Friday, September 18, 2015

Digital Politics: PredictIt

In 2012 the company Intrade made a splash in political prediction markets. It closed after several difficulties (new regulations made accepting American customers illegal). Today prediction-market trading is back on American soil with PredictIt! What is this--and how does it work?

Prediction Commodity Markets

The way markets work is that buyers are matched with each other at a ratio that changes as people enter and leave. If you think something has a 60% chance of occurring you pay 60 cents per share and are matched with another buyer who thinks the opposite (they paid 40 cents per share). If the aggregate reaches 70% Yes, you could sell your 60's and get a .10-cent profit per share.

If the event happens, you get paid off at $1.00. If it doesn't, you get zero (the NO buyers get the $1.00 per share).

Let's look!

Who Is The Next To Fall?

There are all kinds of markets--but the one that warms The Omnivore's heart is "Next Republican to drop out." On the page it looks like this:
Right now the big money says George Pataki is the next out--followed by . . . Bobby Jindal. The Omnivore, though, thinks maybe it's Scott Walker.

Walker's next-to-fall star is rising: +1 cent (note: that's not good for Walker). On the other hand, YES buys are comparatively cheap at .19. If The Omnivore bought a dollar's worth, he would make five if Walker quit tomorrow.

When you actually click on the buy you get this:
A few things to note here:
  1. There's little track record here--this market hasn't been around all that long.
  2. The $23 price seems to be the average. The detail says you can buy for 15 (YES) or 93 (NO).
  3. The volume (19) and total shares (134) is pretty low. Compare to Bush Wins The Nomination at 1,400 volume and 33,478 shares.
Let's also not forget the ever-popular comments section (see Larry The Betting Guy down there). This is where people pretend to be discussing the event but are really trying to manipulate readers into taking what they think are sucker-bets.

The Madness of Crowds

We are told that betting markets are super-good at predicting elections. Here's a Huff-Po piece that compares "polling" to "betting" in October 2012:
Added Paulick Report editor Ray Paulick, one of America's top horseracing handicappers and a political prediction markets aficionado, "Gamblers have more experience with cheaters. They take voter fraud into their metrics. Polls don't. Nor do polls take into account intangibles like how each state's secretary of state factors in or systems within a state designed to eliminate voters."
Frankly speaking, this is bullshit. Gamblers trying to take election fraud into account should also be using astrology. Intangibles like the state's secretary of state? There is a gambling term for people who use a "system" based on "intangibles" to win: Losers. Whatever is going on with gambling's track record, this explanation isn't it.

The Omnivore submits the following:
This is the Princeton Election Consortium's Electoral Vote predictor for 2012. The red bar is the "strike zone" (68% likely). The Yellow bar is the 95% accuracy. It is all above the red-line (above the red line is Obama Wins).

What this means is that the polling aggregate for 2012, on the eve of the election, put Obama's victory chances above 95%. If you follow the black line (and the gray 95% probability shadow around it) Obama was always winning--and save for a short period of time after the first debate, winning within a 95% zone of accuracy or more.

For a gambler who just has to pick YES/NO, this is one of the most predictable bets possible. It should be no surprise that gamblers were more willing to invest in an Obama victory than an individual poll would show.

So the question that's outstanding here is whether or not the prediction markets are more predictive than the polls. Let's check a few:

  1. Trump Polling above 25% on Sept 30th? 78% YES
  2. Government Shutdown on Oct 1:  64% NO
  3. Biggest Bounce From Debate:67% FIORINA, 15% TRUMP, 11% CARSON
  4. Worst Slide From Debate: 50% TRUMP, 37% CARSON, 7% JEB
It should be noted that each of these has a very specific RULES tab. For the polling bounce/loss the standard is the RCP average on Sept 23 (in five days). So let's see what happens between now and then!


Right now the betting money says Jeb is the candidate (34%, followed by Rubio 32%)  and Clinton wins the White House (46%) followed by Bush (29%). These are safe-money bets--but whether or not they are predictive is anyone's guess.

Questions for PredictIt:

  1. How, historically, does volume relate to accuracy across these prediction markets?
  2. Are there cases where prediction markets have predicted a "black swan" event? 
  3. As the numbers change, what's the average rate-of-change? This is interesting because right now we are at a 64%  NO for government shutdowns--as new information comes in, this will presumably shift. The question is how "sensitive" are these markets in general? How much of a leading indicator is the betting market (another way of saying this is how fast do the markets move before the polls).
  4. Is there deeper data that PredictIt has that they'll sell to campaigns based?
  5. Can we see exact time-stamps for the moves? Can PredictIt track these movements to things like debate events, specific TV coverage, or Twitter activity?