Prediction Market Based Electoral Map Forecast

Slashdot points to ElectoralMap.net, a site which aggregates the Intrade prediction market results for individual states in the fall U.S. Presidential election, to predict aggregate electoral college totals and the election winner. I’m bookmarking it as an alternative to media coverage of the race, which will put their own spin on likely outcomes. Note the arrows at the bottom of the map will allow you to track the changes in sentiment over the weeks and months leading up to the election.

A somewhat more elaborate alternative is Electoral Vote Predictor, which uses polls rather than prediction markets. However, they are almost always the same, to tell the truth. I followed this site closely in the 2004 election. Site operator Andrew Tanenbaum includes a frank discussion of the accuracy of the 2004 predictions.

Robin Hanson argued earlier that closely following the news is not that great an idea, and I have to agree. It is something of a bad habit, but I find myself doing it anyway. We have this instinct that choosing our Leader is as important to our lives as it was when we were a tribe of two dozen, and that we have similar influence over the result. Following elections and participating in politics activates these vestigial tribal instincts in much the same was as sports, with similarly futile results. It can be seen as another form of “information porn“, exciting and titillating but ultimately unproductive.

I wish they used a different algorithm in the ElectoralMap site. It appears that the methodology is to take the individual state Intrade predictions, and assign all the electoral votes for a state to the party that has a 55% or greater predicted probability of carrying the state. States with less than a 55% probability for either party are shown as undecided. What I would like to see instead is to multiply the number of electoral votes for a state times the probability that the party will carry that state, and sum those results per party. This way, states that are seen as solidly preferring one party will be given greater weight towards that party than those which are more uncertain.

Then to make it even more useful, run a Monte Carlo simulation where each state is randomly assigned to one party or the other, with the odds taken from the prediction market, to create a simulated election result. Repeat this many times to calculate the mean and standard deviation for possible outcomes. In this way, confidence intervals can be given for the predictions. It’s possible that we might see the predicted electoral votes stay roughly the same for some time, but the confidence interval might narrow or widen. I don’t know if there is some statistical trick to compute the C.I. without actually running simulated elections, just from the distribution of probability estimates, but if so, that would be even better. My suspicion is that realistic confidence intervals are going to show that the election will be too close to call and could easily go either way – at least, that would have been the case in the last few elections.

Why not do this programming myself? Well, I might have the ability to do that. Intrade has an API, but it’s not clear if it allows retrieving the necessary data. I wouldn’t know how to display it on a colorful map, but the math should be simple enough, if you can retrieve the market prices, so producing a textual result would be straightforward. Maybe I will try to work on it in my spare time.

We don’t often post these kinds of pointer articles here, although they dominate many blogs. In that earlier posting about the news, Robin wrote, “I avoid posts that should not be nearly as interesting a year before or after.” This one does not really qualify, although there may be some historical value to be able to go back and look at how market sentiment evolved over time.

GD Star Rating
loading...
Tagged as:
Trackback URL:
  • Jeb

    “Then to make it even more useful, run a Monte Carlo simulation where each state is randomly assigned to one party or the other”

    I believe something like that methodology is used at fivethirtyeight.com. Look at the simulation results on the rightmost bar.

  • Sociology Graduate Student

    Wouldn’t we have to account for correlated error terms? If a prediction market is overestimating Obama’s share of the vote in some states, we should infer it is more likely than not that the market is overestimating his share in other similar states. Maybe I’m missing something but it seems to me that this would underestimate our actual uncertainty about who will win the electoral vote.

  • http://goodmorningeconomics.wordpress.com jsalvati

    The thing you need to remember is that the outcomes of the state electoral votes are NOT independent (people’s psychologies are similar, they have similar biases, the candidates may make a slip that makes them less attractive in all states etc.). For example, in 2006 the TradeSports simultaneously called every individual state’s Senate election correctly, even though they were predicted at fairly even odds. David Pencock makes this point well here: http://blog.oddhead.com/2006/11/10/can-prediction-markets-be-right-too-often/

    Therefore, I don’t think aggregating the prediction market bets like that would be a very good idea.

  • Silas

    Here’s something I want to do (or see done) with the data that I suggested before: compare the time history of a given political bet’s price (Obama’s chance of winning, Democrats’ chances of gaining seats, etc.) with the time history of some economic measure, like the S&P, oil futures, some energy/commodity index, etc. That coud allow you to extract the market’s estimate of how one affect’s the other. (i.e. as the market estimates a higher chance of Obama winning, does it also discount the value of stocks?)

    And to reiterate the issue from last time, no, just looking at election day isn’t enough: other factors could influence the markets or election, it could be decided already, other events could happen that day, and it’s too much for one data point to handle.

  • Douglas Knight

    You could always ask them for code.

    I think the keyword for making maps is GIS, eg,
    GIS-Python. ESRI (ArcInfo) is the dominant brand, so you’ll probably find maps in their “shape” format.

  • http://www.NationalPopularVote.com susan

    The real issue is not how well Obama or McCain might do in the closely divided battleground states, but that we shouldn’t have battleground states and spectator states in the first place. Every vote in every state should be politically relevant in a presidential election. And, every vote should be equal. We should have a national popular vote for President in which the White House goes to the candidate who gets the most popular votes in all 50 states.

    The National Popular Vote bill would guarantee the Presidency to the candidate who receives the most popular votes in all 50 states (and DC). The bill would take effect only when enacted, in identical form, by states possessing a majority of the electoral vote — that is, enough electoral votes to elect a President (270 of 538). When the bill comes into effect, all the electoral votes from those states would be awarded to the presidential candidate who receives the most popular votes in all 50 states (and DC).

    The major shortcoming of the current system of electing the President is that presidential candidates have no reason to poll, visit, advertise, organize, campaign, or worry about the voter concerns in states where they are safely ahead or hopelessly behind. The reason for this is the winner-take-all rule which awards all of a state’s electoral votes to the candidate who gets the most votes in each separate state. Because of this rule, candidates concentrate their attention on a handful of closely divided “battleground” states. Two-thirds of the visits and money are focused in just six states; 88% on 9 states, and 99% of the money goes to just 16 states. Two-thirds of the states and people are merely spectators to the presidential election.

    Another shortcoming of the current system is that a candidate can win the Presidency without winning the most popular votes nationwide.

    The National Popular Vote bill has been approved by 18 legislative chambers (one house in Colorado, Arkansas, Maine, North Carolina, Rhode Island, and Washington, and two houses in Maryland, Illinois, Hawaii, California, and Vermont). It has been enacted into law in Hawaii, Illinois, New Jersey, and Maryland. These states have 50 (19%) of the 270 electoral votes needed to bring this legislation into effect.

    See http://www.NationalPopularVote.com

  • http://www.geocities.com/electionmodel/2008ElectionModel.htm truthisall

    2008 Election Model
    Monte Carlo Electoral Vote Simulation
    TruthIsAll
    Updated: June 12

    http://www.geocities.com/electionmodel/2008ElectionModel.htm

    The model projects that if the election were held today, Obama would win by 314- 224 electoral votes.
    The State model projects that he would win 52.3% of the 2-party vote.
    The National model projects that he would win 52.6%.

    Obama leads the latest state poll aggregate average by 45.7- 43.1%.
    He also leads the latest national poll aggregate by 47.4 – 43.9%.

    The electoral vote is the average of a 5000 election trial Monte Carlo simulation.
    Obama won 99.8% of the trials; that’s the probability he would win.

    The model executes five scenarios of undecided voter allocation.
    In the most likely base case, 60% of undecided voters were allocated to Obama.
    In the worst case, 50% were allocated: Obama had 51.2%, 295 EV and 93% probability.

    The polls indicate that these states will flip to Obama: CO, IA, MO, NM, OH, VA

    These graphs display the trends:
    Aggregate poll shares
    Electoral vote and win probability
    Electoral vote and 2-party vote share
    Latest battleground state polls
    Battleground state win probability
    Projection sensitivity analysis for 5 undecided voter allocation scenarios

    But there’s a catch: the projection assumes a fraud-free election.
    It assumes that the recorded vote will be equal to the True Vote – but it never is.

    Approximately 3-4 million Obama votes will be uncounted (70-80%).
    There will likely be vote-switching on the DREs and central tabulators.
    The Democrats need a very heavy voter turnout to overcome the fraud.
    Obama will probably need at least 54% of the True Vote (2-party) to win.

    Now for the good news: Obama should get the 54% – at a minimum.
    It’s only June and he is leading despite all the media-driven negatives from the primary.
    Bush is at 25% in the polls – and McCain supports his policies.
    His poll numbers vs. McCain can only go up as the focus turns to the election.

    The 2004 Election Model
    Review the methodology in the Election Model.

    The Election Model aggregate state poll closely matched the national polls.
    The aggregate pre-election state polls also matched the pre-election national polls.
    Kerry led the aggregate final state pre-election polls by 47.7- 47.0%.
    Kerry led the national 18-poll average by 47.3 – 46.9%.

    The Election Model state model projections closely matched the national average.
    The State projection had Kerry winning 51.8% of the 2-party vote and 337 electoral votes.
    The National projection had him winning 51.6% of the 2-party vote.

