http://www.geocities.com/electionmodel/ElectionForecastingGame.htmThe Election Forecasting Game
TruthIsAll
Academics and political scientists create and execute election forecast models months in advance of an election. The models utilize time-series data for economic growth, inflation, job growth, interest rates, foreign policy, historical elections, incumbency, approval rating, etc. The models are interesting and elegant but they don’t account for changing voter psychology best expressed in pre-election preference polls. Academic regression model forecasts are executed months before the election, whereas poll-based models are run frequently right up to Election Day.
Polling and regression models are analogous to the market and theoretical value of a stock. A stock’s intrinsic value is based on discounted forecast cash flows and is rarely equal to current market value. Polling data can be considered as the current share market value; the regression model vote share is equivalent to intrinsic value. On Election Day, the current share price is all that matters.
Inherent problems exist in all election forecasting models. The most important is never discussed: election fraud. The assumption is that the model will be judged based on its success in predicting the recorded vote share. Of course we know that the recorded share is never the same as the True Vote. It’s amazing that sophisticated political scientists fail to even consider fraud in their models.
But let’s move on to other issues. As I see it, a major drawback in academic regression models, apart from those just mentioned, is that they are designed to forecast a popular vote share, but not the all-important electoral vote. State-based polling models utilize the latest polling results in calculating a corresponding electoral vote. The Election Model goes further: it calculates state win probabilities that are input to a 5000 trial Monte Carlo simulation. The objective is to determine an electoral vote win probability. Calculating the expected electoral vote does not require a simulation. It is simply the sum of all 51 state electoral votes times the win probability.
The 2008 Election Model executed a Monte Carlo simulation just before the election. Obama's projected share was 53.1% with 365.3 expected electoral votes. The simulation mean EV was 365.8, the median 367 and the mode 371.
Obama won the recorded vote with a 52.9% share and had 365 electoral votes. His margin was 9.5 million votes.
BUT THE MODEL WAS WRONG.
Obama did much better than that. The state pre-election polls used in the simulation were likely voter (Obama led LV polls by 7%), not registered voter polls. (he led the RV polls by 13%).
Obama won the True Vote with 57% and had over 400 EV. He won by over 22 million votes.
Although the model predicted the recorded vote, it failed to project the True Vote.
Presidential election forecasting models should have the following disclaimer: The forecast is expected to deviate from the official recorded vote. If they are nearly equal, it indicates one or more of the following: a) input data errors, b) incorrect assumptions, c) faulty model logic and/or methodology.
Printer-friendly format
Email this thread to a friend
Bookmark this thread
(150 posts)
Send PM |
Profile |
Ignore
