Thursday, September 04, 2008

Simulating the presidential election - state by state is best

There are a number of sites out there that attempt to predict the presidential election. I prefer the sites that add some simulation to the process so that we can see not just a prediction but the relative accuracy and sensitivity of the results to the input data.

The Presidential Vote Equation derived by Ray C. Fair uses the econometric approach which predicts just the Democratic and Republican popular vote share. Fair's simple regression showed that the important variables are incumbency for the person and party, typically a negative effect, and economic good times are good bad times are bad. This year if the economy stays bad the Democrats are favored, but it could be a dead heat if the economic bad news is not as bad as expected.

I think the best or most interesting predictors are the ones that examine the polls from many states to calculate state by state electoral votes because that is the way the president is chosen, not by popular vote. The best I have seen thus far is (named for the 538 electoral votes). They work through the polls in every state and group like regions together to estimate polls for states when there is a long time between polls (Delaware, for instance, is in the "Acela" Mid-Atlantic group). They predict 307 electoral votes for the Democrats today (up from 303), comfortably more than the 270 needed to win. They also perform simulations based on the error in the polls to generate various scenarios. They have a democratic win in ~70% of the scenarios.

I took a crack at this approach using the Intrade prediction market data. Users of the site can purchase stock in a particular outcome, for instance, Delaware's electoral votes go to the Democrats. Users buy and sell their shares at values based on their confidence in the outcome. It has been shown that this is a predictor of the group's estimation of the probability of this occurrence. I took the probability of the democrats getting the electoral votes and divided by the sum of the either party getting the votes to generate a Democratic probability. Then I used the rand() function in Excel, if it beat the probability, votes went to the Democrats otherwise to the Republicans. Adding up all of the states yields the total electoral votes for the Democrats who need 270 to win.

(click picture or here for larger)

Thus my simulations show that, using the Intrade Sept 2nd data and assuming winner take all in each state (even Maine and Nebraska), the Democrats garner 270 electoral votes or more in 75% of the simulations. The most likely being 277, the median is 289. After the Republican convention is over we can see how this changes. Great care should be taken with these estimates since some of the states have only a few trades and thus the error in the estimated probability is high. I could try to factor that in to the simulation, but I feel the point is proven in principle and we can leave the heavy lifting to the folks that do it for a living, instead of for fun.

No comments: