On Tuesday Cruz took more than 50% of the vote in Utah, netting him all of Utah’s 40 delegates. In the three democratic contests, Sanders received about 20 more pledged delegates than Clinton. How well did the markets predict these outcomes? Not at all, because we only had markets on the winners of the primaries but not the spread. The markets predicted all five of Tuesday’s winners correctly but that didn’t give major insight into how the delegates get allocated, except for the winner-take-all republican primary in Arizona.
It would have been nice to have had a market on whether Cruz would top the 50% mark in Utah but would you create different markets for each state based on the rules and the current positions of the candidates? There is an all purpose solution: Right now a security SANDERS.WIPRMRY16.DEM, to use the Predictit naming scheme, pays off $1.00 if Sanders wins the Wisconsin primary and $0.00 if he loses. Instead suppose we had a scaled security that pays off the percentage of pledged delegates that Sanders wins in the primary. For example SANDERS.IDPRMRY16.DEM.SCALED would have paid off $0.78 since Sanders received 18 pledged delegates out of the 23 pledged delegates available in Idaho. This method works across different state allocation procedures, CRUZ.UTPRMRY16.REP.SCALED and TRUMP.AZPRMRY16.REP.SCALED would still both pay $1.00 since Cruz and Trump received all the delegates in Utah and Arizona respectively.
The value of the security SANDERS.WIPRMRY16.DEM.SCALED would be the expected percentage of delegates to be won by Sanders in Wisconsin. Expectations have nice properties. We can multiply by the number of pledged delegates to get to expected number of delegates to be won in a state. We can add the expected number of delegates in multiple states together to get the expectation of the sum, even if the results are not independent of each other. So we could use the scaled security to truly see the predicted effect of a week’s worth of primaries and caucuses on the delegate count.
Scaled prediction markets do carry their own challenges. It’s harder for a bettor who would have to understand the sometimes complicated rules for a state primary/caucus to make a good prediction. Many of the primaries allocate several delegates by congressional district so state-wide polls don’t always give the full picture. In a proportional primary the payoffs might be a relatively narrower range limiting the return to the bettor. The final delegate count for a state may take several days or weeks to become official.
More importantly the public and the media seem to care more for the winners than the more important effect on the delegate count. Even the vaulted New York Times titled their results article Clinton and Trump Win Arizona; Cruz Picks Up Utah; Sanders Takes 2 though to their credit the article discusses early the delegate counts.
This is just a general tension we have: prediction markets get popular by having securities that people want to vote on, which may not be exactly what we can use to make good predictions about what matters. Life will get better for the general election since most states use winner takes all for the electoral college.
Lance Fortnow is a professor and chair at the School of Computer Science at the Georgia Tech College of Computing