Predicting the 2008 Presidential Election
I am a fan of prediction markets. They have typically done much better than polls at predicting the outcome of elections. Why? Here’s a thought experiment. Consider who you think is going to win the election (not who you want to win). Now consider that I was going to bet you $10,000 of your hard earned money on whether your prediction comes true. Did that change your thinking at all? Some of you might have even switched candidates once money was on the line. That’s the difference between a poll and a prediction market.
Two big prediction markets (Betfair and Intrade) currently show Obama with around a 65% chance of winning. Not surprisingly (given that there’s nothing at stake for the people polled) CNN’s current poll calls it much closer with Obama leading 47% to 43% for McCain (with 10% of those polled not sure). [Update: I just realized I compared apples and oranges when I wrote this since CNN’s is a popular poll not a projection of who will win based on electoral college. But this won’t affect the punchline….]
There’s an interesting new twist in forecasting analysis though that I’d like people’s opinions on. It’s the simulation approach as embodied in FiveThirtyEight.com. This analysis has Obama currently at a whopping 83% to win!
Of course one’s first instinct might be to question the validity of the analysis. After all, the site is public information, and if the analysis is valid it should be priced into the prediction markets. But I’m not so sure we can dismiss it that easily. For one, the prediction markets themselves have shown long periods of “mispricing” where one could arbitrage between two markets. Secondly, it stands to reason that if investors don’t understand how the analysis being done is deeper than other publicly available information, then they will discount the analysis. Finally — and this gets into the heart of complex systems thinking — markets are averagers, whereas the simulation approach being used by FiveThirtyEight.com seems to model the system being predicted very well. This includes the various non-linear convergences and divergences, tipping points, info cascades, etc, that is beyond the ken of an averaging model (like a market).