Diving into intrade

It is rather difficult, in general, to even tell who is winning the presidential election, or what a candidate's chance of winning the election is. A national poll is borderline-irrelevant, since that's not how the race is decided. 51 state polls is a better idea, and electoral-vote does a nice job of showing you that information, as it evolves.

But you get rather infrequent updates of any particular state. And polling in general is subject to many problems. The pollster applies typically some adjustment to account for party affiliation, but that is a bit of an art and in the end something whose properness is always argued. And not least of the problems is that people will not tell the truth about who they plan to vote for.

A neat solution to all those problems is the intrade markets. There, you will find a market for each electoral college race. You can buy a contract that, for example, a Democrat will win the Pennsylvania state election for US President. Right now, the price of that is $69. If Obama ends up winning Pennsylvania, you collect $100, otherwise you collect $0. So it is fair to say that invested participants of this market believe Obama has about a 69% chance of winning Pennsylvania, as of today.

One has to put their money where their mouth is (which tends to make people come up with their true opinion) and to project into the future everything they imagine might happen between now and the election. Furthermore, the markets are all relatively liquid and in any given market you get price movement every two or three days on average. The participants in the markets obviously have the polls, but also the news, and their own ideas about how the campaigns will evolve, and ultimately their own ideas about who will show up to vote and for whom will they vote. Given enough market participants taking those questions seriously, you ought to have a really good idea of what a candidate's chances are.

I am working on using this to extract a truer and more real-time idea of who is really winning the presidential election, by how much, and how exactly (in terms of the electoral college). It is a lot of work though, to do correctly.

One interesting piece of information I thought I would share right now, is the answer to the question how correlated are the 51 electoral college races? Even for such a simple-seeming question, there are a lot of potential pitfalls I haven't considered really. But what I did is measure the correlation of returns, ln (price_today/price_yesterday), using the last 90 days of trading. I studied the markets only on the Democrat winning, just to cut down on the work involved. The price of the corresponding Republican contract is usually very close to $100-price of the Democratic contract.

The surprising answer is that the markets appear very loosely correlated. I would have thought that national events drive all markets together to large degree, but that does not appear to be the case. The mean absolute correlation in the 51x51 matrix is 0.15, and the mean correlation is .06.

90% of the correlations are between -.3 and .3
75% of the correlations are between -.2 and .2
45% of the correlations are between -.1 and .1

Electoralmap.net does a good job of bringing the idea of intrade together with the idea of electoral-vote, but you are still left unsatisfied if you want to put a number on the chance of one candidate winning the election. There has to be a way to combine who's winning each state with how much they are winning by (and how many electoral votes that state has), to yield finally a probabilitythat a candidate will win.

I'll work on modeling the dynamics of the markets, independently or possibly jointly, to get a clearer picture, updatable in real time, of the election horse-race. Essentially, given the time series data, one can simulate the next 50 days of market prices and get to a distribution 50 days hence (when the election is) of electoral votes for a candidate. You can extract from the markets info like:

Chance Obama wins more than 250 electoral votes: x%
Chance Obama wins more than 260 electoral votes: x%
Chance Obama wins more than 270 electoral votes: x% (and therefore wins the election)
Chance Obama wins more than 280 electoral votes: x%
etc...

This method takes into account how strongly people feel the candidate is winning each state, and boils it down to one number stating the chances the candidate wins the whole election.

Treating the states as independent then, and drawing 53 times (53 days left to election) from a LogNormal distribution of mean 0, standard deviation of that state's past 90 days returns (capping each day at $100), we get an ending price for each contract. If >$50, assign electoral votes, otherwise assign zero. Do this 1000 times.

I DO NOT ENDORSE THESE RESULTS AS BUG-FREE YET, BUT...
Obama, as of right now, has a 40% chance of winning, using this method

EV chance
240+ 89%
250+ 76%
260+ 56%
270+ 40%
280+ 29%
290+ 21%
300+ 16%
310+ 11%