A Guide to Presidential Prediction Models
(image)
I don’t enjoy politics.
It’s a shame too. Politics has some of the best data and forecasting around. And great gossip and news. Every four years I take a look at the election and wish I could get as excited as my friends.
Through Lumesis, I had to follow elections and politics more closely, since it impacted our business. Now at CrowdTwist, I don’t interact with politics as much, but it still rears its ugly mug occasionally.
Predictive models, on the other hand, I can get into. And boy, are there a lot of them! Most of my friends love politics but have no interest in digging through political papers. So for all of you: here’s a cheat sheet for who is calling which candidate, why, whether it measures popular or electoral votes, and how seriously I take them.
Model: 13 Keys to the White House
Creator: Allan Lichtman
Who is this guy?: Professor atThe American University in Washington, D.C.
Learn about it: http://www.informs.org/ORMS-Today/Public-Articles/June-Volume-38-Number-3/Election-2012-The-13-keys-to-the-White-House
Type: Popular
Prediction: Strong Obama
Mudslinging: Nate Silver at 538 has a teardown.
How does it work?: Lichtman asks 13 ‘True/False’ questions (like ‘is the incumbent-party candidate charismatic?’). It’s best-of-13: if 7 are true, incumbent wins. If not, he loses. That’s it.
Big claim: it has correctly predicted the popular vote outcome of every U. S. presidential election since 1984’
Best thing about it: uh, it’s hilarious that anyone takes this seriously?
Worst thing: Lichtman counted Obama as ‘charasmatic’ in ‘08, but not this year. If that sounds like a reasonable basis for decision-making to you, I’ve got a whole bunch of facebook stock I’d love to sell you…
My thoughts: This gets press? Let’s move on.
Model: Time for Change Model
Creator: Alan Abramowitz
Who is this guy?: Professor at Emory
Learn about it: here
Type: Popular
Prediction: Slight Obama
Mudslinging: In 2008, Abramowirz donated the Obama and didn’t disclose it, according to Wikipedia.
How does it work?: The model combines Obama’s approval rating, recent GDP growth, and the advantages of incumbency
Big claim: The model has predicted share of the popular vote withing 1.1% for the last 4 elections
Best thing about it: Abramowitz has changed his model to account for the increased polarization in the US
Worst thing: Abramowitz claims that the results around the polarization adjustment ‘out-of-sample’, which seems like an aggressive claim
My thoughts: For what it’s worth, Abramowitz’s model seems hacked together. It’s no ‘keys’, but I consider its predictive power to be suspect.
Model: Presidential and Congressional Vote-Share Equations
Creator: Ray Fair
Who is this guy?: Fellow, International Center for Finance at Yale
Learn about it: http://fairmodel.econ.yale.edu/vote2012/index2.htm
Type: Popular
Prediction: Even/Slight Romney Edge
Mudslinging: Again from 538, the catty reality star of election modeling
How does it work?: The model is a best fit regression that takes into account a few non economic variables (are we at war? Is a sitting president running? etc) and adds in GDP growth, inflation, and the number of quarters of strong economic growth during the administration. Strong, consistent growth is good for the Prez.
Big claim: Fair claims his model has been right 21 or 22 times in the last 24 elections (that’s starting in 1916 for the English majors) depending on how you classify Bush-Gore. It missed badly with Bush-Clinton in 1992 and he counts 2000 as a miss too.
Best thing about it: Fair’s model is clear and rigorous: the math makes some intuitive sense even. And he’s got a handy calculator so you can see just how much you can effect the outcome by changing economic statistics here. Just like the Bureau of Labor Statistics does! (I kid, I kid…)
Worst thing: Fair calculates his model’s standard error at 3%, so the model doesn’t really tell us anything. Any it hasn’t worked well in recent elections. But other than not making a decision and then often being wrong when it does, it’s a pretty sweet model.
My thoughts: I am something of a Prof Fair fanboy- I really admire his work around aging in Sports and Chess. His election model is classic econo-physics: elegant and reasonable, but ultimately limited.
Model: Bread and Peace
Creator: Douglas Hibbs
Who is this guy?: Retired Academic, consultant
Learn about it: http://www.douglas-hibbs.com/
Type: Popular
Prediction: Strong Romney
Mudslinging: You can fit this in with the other 538 criticisms
How does it work?: Hibbs’s model is one of the simplest: how fast are incomes growing, and how many people are dying at war? With no war deaths, income growing at about 1% makes for a 50-50 election
Big claim: Hibbs doesn’t make huge claims: I have seen others say he claims 90% of variability in his model but I haven’t see the quote. He started releasing in 1992 and has had some misses but has been directionally right each time.
Best thing about it: It’s simple, elegant and it seems to work pretty well. Doesn’t get much better than that.
Worst thing: As 538 points out, Hibbs’s model shows extreme signs of ‘over-fitting’ or describing our specific recent experience instead of the underlying mechanisms. I think Silver is being a little unfair, but the point is well taken.
