Saturday, September 1, 2012

Election anti-bounce - economists do it with models


A very brief observation on economists, models, and the US presidential race.

Nate Silver, an economist (who successfully escaped the University of Chicago after completing his BA there), and a person who got hooked on econometrics, has the best electoral race forecasting model I am aware of.

During the 2008 race, he forecast the outcome within 0.1%.

His current model (as of 9/1/12) has President Obama's probability of winning at 72.0%. That is an increase of 2.7 percentage points during this last week of the Republican national convention. The Republican anti-bounce? Maybe they needed more of Dirty Harry, er, Clint Eastwood?

More importantly, this lead has been fairly consistent, perhaps slowly widening, since June. Those who call the race close are paying too much attention to the national horse-race polls. There, Obama's current projected lead is 1.8%, which is fairly close. But at this point has little bearing on the forecast outcome.

For the more wonkish of you out there, check out Nate's blog FiveThirtyEight.com here. Essentially, Nate's model is a (Bayesian?) weighted average of poll results and economic indicators. His results last time were impressive. It's the best political forecasting model I am aware of.

Update 9/3/12. Happy Labor Day! Some have suggested, including an anonymous comment here, that I am being too optimistic in interpreting the Silver models' outputs. No, I am interpreting the models' output. If they are trending toward Obama, that is what the model is saying.

So it is incumbent that if the model goes negative, I also report that output. Silver's Now-cast went negative early today, indicating a 3.1 percentage point decline from the intra-RNC peak to 71.0%. Note that level is 0.3 percentage points lower than the model indicated at the beginning of the RNC, so interpretation is sensitive to choice of starting points. Some, including Silver wearing his hyper-conservative hat, say this is evidence of a Romney RNC bounce.

OTOH, the paired Nov. 6 Forecast model, which has an adjustment for relative convention bounce as well as an adjustment for relative economic indicators, indicates a further widening in the probability of an Obama win, to 74.5%, a model high.

So the convention period was, in terms of these two models, a toss-up, or as Silver says a split-decision. No one knows which model is more accurate at this point in time; I simply point out that Silver built the Nov. 6 Forecast model to specifically adjust for relative convention performance. (Don't go too wobbly, Nate Silver). What is clear right now is that Romney has underperformed recent convention performances.

On to Charlotte to continue the saga.

4 comments:

  1. Ray Fair's model is here http://fairmodel.econ.yale.edu/vote2012/index2.htm You can make your own predictions. It's much more narrow in scope than Nate Silver's model.

    ReplyDelete
  2. The Fair model, as I recall, is mainly economic indicator based. Silver has weighted adjusted poll results and economic indicators based on empirical work to establish the better explanatory variables and their appropriate weights.

    To be, um, fair, the Ray Fair model has been around for quite a while and part of its virtue is that long historical record.

    ReplyDelete
  3. You need to be a little careful with the differences between Nate's fore-cast and now-cast models. The fore-cast models actually take into account modelling convention bounces, economic modelling and a lot of other tweaks. The now-cast is more of a 'if you stopped everything and took a vote right now, who would win?'

    So what we're seeing is that the now-cast is flat, and there seems to be an absence of any convention bounce for Romney. Because the fore-cast has a factor in it which assumes that there should be a post-convention bounce which then decays it is starting to penalize Romney pretty harshly.

    It'll be interesting to see if Obama can pickup a bounce in the actual now-casts -- if those results are flat as well, it could just be evidence that swing voters just aren't being swayed by pageantry this time around and that the post-convention-bounce-model in the fore-casts aren't reflecting reality.

    (I'm certainly with you in *hoping* this all means that romney is going to get absolutely crushed in november, I just think we need to be cautious in calling it early...)

    ReplyDelete
  4. Delighted to see someone paying attention to Nate's blog. The Now-cast is in fact up 1 percentage point during the RNC, so yet another anti-bounce. So both models tell the same directional story.

    As to which model is most appropriate, Nate discusses that at length today.

    My take is that if the current economic indicators and convention weight adjustments in the Nov 6 Forecast model are sufficiently strong to show that size trend (up another 1.1 percentage points since my original post) then we need to pay attention since they are wiping out any actual positive convention bounce by a large margin.

    What happens to these sophisticated models during the DNC - and after the Friday jobs report - will be fascinating.

    ReplyDelete