XKCD's Randall Munroe runs Monte Carlo on Intrade's per-state markets

http://xkcd.com/intrade/

by swombat • 18 years ago

38 31 18 years ago

xkcd.com

clay• 18 years ago

So here's the question: why does this give odds of 91% to Obama while Intrade has him forecast at 68%?

andreyf2• 18 years ago

Because markets are imperfect.

"It is based on Intrade.net state-by-state prediction market odds..."

When betting state-by-state, people think Obama is doing better than when they bet nationally.

http://en.wikipedia.org/wiki/Monte_Carlo_method

paulgb• 18 years ago

I bet there's some money to be made on prediction market arbitrage if you know what you're doing.

Also: http://www.fivethirtyeight.com/2008/09/intrade-betting-is-su...

cperciva• 18 years ago

Because the xkcd odds are computed based on the completely bogus assumption that Obama and McCain win or lose states independently.

When we get to election day, the states become pretty independent; but until that day, there are very strong correlations between the results in different states.

andreyf2• 18 years ago

Huh, what?

First: early voting has already started in some states, second: there is always a correlation between how states vote, because they're based on how the candidates have performed.

time_management• 18 years ago

I really hope Munroe's right, and that Obama wins, but I think that 91% is too high. An October surprise could drive a large move, which would make the states very highly correlated. This is much like what often occurs in the stock market. Individual stocks might be only mildly correlated (0.1 - 0.4) in general, but in a crash (or rally) the correlations converge upon 1. Elections are full of last-minute "crashes" and "rallies".

My market on an Obama win is 70-74. I'm including the very serious possibility of election tampering favoring the Republicans, as many allege occurred in 2004, as well as (alas) the Bradley effect. It's difficult to place probabilities on unprecedented, infrequent events, and this is an unprecedented presidential election, not only because Obama is black-- that's only a small component-- but because he's an entirely different kind of political candidate than we've seen in the past, and because we're frankly in a very different country in 2008 than we were in during 2000 or '04.

toby• 18 years ago

Thank you. I actually did this same simulation (I have the Python code if you want it), and got about 80% for Obama -- this was a couple of weeks back.

I sent it around and people rightly pointed out that the states are definitely not independent. The most interesting thing you can do with this is figure out what the implied dependence is.

MaysonL• 18 years ago

Well - the 68% is not really a percentage: it's the price of a contract which will pay 100 to the owner if Obama wins. It's the market's consensus of what the probability is, so yeah it's a percentage. Given that the state-by-state percentages are also contracts, it seems that somebody with a little bit of time on their hands, and some programming smarts, could make a little money doing some arbitrage between the state contracts and the overall contract.

robg2• 18 years ago

See fivethirtyeight.com for much better simulations - 10,000 run each day based on weighted national and state polls. Plus their graphic design is excellent in being both very clean and informative.

Munroe links to them at the bottom.

mdxch• 18 years ago

Betting sites like Intrade are lagging indicators with respect to the polls. Bettors adjust their bets whenever new polls are released, but it doesn't generally work the other way around (people responding to pollsters don't do so on the basis of what they saw on Intrade).

tptacek• 18 years ago

That sounds like a valid point, but bear in mind that a lot of research has been done on the IEM's (more credible than Intrade), and they are more predictive than polls.

steveplace3• 18 years ago

I disagree. I think prediction markets are better because people are risking money. Of course they adjust based on sentiment and other external factors, but I do think it's superior to traditional polling that has significant sampling and measurement problems.

steveplace3• 18 years ago

The "stock" for Obama is at 67.9% on inTrade. Could make 42% on money risked if you're right.

sharksandwich• 18 years ago

If you buy at 68 and the prop is true, you get ~47%

steveplace3• 18 years ago

I've never played intrade's markets, so I don't know how liquid the contracts are... but there is an arb play here if you can time it properly.

Eliezer• 18 years ago

Conditional independence FAIL.

randomwalker• 18 years ago

Thank you, that's what I came here to say. All it takes is one gaffe from Obama to flip all the swing states in one shot.

I'm not sure the 90% number means anything at all. Perhaps it is an estimate of Obama's chances if the election were to be held NOW?

jrockway• 18 years ago

I'm not sure the 90% number means anything at all. Perhaps it is an estimate of Obama's chances if the election were to be held NOW?

This is exactly what the text at the bottom of the page says.

mattmaroon4• 18 years ago

No. It means he's ahead in so many states, that he'd pretty much have to do something big (and it's going to take more than a little faux pas) to fuck it up, and there's about a 10% chance of that.

robg2• 18 years ago

But it's not quite that dependent either. There are stages of dependencies. Some states hang together more tightly (Northeast, West Coast, and Deep South) than others (Midwest and Southeast). And gradual shifts can be detected through polling and pinned to real world events (conventions, crises, etc). The campaigns use game theory to move resources accordingly (see McCain now in Minnesota versus pulling out of Michigan). Obama has a very strong firewall.

That's where, I guess, 538.com has taught me so much. The polls are a constant flow of changing opinions. But if your models are sensitive and well-designed you can see the slow wave shifts developing. It's not predictive, but history helps too. For instance, no candidate has ever been this far up, with this little time left, and lost.

mattmaroon4• 18 years ago

I'm curious how he modeled it. Do actual results follow a bell curve with prior polling as the median and some well-known standard deviation? If so, it would explain why a 10% lead in the polls is considered nearly insurmountable in a given state. I know that if someone is polling at 60-40 in terms of voters, their odds of winning the state are much higher. I've heard 90% thrown out, but don't have any hard data.

Perhaps a better Monte Carlo would use polling data and that.

kul• 18 years ago

Monte Carlo simulations is what they used to work out the odds of sub-prime defaults in their CDOs.

byrneseyeview• 18 years ago

I can assure you that they use addition and subtraction far more often than they use the Monte Carlo method. Adjust your behavior accordingly.

jfarmer• 18 years ago

As other have pointed out, this is silly because the states aren't independent. There's no way Obama is going to lose OH but win IN, for example, although probabilistically speaking there's some chance.

538 models the dependence along demographic lines, so that demographically similar states, districts, and counties are assumed to vote similarly.

mattmaroon4• 18 years ago

This is interesting because it points out yet another opportunity. If you believe 538, you should pretty much be buying Obama to win on Intrade and shorting him on the swing states.

It would hurt my brain too much to do so, but you could probably figure out a betting strategy with a small or moderate assured return.

davidmathers• 18 years ago

I just downloaded Randall's code and modified it to match my analysis, which is: there are only 3 states that matter (FL, OH, CO) and if McCain wins all 3 then McCain wins but if Obama wins any one of them then Obama wins.

Here's what I got: Obama 94.6%, McCain 5.4%

Assuming my analysis is right, am I doing anything wrong mathematically?

chairface• 18 years ago

Am I correct that CO is one of those states that can divide their electoral votes between candidates, and if so, did you take this into account?

evgen• 18 years ago

Nope. Maine and Nebraska are the two that do this.

XKCD's Randall Munroe runs Monte Carlo on Intrade's per-state markets

Related Stories