the incumbent is popular (like, say, Hillary Clinton). But that new variable resulted in a worse prediction than his regular model made, so he's back to the original.

Again, this isn't a criticism of Abramowitz specifically, or indeed of any modeler. This stuff is hard. And to his credit, Abramowitz acknowledges that, and tries to identify cases where inferring from limited past patterns can steer you wrong. Candidates have under and overperformed relative to what his model would expect before, he notes. He identifies three cases in particular: the underperformance of George McGovern in 1972, Michael Dukakis in 1988, and Al Gore in 2000.

He interprets the McGovern case as indicating the importance of party unity, and the Dukakis and Gore cases as indicating weak candidates and poorly conducted campaigns. "You can't really build these into the model, because they're pretty subjective," he notes. But just because something's not readily quantifiable doesn't mean it's unimportant.

Some quantifiable factors also give Abramowitz pause about the model. Past nominees have followed normal trends in the partisan preferences of different racial groups. But Donald Trump's vocal anti-immigration rhetoric has led to abnormally low levels of support from Latinos and Asian Americans. Abramowitz has argued that Trump's true level of support from Latinos is in the low 10s, if measured from high-quality surveys with Spanish-language options. That seems almost certainly like a consequence of Trump's specific rhetoric and reputation on racial issues, a candidate effect that normal political science forecasts couldn't capture.

Of course, all this doubt might be overblown. The entire 2016 primary cycle involved pundits and political scientists discounting what the numbers were telling us — Donald Trump was the most popular Republican candidate and set to win by a wide margin — on the basis that Trump was different, that his lead was less real than past candidates' leads because he's too extreme or inexperienced or alienated from the Republican Party. And yet at the end of the day, the simple interpretation was right: The poll lead wasn't a fad — it was a sign that Donald Trump was going to be nominated.

Similarly, the best political science models, like Abramowitz's, have served us well in the past, providing fairly reliable predictions of the victor in November well ahead of time. It's tempting to look at Trump and say that this time is different. But maybe we're once again trying to explain away perfectly valid numbers. Maybe Abramowitz's model is just right, and Trump is set to surprise all of us.

The political science that predicted Trump's rise