    Exit Pollsters Edison-Mitofsky released their 2004 Evaluation report in Jan. 2005.
    It indicated that Kerry won the unadjusted (WPD) aggregate state exit polls by 51.8 – 47.2%.
    The 12:22am National Exit Poll update indicated that Kerry won by 50.8 – 48.2%.
    E-M stated that respondents were randomly selected and that the margin of error was 1.0%.

    Professional pollsters allocate undecided voters in every election.
    The majority are allocated to the challenger, especially if the incumbent is unpopular.
    Bush had a 48% approval rating on Election Day.

    The Gallup Poll allocated 90% of undecided voters to Kerry. Harris and Zogby gave him 67-80%.
    The Election Model final base case projection allocated 75% of undecided voters to Kerry.

    The Election Calculator Model
    This model calculates the True vote based on estimated shares of returning voters.
    The National Exit Poll “How Voted” category is a reasonable best case estimate.
    Returning prior election voters equal total votes cast less mortality multiplied by the turnout.
    In 2000, the Census reported 110.8m votes cast, but only 105.4m were recorded.
    In 2004, the Census reported 125.7m votes cast, but only 122.3m were recorded.

    The Election Calculator determined that Kerry won the True Vote in a 67-57m landslide (53.2 – 45.4%).

    The Calculator determined that Obama should win the True Vote in a 71-59m landslide (54 – 45%).
    The 2004 True Vote was the basis for calculating the 2008 model.

    Pre-election and Exit Poll Confirmation
    Final pre-election polls showed Kerry and Bush in a virtual tie.
    The unweighted state average favored Bush, but Kerry led the aggregate weighted average.
    The pre-election aggregate state polls closely matched the national pre-election poll average.

    Some naysayers maintain that polling analysis cannot prove that the 2004 election was stolen.
    They claim that the Final National Exit Poll was correct and the unadjusted exit polls were wrong.
    But after undecided voters were allocated, the pre-election polls matched the unadjusted exits.

    The Final National Exit Poll was forced to match the recorded vote using impossible weightings.
    The Final NEP indicated that Bush 2000 voters comprised 43% (52.6m) of the 2004 electorate (122.3m).
    But Bush only had 50.5m recorded votes in 2000.
    And approximately 2m Bush 2000 voters died prior to 2004; another 2m did not vote.
    The Final Exit Poll had to inflate returning Bush voters by 6m (52.6-46.5) to match the recorded vote.

    Where did Bush find 16 million new voters?
    Kerry won 64% of returning Nader voters. Bush had 17%.
    He won 57-60% of new voters and others who did not vote in 2000.
    He won at least 91% of returning Gore voters who saw their votes nullified in 2000.
    He won at least 70% of late undecided voters.

    Kerry also won the final 5m recorded votes by 54.3-45.7%.
    These were late votes recorded a few days after the election.
    They consisted primarily of absentee and provisional ballots.

    Bush won the 2004 recorded vote by 62.0-59.0m, 50.7- 48.3% and 286 EV.
    Of 3.4 million uncounted votes, approximately 2.6m were for Kerry – a net 1.8m margin.
    Bush’s 3.0m margin is therefore reduced to 1.2m – and that’s before vote-switching.

    To believe that Bush won, one must believe that the pre-election and unadjusted exit polls were wrong.
    And also believe that the Final Exit Poll, although mathematically impossible, was correct.

  • http://kenzoid.com Ken Kennedy

    Another Monte Carlo simulator is at Hominid Views: http://hominidviews.com

    Per the FAQ, (http://hominidviews.com/?page_id=1354), his simulations are “Monte-Carlo simulations of the Electoral College outcome based on state head-to-head polling data.” His most recent analysis…”showed Sen. Barack Obama leading Sen. John McCain by (on average) 277 to 261 electoral votes. If a general election was held now, Obama’s chances of winning were about 67.6% (if we give ties to Obama).”

  • Evelyn

    GoogleCharts now allows a colored geographical map:
    http://code.google.com/apis/chart/#maps

  • http://electoralmap.net/index.php Doug

    I’ve added the state by state summation you suggested, at least for today (6/16). If it continues to be interesting I’ll make it a permanent feature.

    I think the Monte Carlo is a good idea too, and I’ll probably add that when I have some time.

  • http://www.anvin.org Suzi

    FYI, another predictions market, hubdub, actually does things by state (predictions voted by state) and can give you proportional coloring on their map. Check it out at http://www.hubdub.com/election_map

  • Pingback: The addiction to politics » Against Politics