Update: Reader Sourav takes a different view: “In this age it is hard for a Black Swan event to happen without any indication….People touting the ‘But Herman Cain was ahead this time 4 years ago’ theory were misreading the data. No one in 2012 had a consistent lead for so long. Trump has been the most retweeted candidate, highest Google search volumes. If the data pundits ignored all this, they did it at their own peril.” Sourav also has some analytics of Trump’s unusually large announcement bump here.
Like I said two weeks ago, Donald Trump is the strongest GOP Presidential front-runner since George W. Bush. Nothing’s happened to change that.
In this week’s news, Sarah Palin endorsed Trump. Then Bob Dole said Trump would be better than Ted Cruz. There are exceptions, but that’s quite a range. Belatedly, The Upshot and Nate Silver are coming around, which is quite a change.
This leads me to a question I have been pondering: what can we learn when quantitative punditry goes off track?
At its best, “data punditry” can help us see past the noise of individual perspectives. Good analysis can help us separate concrete evidence from our biases. But the problem is that we still have to have some prior framework for interpreting data. In this case, Silver and The Upshot weighed in at a time when polls were not very predictive. They did the natural thing: examine other indicators. Now Silver is finally catching up to the polling data, and The Upshot is starting to reflect Trump’s dominance.
What if our biases are so strong that we are unable to recognize a major phenomenon even as it stares us in the face? I am familiar with this experience, having committed a significant error myself in summer 2014.
This year, the major phenomenon is the durability of Trump in a highly divided field. Last year, analysts faced a conflict: (1) polls were starting to point strongly toward Trump as more than a flash in the pan, yet (2) traditional markers of a strong candidate (“The Party Decides”, fund-raising) pointed toward someone else. And mentally committing to “The Party Decides” made it hard to accept concrete evidence of public opinion. Which is how some data pundits ended up lagging behind more traditional journalists. Will Trump get the GOP nomination? I can’t say with certainty, but it seems probable. Now, essays like The Six Stages of Trump Doom sound a lot like “non-data” punditry. Very few analyses from summer are wearing well.
As I said, I am no stranger to this kind of mistake. In August 2014, I didn’t see that year’s wave election coming. About a month later, this reality was forced upon me, and I gradually changed my tune. So yes, predictions can fail. The question is: what can one learn from failures when they happen?
Let me step back and just offer a few thoughts dating back to 2004, when I first got into this game. At that time, a state-poll-based snapshot of Kerry v. Bush gave a helpful picture of the campaign’s ups and downs. Electorally speaking, things went right down to the wire, a tough situation for making predictions. There was some media attention that year, but not that much. The big boom in election data analysis, 2008, was a fairly easy year, since Obama was pretty obviously headed for victory. In such a safe environment for forecasting, there was room to drill into the details of pollster reliability, that kind of thing. For some people, it was nerd heaven. The same was true for 2012.
In 2016, things may not be so favorable for the data nerds.
Most predictions are predicated on the idea that past history can be extrapolated to predict the future. Ray Fair’s Bread and Peace model, “The Party Decides” – these decades-long trends are used to extrapolate into the near future. But these models are only as good as the likelihood that the trends will continue. They are only useful under “normal” conditions – and can’t capture “black swan” events that upend the whole playing field. And once in a while, one of the major political parties undergoes a major realignment, as XKCD’s Randall Munroe has graphically illustrated.
Since the 1980′s, the normal condition has been that the Democratic and Republican parties have been stable, even as they drifted apart from one another. 1994 was a big turning point because it ushered in Gingrich-style polarization. That led to rightward movement, which reached crisis proportions during the Obama administration. Limbaugh begat Gingrich, Gingrich begat Palin, Palin begat Trump. Now here we are, with the most authoritarian one-eighth of the population calling the shots for one of the two major political parties. At the national level, the Republican Party seems to be at a breaking point (though it remains strong at lower levels).
Data punditry isn’t equal to the task of capturing this year’s weird events. For this reason I have been found prognostication by FiveThirtyEight and The Upshot to be an incomplete description. U.S. politics is in incredible ferment, and data punditry isn’t equal to the task of capturing it. We seem to be entering a phase where like other commentators, data pundits are running alongside the crowd, trying to make sense of events. I think the two types of reporting are going to need one another. At least until the general election campaign, when things should get predictable again.