This shit just makes me nostalgic for the glory days of Qanon indictments markets. .07 is nothing, We had biannual"Will Hillary/Comey/Rosenstein/Clapper/Brennan be indicted" markets in the 10-20% range and one of them got up to 60% at one point, it was glorious. I got so much absolutely free money.
1 Like
I think we basically agree although I don’t think what happened in Wisconsin was some completely random polling miss, similar things happened across the Rust Belt, it was some sort of systemic error in that region. Maybe that has been corrected now but it’s a little speculative to assume it has been. I guess my argument is that to the extent, say, Michigan 2016 and Wisconsin 2016 were similar, it’s quite possible Wisconsin 2016 and Wisconsin 2020 will also look similar. I’m hesitant to draw conclusions from midterms which have a very different voter profile.
The other thing is, and I can’t really express this in a coherent mathematical way, that I have a hard time believing that the polling average represents an actual average outcome. What I mean by that is, OK, the average is Biden +6.5 roughly, but an actual result of Biden +2 seems WAY, WAY, WAY more likely to me than an actual result of Biden +11, which feels borderline impossible. It seems to me like while the polling outcomes are normally distributed thanks to polling error, the range of actual possible outcomes is skewed much more back towards a close race.
1 Like
To try to express myself a bit more clearly, the variance around the average in polls is down to sampling error, just random deviation from the correct result. But deviation from polling average on election day is down to events just before the election - which I’ll ignore here - and systematic polling error.
To use an analogy, suppose I show you a weather forecasting service and tell you that it’s quite possible, although not certain, that the service has a systematic deviation from what the true temperature will be. If it’s predicting temperatures like 70F for several days next week, that implies nothing about what the systematic deviation, if any, will be. But if it’s predicting temperatures like 115F, OK, maybe that’s a true forecast, but if there does turn out to be a systematic error, it’s much more likely that it’s predicting temperatures that are too hot than too cold, because temperatures as high as 115F are already uncommon.
Similarly in a Presidential election, we just don’t live in a world anymore where people can win 60+% of the popular vote. When polls are predicting Biden +6.5% in Wisconsin, the range of possible polling errors can’t look anything like a normal distribution. Either there’s no systematic error, or there’s one that is making Biden’s numbers look better. That’s my concern. The fact that there was exactly such a systematic error last time and I don’t know if it was corrected becomes more concerning given this backdrop. I have no clue if models take this into account already.
538’s current prediction is 77% chance of Biden win in Wisconsin, but look at their analysis:
Let’s breeze straight past the dubious adjustments for recent convention bounces and go straight to the bit where they allocate undecided voters 50/50? Like what odds would you want to bet that undecided votes will break for Biden rather than Trump?
Ok, Wice’s post was mainly just zero-analysis wish-casting garbage. I’ve changed my mind on him.
The markets have priced in Trump stealing the election, which is probably explains why Biden’s odds don’t reflect his poll numbers: Betfair, which isn’t available to American Trump button mashers, has Biden as only a 55% chance of being the next president.
I see what you’re saying here but I think it generally reflects a bit of recency bias. I agree with the framing too, when I built models for 2012 and 2016 I thought of uncertainty in 3 different ways:
- Sampling uncertainty - the probability that the sample just doesn’t reflect the population
- Systemic uncertainty - the probability that the way the polls is conducted (via questions asked, weighting, etc.) doesn’t accurately reflect the race
- Temporal uncertainty - time will pass between now and election day and stuff can happen
#2 is clearly what happened in 2016, and it happened in the same way in similar states. While it is reasonable to think it might happen again, I’m not aware of historic evidence to support the hypothesis that when there’s systemic error it either favors republicans, the incumbent, or the candidate currently ahead. Are you? I get the feeling but I think you’re over-litigating 2016. I’m not sure why we couldn’t also see a systemic error in Biden’s favor given how embarrassed pollsters were in 2016.
TBH I find their Step (4) more dubious than step 2. I think I looked into this and couldn’t find any data to support undecideds breaking in a specific way that’s predictable (obviously, in 2016 they broke for trump probably due to just not liking clinton + the comey letter).
I’m not sure “economics & incumbency” shouldn’t be fully dropped at this point, but I imagine Nate had his reasons for still weighting it at 17%.
I thought his temporal uncertainty was too high - I think when he posted the model he said Biden would be 90% in an election held today, and that the big difference was time passing and stuff happening. I think everyone is pretty saturated with news and even traditional “big” stories might not move things much. Plus so many voters are just locked in either pro or anti trump no matter what. That said, the Woodward tapes seem like a pretty big deal so maybe I was wrong.
1 Like
I don’t have a mathematical argument here but I think there are more embarrassed Trump voters out there than embarrassed Biden voters. I suspect a Bradley effect. I’ll lay you +110 that undecideds break for Biden if you like.
One thing that was crazy to me when I was doing some googling for figures and stuff related to this conversation is that mid last year, over 50% of likely voters said they would not consider voting for Trump. But then in a The Hill/HarrisX poll in February this year, 62% of voters said they would consider voting for Trump. The primary reason given was “the economy”. Swing voters are totally brain dead so it’s hard to predict the absurd things that might get them to change their minds.
1 Like
It’ll be forgotten in less than a month.
The “Will Donald Trump lose any state he won in 2016?” market seems quite juicy. I bought some today at $0.82.
I feel like you’d get better value betting on Biden to win Michigan at $0.72. I don’t think Trump wins MI unless he also wins all the states he won in 2016.
1 Like
I have some of that too. But if rust belt goes ugly, could still hit AZ, FL, or NC. And it’s possible to lose MI but win WI or PA, though unlikely.
Yeah, if you’re looking for a Biden state, it’s MI at this price, or WI, imo - sure as hell not Ohio, Florida or Texas.
If Biden doesn’t obliterate the rust belt, it’s over.
1 Like
Yup. If Trump wins MI, Biden isn’t winning FL or NC, AZ is maybe possible because of demographic shifts, but I doubt it.
Minnesota might be the best value right now imo.
Trump barely won FL. And 2018 it was basically a tie. And it’s very different from the rust belt, it’s still live in this scenario.
it’s Florida man
no it’s not
I mean, that just sounds kind of annoying to specify - whole country? aggregate specific states? It’s a bet I’d take but I dunno, I can just agree to acknowledge being wrong too, might make it easier.
The election is only 7 weeks away though, could last that long. But you’re probably right