On correlation

Hi Daniel,

I can indeed comment on correlation between states. There is no need for guesswork here, the history of how the markets move together is all right there in the intrade data, and one can imagine they will continue to move forward together to a similar extent to how they've moved together in the past. My simulation uses the detailed pairwise correlation of intrade markets - and indeed it makes a huge difference and seems to produce more realistic results.

For example, let's pick two states with a pretty large correlation, Ohio and Pennsylvania. Over the last ninety days, their prices have exhibited a positive correlation of about 0.3. The chances of Obama winning Ohio and Pennsylvania respectively are 47.8% and 68%. So let's look at a table of the four possible outcome for OH & PA as a group - first assuming independence, and then from my simulation (which also models the individual variances of OH & PA daily prices and projects those forward).

scenario	Independent	Simulated
OH but not PA	15.3%	0.9%
PA but not OH	35.5%	54.2%
OH & PA	32.5	41.6
Neither	16.7%	3.3%

So notice that in practice, there's a 95% chance Obama wins PA and a 42% chance he wins OH. Nate at fivethirtyeight.com also produces state probabilities every day, and he has PA at 78 & OH at 52. Why does my simulation not produce an Obama victory in PA something more like 68% of the time? That's a subject for a whole other column. Basically, I'm predicting what intrade will be predicting come election day. I think there's a 95% chance that on election day intrade will be selling the 'Democrat wins PA' contract for more than the 'Republican wins PA' contract.

I've noticed that my model broadly agrees with the polls-based approach they have up at fivethirtyeight.com, although I haven't yet produced a graph of what my model would have said at each day in the past.