The best kittens, technology, and video games blog in the world.

Thursday, November 05, 2020

Prediction Markets were Batshit Insane this Election

. by Runs With Scissors from flickr (CC-NC-ND)

The elections are over. The biggest loser are the prediction markets.

As I'm writing this, Arizona has already been called for Biden. And PredictIt has it as 30% chance of Trump.

Trump has no mathematically possible way of winning. PredictIt has it him at 20%.

I talked shit about 538, and I was absolutely right, but even they in the end had Trump at just 10%. Models I'd consider more reasonable like The Economist's had Trump at 3%. Prediction markets had him at completely batshit insane 40%.

And as votes started coming, when Trump did a tiny bit better than expected (but at no point was anywhere near having any chance of winning), Trump peaked at completely ridiculous 72%! It was known long in advantage that election day vote will be a lot more Republican than mail in vote, and states were clear how much of each kind outstanding there is. Prediction markets completely missed that, and had Trump as a favourite just because Republicans mostly went there in person while Democrats mostly sent their votes by mail, and it was too hard for them to understand.

Everything I said turned out to be right. If polls were right, elections would have been very easy Biden victory, and the "normal polling error" benefiting Biden would turn them into a Reagan style landslide.

But not only that - "normal polling error" towards Trump, would still mean easy Biden victory. Even unusually high polling error - what we had this year - didn't bring Trump particularly close to competitive territory. Everything that possible could go right for Republicans were perfectly for them - and Trump lost anyway, because his chances were nonexistent.

Polls weren't great, but they were fine. They predicted Biden to win by about +8%. By the time counting will end, he'll end up closer to +4 or +5%. We won't know for a while, as California takes forever to count their votes, so popular votes always ends up more Democratic than it seems on election day.

It turns out Trump was indeed closer in tipping point states, so Electoral College gave me a few points, but it never managed to overcome that kind of lead. There's no reason to predict which party will be at an advantage in 2024.

Real markets work because they're dominated by sophisticated institutional investors, whose job it is to figure things out, and whose predictions are tested every day. We hoped that small prediction markets dominated by small scale amateurs tested once every 4 years would be just as good, but it turns out they failed utterly.

I'm definitely disappointed as for a while it seemed that we might get great rationalist tool for very little money. It turns out it takes more than that.

EDIT: It's November 13th now, Trump lost nearly two weeks ago, every state has been called long ago, most betting sites paid out on Biden (including one where I bet), and Trump is still 12% on PredictIt. Some people must really hate their own money.

2 comments:

ElephantofDoom said...

To be fair, Biden was only about 20,000 votes in Nevada away from having to take this election to Congress, so I would say that 538's model was pretty good. But Trump getting 40% was stupid.

taw said...

538 model had mixed track record this time. Some of the key assumptions why they had Trump at 30% back in August:
* they had outright pro-Trump bias in fundamentals (because pandemic will be over and economy will largely recover by November)
* they had much bigger error bars than historical data showed
* they assumed that electoral college bias would stay even if big polling error happened

The worst part of 538 model were definitely the pro-Trump economy assumptions, and that's how they had Trump at 30% in August. Fortunately for them they designed their model to remove fundamentals gradually as election comes closer, so by the time of the final forecast they basically removed the worst part of their model, and Trump was down from 30% to 10%.

Those assumptions were never defensible, and in previous years they even showed "polling-only" model without this kind of assumptions.

I don't think big error bars assumption made much sense - even though error was unusually big this time, I think it was just a lucky fluke for them. Sample size of elections is small, and I fully expect accurate polling next time, so I still think they were more wrong than right, but not by that much.

Their assumption that Trump's electoral college advantage held fairly well, at least this year. I thought big polling error would basically randomize electoral college advantage. What we got was fairly uniform swing, plus localized Trump surge in one state that just so happened to be a swing state (Florida), so Trump ended up with even bigger advantage than predicted.

So decent win for their final prediction, but still a major fail for their full model including fundamentals.