Next year – suddenly out of nowhere – we’ll have a slew of new political prediction models all of them claiming to have been able to predict the Rise of Donald Trump, the Sanders Movement and even the recent exit of Britain from the European Union. Of course, all of them will be after the fact which has always been an easier thing to accomplish – and in the next cycle we’ll throw them all out and go back to the standard models that assume that most people will vote the way that they’re expected to vote under normal circumstances.
We’re talking about some major models breaking down in the face of unexpected changes using standard assumptions, here! For example, the “Time for Change” model created by Emory’s Alan Abramowitz (which just to be clear, seems to be well-known in the public sphere, but holds no real academic credentials judging by number of citations alone – though Alan is well cited as a researcher) by his own account predicts that Donald Trump will win the US presidential election by a comfortable margin, but that he does not believe his own model is relevant – and that Hillary Clinton will actually win by a comfortable margin. This is actually quite a big deal in the modeling world – not because a modeler predicts that his model will fail in predicting certain events – that’s just having a solid understanding of the limitations of your own model – no, but because there was no real alternative model or explanation provided. Why will it fail!? [silence]
There are a lot of explanations on why some well-known models will fail this year – many of which will no doubt surface after today’s #Brexit vote in Great Britain – which basically has dumbfounded every political expert on two continents. So dumbfounded that even some of the leading figures of the Brexit movement had no idea (even to the last moment) that they would win the referendum until it actually happened.
The truth is however, that there can only be one real explanation for why these models are failing – they’re just flat out wrong – and yes we can all go back to the common aphorism known in the modeling community, “all models are wrong but some are useful” and try to explain away the fact that these predictive political models aren’t just wrong but, right now, aren’t worth the paper (or computer) that they’re written on – some of which had millions invested in them not including the time taken to build them. Really, this is beyond the simple ‘useful but wrong’ modeling. It’s not just that these models are wrong for this political cycle or that they retain some small usefulness in every cycle even though they’re wrong most of the time – it’s that it’s very possible that they’re fundamentally flawed to begin with!
The question then becomes: Well, how do we explain that they work sometimes and have predicted past election results?
Luck! Yes Luck! Let me say it one more time – LUCK! Or as we scientists call it – statistics. We all know that it’s super easy to regress the next data point on a linear curve (which we know to go ahead and assume as an infinitesimal point on a curve – calculus 101) when you have enough data points lined up further up the curve – and so many of these models rely almost exclusively on regression analysis and have no intention or make no active effort in understanding, modeling, or predicting individual level behavior – and until they do they will never predict these types of important scenarios.
You see, from the perspective of fundamental statistics Trump and Brexit – even Bernie – are outliers of the standard curve, and traditional statistics aren’t designed for that kind of prediction. However, if we attempt to predict the individual level behavior of actual voters, their cognition, their perceptions and simulate them in silico we might just have a better chance of creating models that predict the so-called outliers, because to a simulation an outlier is not an outlier but rather it is emergent behavior built right into the model or simulation by coding so-called normal behaviors directly into individual agents.
That’s right! I’m saying that the use of agent-based modeling could really help in building political prediction models, but unfortunately, at the moment there just aren’t enough experts working on making that happen. Why? Well a number of technical issues exist – like having to create some pretty accurate synthetic populations, needing the computational resources to run the model a single time, and of course, the assumptions that go into the model itself are quite elaborate.
Still, it is possible that the ideal model is right around the corner because in my opinion, if nothing else, what all these sudden and unpredictable changes do is motivate our community of modelers in creating more perfect models!