My thoughts: Hibbs’s model is really interesting and performs pretty well. It amazes me how well the data from 1952 to 1996 fits his model. Hibbs’s model likes Romney’s odd the most, giving him about a 90% chance of winning
Model: History and Primary
Creators: Helmut Norpoth and Michael Bednarczuk
Who are these guy?s: Professors at Stony Brook and UW-Milwaukee
Learn about it: http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=8700583&fulltextType=BT&fileId=S1049096512000868
Type: Popular
Prediction: Strong Obama
Mudslinging: No one really pays attention to this model, apparently
How does it work?: Norpoth’s model counts incumbency and how the two candidates did in the New Hampshire primary: larger percentages of the vote in the Granite State in January correlate with a win in November.
Big claim: That New Hampshire matters.
Best thing about it: It’s a great excuse to listen to Granite State of Mind
Worst thing: It suggests that the leader of the free world is chosen by a bunch of burly dudes fishing Lake Winnipesaukee and their wives
My thoughts: History and Primary believes in Obama the most, giving him an 88% chance of a win. Also, oh my lord I hope this logic is wrong.
Model: State Level Economic Indicators
Creators: Kenneth Bickers and Michael Berry
Who are these guys?: Professors in the University of Colorado System
Learn about it: http://www.colorado.edu/news/releases/2012/10/04/updated-election-forecasting-model-still-points-romney-win-university
Type: Electoral
Prediction: Strong Romney
Mudslinging: Probably just me, below
How does it work?: Like the Fair model, this model depends on economic inputs (income and unemployment), but it instead uses them at a state-by-state level to call each state. The authors look at previous year vote splits to set a baseline for each state.
Big claim: would have predicted how 90% of states have gone (and 100% of presidential winners) since 1980
Best thing about it: The model seems to make sense in theory, and provides specific scenarios. It has Obama losing Ohio and Florida badly- Wisconsin becomes the new swing state due to Paul Ryan’s involvement.
Worst thing: This model is conservative, assuming that states lean the way they have in the past. Given that, Bickers and Berry are going to miss any surprises or changes in a state’s electorate.
My thoughts: I’m pretty sure this model is massively overfit. Using 13 variables to predict 8 elections guarantees that there are no remaining degrees of freedom- I would be shocked if they hadn’t predicted previous elections! The authors would argue that there is more uncertainty due to 50 states, but based on correlations between states I would never consider them to be independent. I would expect this model to perform slightly better than the Fair model since it’s electoral, but very similar
Model: The Jobs Model & The Proxy Model
Creators: Michael S. Lewis-Beck and Charles Tien
Who are these guys?: Professors at the University of Iowa and Hunter College
Learn about it: http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=8700619&fulltextType=BT&fileId=S1049096512000984
Type: Popular
Prediction: Romney and Obama. Wait, what?
Mudslinging: uh, themselves?
How does it work?: Not to be outdone, Lewis-Beck and Tien have two separate models that they use, both regression models. The Jobs Model looks at presidential popularity, economic growth, jobs growth, and incumbency, and is what they are better known for. The Proxy Model depends on business sentiment. Right now, the Proxy Model leans towards Obama, while the Jobs Model shows Romney ahead.
Big claim: At the point where you feel compelled to build a second model, you probably aren’t claiming crazy results with the first one.
Best thing about it: Both are simple models that look to mix sentiment and economic data- the business sentiment report is really a poll in economic data’s clothing.
Worst thing: The Jobs Model nailed 2004, but was way off in 2000 and 2008. Maybe it works best when a sitting president is running.
My thoughts: Given the poor results in the last few elections, I’m skeptical. The reasoning behind the model makes sense, but I suspect it doesn’t work better than just forecasting using presidential popularity alone.
Model: Leading Economic Indicators and the Polls
Creators: Robert Erikson and Christopher Wlezien
Who are these guys?: Professors at Columbia and Temple
Learn about it: http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=8700589&fulltextType=BT&fileId=S1049096512000881
Type: Popular
Prediction: Strong Obama
Mudslinging: nobody, really. Even Nate Silver thinks these guys do good work.
How does it work?: It uses the index of leading economic indicators (if you aren’t familiar with economics, this is a known index that comes out quarterly), along with the most recent polling data.
Big claim: Using data from 1 quarter before the election, the model has gone 13/15 since the 50’s. Also they think that debates don’t make any difference in the election.
Best thing about it: You can bring it up to political scientists without getting laughed at. These guys have great reputations.
Worst thing: Trying to pronounce Wlezien. Also, since they use quarterly data, they lag behind everyone else on updating: they are still using economic data announced pre-convention.
My thoughts: The LEI+P Model likes Obama’s chances, giving him an 80% shot. But this model has over-estimated the incumbent party vote 4 elections in a row (thought within margin of error). That’s probably just coin-flip noise, but it does make me a little hesitant.