Monday, September 22, 2008

What are the current odds of an Obama victory?

Reader, I know not what to think!

Like a lot of other people, I have for a while had the sense that Nate Silver and friends at FiveThirtyEight.com and the gloriously impersonal markets at Intrade gave me my best reading of the winds of political opinion.

The two sites use radically different methods, but they attempt to answer the same kinds of questions, and for present purposes, the question is precisely the same: what are the chances that Barack Obama will win the election in November? (Naturally, I could use McCain's name throughout this post, but I find it easier to use the numbers for the current favorite.)

FiveThirtyEight is driven by Nate Silver's attempt to aggregate poll numbers and to interpret them based on the history of polling and election results. The site combines state and national polling, tweaking the model along the way and running simulations to estimate the most likely results given current conditions.

Intrade, however, uses the emergent collective wisdom of a market to answer the same question. In theory, the simplicity of the market might do as well or better than Silver at taking all information into account. For example, Silver has been trying to model the distortions of cellphone usage in this year's polls. Because the possibility of cellphone distortions is well known, a real-money market has the potential, at least, to account for their effects without relying on a controlling modeler to estimate them: in many cases, the aggregated crowd can be sharper than the market. (See the work of Robin Hanson for more on the power of markets in politics.)

In spite of these differences, the two ways of answering the questions have produced similar predictions, with small variations due in part to Silver's model deliberately reacting slowly to new developments. Until now.

As I write, FiveThirtyEight's "win percentage" for Obama has rocketed up to 74.4%, although Obama was a slight underdog shortly after the Republican convention. Intrade's bettors also made Obama a slight underdog at that point, and they too think that Obama has regained his status as the frontrunner, but only barely: there, Obama leads 51.5%-47.4%. (For reasons not worth explaining here, FiveThirtyEight's win percentages add up to 100, whereas Intrade's probabilities will be slightly lower.)

I do not see an easy explanation for such a huge divergence. Both sites attempt to look beyond a current snapshot to project the November result, and both attempt to use all the relevant public information (Silver explicitly, Intrade by means of the market).

I welcome correction on this, but I think the plain-language way to sum up the difference is this: the market is making a big bet that McCain will perform better than candidates who had similar poll results at this point in past campaigns.

The difference must rest on a claim about the present or a claim about the future. A claim about the present would involve something like the Bradley effect--the notion that polls will artificially favor Obama because voters don't want to admit they are voting against the Black candidate. Silver has written a series of posts discounting the Bradley effect in this election, however, and even a strong belief in the effect would not explain a divergence as sudden as what we've seen.

If the implicit claim is about the future, the logic might run something like this: we're seeing a close race, and the spectacle of the current Wall Street meltdown has directed the race in a direction favorable to Obama, but this is a bubble. That is, the polls we're seeing today are like post-convention polls, which involve a predictable but fleeting bump. Today's polls are a parenthetical remark, not the story itself.

Whatever explanation you favor, this is certain: the wisdom of the betting crowd sees something behind the numbers that favors McCain--not enough to make McCain the favorite, but enough to keep the race close to a coin flip. The divergence between FiveThirtyEight and Intrade has given us a window into the differences between analytical modeling and market mechanisms.

4 comments:

tjarrett said...

For what it's worth, http://www.electoral-vote.com does something similar to FiveThirtyEight.com and comes up with numbers that are closer to Intrade's: Obama 273, McCain 265. What do you think of their methodology?

Erik_Simpson said...

I don't know how you'd translate electoral vote predictions to probabilities--the EV numbers swing around wildly as Ohio and Colorado flip back and forth (among others), whereas the win probabilities give a clearer big-picture view.

In the Bush-Kerry race, I thought electoral-vote.com was the best polling gizmo I'd ever seen, even though it had some obvious oversimplifications in the way it weighted polls and stuck to an arbitrary one-week timeframe, no matter what polling had been done. As far as I can tell, electoral-vote.com has stuck with the same methods, while Nate Silver has recognized and fixed the most obvious flaws--and added a lot of subtler demographic modeling. That's why I'm a convert to 538, but I could be convinced to pay more attention to other sites. (I know some Real Clear Politics devotees, too.)

Michael said...

Another thing to do if it were ethical: place massive sums of money on your observations of market inefficiencies before they spread to widely read news outlets and force market readjustments.

Next time, Gadget!

Michael said...

Another trick for next time: make a list of all left-leaning transportation, labor and agricultural policy experts who played basketball in college.