Read on Twitter
Bringing my replies over to my timeline, @thogge asks a good question: why are betting markets so different than polls? And because betting markets have skin in the game, are they a better predictor? https://twitter.com/thogge/status/1323461075170324481
Short answer: in theory yes, in practice, probably no more reliable. There are a bunch of reasons. This Predictit data is useless, volume here is negligible, I think it's $25K bet on Trump today = nothing.
However, my home team Befair is doing legit volume here -- over $300M -- so this is a liquid market. It's not quite as different from the pools but still implies 65% for Biden, much lower than most of the polling / pundit %s. Should Biden voters be concerned?
While clever people like to fetishize prediction markets and markets in general as being more predictive, I do not think this is a particularly meaningful piece of data.
I think of prediction markets as having a couple ways of revealing insight:
I think of prediction markets as having a couple ways of revealing insight:
Most powerful is ferreting out material non-public information. We can imagine an event where some people know the winner but the public does not -- an Oscar market. If the prize got big enough, eventually someone would bet with their inside knowledge and the markets would move.
But, no one knows what will happen Tuesday. There really is very little to no inside information that can be extracted by the incentives of the market.
The second way is that sophisticated math-y people invest time, energy and money to create superior methods of modeling the result to where they can consistently predict the winner better than the casual or even serious bettor. These people win CONSISTENTLY and move markets.
This involves modeling, backtesting, finding proprietary data sources, etc. It needs repeated events. Because of the investment you want to be able to deploy it repeatedly to generate a stream of profits. None of this is really possible on a US President election market.
It happens once every four years, there is no possible backtesting, different candidates ever time, changing rules and turnout, etc. In theory could someone with a giant profit motive model better? Yes. Is it happening here? I seriously doubt it.
So, let's look back. With over $100M bet before the 2016 election, it was liquid. How did the markets do? Day before: "83% chance of Clinton"
https://www.cnbc.com/2016/11/07/betting-sites-see-record-wagering-on-us-presidential-election.html
https://www.cnbc.com/2016/11/07/betting-sites-see-record-wagering-on-us-presidential-election.html
8am election day: Markets say 82% Clinton. So, we think these are the geniuses who are proving the pollsters wrong this year? https://twitter.com/jdh/status/796022033289134080?s=20
1:38pm election day: Clinton still at 82% https://twitter.com/jdh/status/796104575107219456?s=20
Here we go, now the markets are onto it: 5pm election day... Donald Trump down to 8% chance of winning! https://twitter.com/jdh/status/796155680918937600?s=20
Oh shit, 6:26pm: Clinton down to 55%.
What's important here: new info is out, polls begin to close, until this it was all noise, now the markets are useful, quickly digesting new info that comes out and weighting it, probably better than your TV dude https://twitter.com/jdh/status/796177062394368000?s=20
What's important here: new info is out, polls begin to close, until this it was all noise, now the markets are useful, quickly digesting new info that comes out and weighting it, probably better than your TV dude https://twitter.com/jdh/status/796177062394368000?s=20
8:27 pm election night: Clinton down to like 2%. There's your prediction market insight for you: it figured out that Clinton lost, well, about the same time everybody else did.
I don't see any reason to believe this year's prediction markets would be particularly more useful.
I don't see any reason to believe this year's prediction markets would be particularly more useful.
Someone shared this, which is interesting and relevant:
http://quant.am/statistics/2020/10/11/taleb-silver-feud/
http://quant.am/statistics/2020/10/11/taleb-silver-feud/
Slept on it and had a couple of thoughts, also @pentagoniac in the replies has some related thoughts worth reading I see.
#1- easy to look at the 82% Clinton in 2016 and say the markets were wrong but we really don't know. Longshots happen, and it isn't necessarily because the model was wrong -- it's not true to think the correct prediction on Nov 8 2016 would have been 100% Trump even though he won
#2- the info from 2016 indicates that markets had no better insight than polls about Trump. But I guess what's different 2020 is: last time, polls and markets were saying roughly the same thing -- Clinton as a prohibitive favorite. Here, markets materially more bullish on Trump
Do markets include some MNPI? I suppose it's possible that someone in AZ PA FL knows the early voting totals, it sways more to Trump than expected, and is betting that info. Seems unlikely, difference has persisted for a while - did they know 2 weeks ago?
What's more likely is the market participants feel the pollsters are constrained by survey data, and that they can incorporate non-poll info, for example they think the GOP will use courts and poll intimidation to shift the actual yield from the voter intent expressed in polls
What the different pricing is telling you is probably that those market participants see a significant probability that either pollsters are wrong, or that interventions outside of the voting process significantly moves the market towards Trump
I remain skeptical that there is insight here. I have talked to some of these bettors and so far all have said things like "the polls were wrong in 2016" "people don't want to say Trump on the poll call" "I think Trump will have more success in courts" "SCOTUS is GOP"
These people, in the sports betting world, while they sound smart, they are the dumb money. They are where the profit comes from. Going against the probabilities by backing smart sounding, completely unscientific theories
Consistent winners that move markets don't bet intuition - they are investing in repeatable models, proprietary data sources, sophisticated prediction techniques, and a every-four-year event with different contestants each time does not seem susceptible to that kind of insight
So while it's possible the better odds for Trump in the betting market contains signal, I personally would guess it is just noise.
We'll never know! Either Biden will win, which both the polls and betting markets predict as probably, or Trump will win, which they both say is improbable. With a sample of one, no way to tell whether the odds were right or wrong.
