The proper way to predict Obamas electoral vote count

No Gravatar

I&#8217-m puzzled by the way Intrade projects the electoral vote count on its home page. Two methods are proposed: (a) add up the votes of all the states that are &#8220-leaning&#8221- (&gt-50%) for a candidate, or (b) compute a price-weighted average. The latter is obviously meaningless because electoral votes are winner-take-all in pretty much every state.

But what about the &#8220-leaning&#8221- method? Well, it only makes sense if you believe that market prices do not represent probabilities. In fact, the &#8220-leaning&#8221- method treats the 15 electoral votes from a swing state like North Carolina (65% chance to go blue) the same way it treats the 15 votes from a true-blue state like New Jersey (95% chance to go blue), tossing them both equally in the Obama column.

Now, Intrade has been known to want it both ways: interpreting its prices as probabilities most of the time, but then also claiming that it correctly predicted all 50 states in 2004 because all the contracts priced over 50% eventually expired at 100%. This claim conveniently ignores the fact that if prices are probabilities, then at least some of the states priced &#8220-red&#8221- or &#8220-blue&#8221- over 50% should in fact have gone the other way on election day.

It may very well be that, given the 2004 data, the prices of election markets (on Intrade and elsewhere) should not be interpreted as probabilities. Perhaps our academic friends can come up with another meaningful way of looking at those prices. But in the meantime, assuming the price/probability correlation holds, the proper way to project the electoral vote count from the market prices is to run monte-carlo simulations based on individual state prices.

Here&#8217-s an example using this morning&#8217-s prices on NewsFutures. The histogram shows the results of 1 million simulated elections where each state goes red or blue according to its market-derived probability of doing so. Note how the &#8220-most likely&#8221- outcome, 364 votes for Obama – which is the number the &#8220-leaning&#8221- method would report – is at the same time very unlikely with just 5% chance of happening.

10 thoughts on “The proper way to predict Obamas electoral vote count

  1. Chris F. Masse said:

    Very smart post. I will have to think hard about that. And, yes, we need external opinions from our prediction market luminaries.

    Great, a polemique. It was badly needed. It was too calm, on Midas Oracle, recently.

  2. Lance Fortnow said:

    The price-weighted average is not at all meaningless, it’s the expected number of electoral votes the candidate would get when the prices are interpreted as probabilities.

  3. Chris F. Masse said:

    But what about the “leaning” method, prof Lance Fortnow?

  4. teaNoranges said:

    I second Mr. Fortnow. If the prices are probabilities, the Intrade average value is the mean of the distribution of possible outcomes, e.g. the mean number of electoral votes for the candidates.

    That the first moment of the distribution follows so elegantly from prediction market data is a bit of miracle. How would one create such a number from polls? It’s only possible by making several assumptions and/or running complex models.

    But back to Emile’s point: All the prediction markets would be served by a standardized metric to assess how successful they were in retrospect. For individual markets this is easy, but for a set of correlated markets (e.g. the state races) it gets tricky.

  5. teaNoranges said:

    Emile: Cody Stumpo at Intrade has been using a nice method to solve the problem you’ve come across. He uses all the historical state market prices to write down a 51×51 matrix of state-state correlations, and a 1×51 matrix of state volatilities. Then the 51 state prices are simulated out to election day. This gives a full probability distribution of EC votes for each candidate. His numbers coming from this method have tracked Nate Silver’s pretty closely.

  6. Chris F. Masse said:

    “We accurately predicted the outcome for every single state in the 2004 presidential election and all but one of the Senate races.”…..C2%A0xanax

  7. How to oversell InTrade's predictive power | Midas Oracle .ORG said:

    […] (which event derivative traders rely on to get informed), and they tell gullible reporters that InTrade has magical power. It’s total bullshit. The life of the InTrade prediction market on “Joe Biden as the VP […]

  8. Prediction Markets - The Day After | Midas Oracle .COM said:

    […] You might re-read Emile Servan-Schreiber’s post, which deals with interpreting prediction market prices as proba…. […]

  9. Prediction Markets - The Day After | Midas Oracle .ORG said:

    […] You might re-read Emile Servan-Schreiber’s post, which deals with interpreting prediction market prices as proba…. […]

Leave a Reply

Your email address will not be published. Required fields are marked *