I have little understanding of these models, and I hope posters who have experience working with them weigh in (@goofball) but it’s not like the creators are saying “I personally believe these tail outcomes have these likelihoods.” There’s a huge number of inputs and interrelations. It’s probably hard to make a good flexible model that accounts for those in line with the limited historical data without opening the door to some “impossible” tail outcomes.
3 Likes
They kinda are, aren’t they?
They’re saying that Biden has a 53% chance of winning because he wins 270 EVs on 53% of the simulations. If the rules stated that Biden needs 475 EVs to win the election, then their conclusion would be that it has a 0.5% chance of happening, would it not?
Going to be hard for the American public to get behind a family with felony convictions!
3 Likes
I think they would argue they are not making any conclusions. They are making a probabilistic estimate right?
Isn’t that why everyone was lol 538 when trump won and their model only gave him like a 30% chance? People assume they are predicting when they are not?
Edited for accuracy. Thanks goon.
30%
1 Like
Well I for one am convinced. I am 100% not voting for Hunter Biden this November!
9 Likes
I’m not an expert on this sort of stuff but seems the overall purpose of the model would be to help predict who is going to win and not to make accurate predictions of probability of extreme outlier events, so the model could be good at its main purpose and trash out on the tails. (So they probably shouldn’t publish the tails)
Seems kind of analogous to hurricane models where really it’s the 72hr type range people caring about but some the websites will let you look at the models going out 2 weeks in the future which end up showing the hurricane doing figure 8s over Tennessee or striking the UK or something
2 Likes
I’m not familiar with election modeling but as a mathematician i can tell you that tail events are much more dependent on fine properties of the model than typical events (this is the difference between central limit theorems which are universal and large deviations which typically are not)
So as said above it’s very possible to have a good model for predicting typical behaviour with nonsensical tail events
Also im not sure I agree with your estimates…what is the probability of Trump having a stroke and joining a Marxist cult (so noone will vote for him). It’s small but is it closer to 10^-6 or 10^-12 ? I dont know 
4 Likes
If we think that the model is good and it generates tail events that look nonsensical, how do we make sure it’s not just our human bias improperly labeling it as nonsensical?
For example, a four-team betting parlay where each team is 84% to win will only hit 50% of the time, but it feels like it should hit a lot more frequently. I think we suffer from a lack of imagination when it comes tail events where computers do not.
1 Like
Do you have an example of something like this?
I’m basically asking this question but in reverse.
If we think that the model is good and it generates tail events that look nonsensical, how do we make sure it’s not just our human bias improperly labeling it as a good model?
1 Like
Well as you pointed out, a betting market is a good sanity check. I think your position is that you (or anyone) would basically be printing if they laid 33 to 1 against Trump winning NY, and that that is sort of self-evident.
I can see where you are coming from, but it’s not as self-evidently obvious to me. Hochul only won the gov race by 6 points, and there’s some 3rd party uncertainty (Jill Stein, Cornell West, RFK Jr). None of that adds up to any good argument for Trump actually winning. It also assumes like a bottom 5% type news cycle from here to Nov for Biden. Who knows what that might entail exactly. We aren’t very good at figuring out what 33 to 1 is supposed to feel like.
It seems obvious to me. Governor’s races aren’t a good barometer. Vermont has a Republican governor. Kentucky and Kansas have Democratic ones.
In national races, no Republican has won a senatorial election in NY since 1992. Last R presidential candidate to win was Reagan in 1984. Since 1992, the Democrat has won by at least 15 points. Trump lost by 23 points both times. Trump winning would be a seismic shift. None of the polls on their site have Trump winning, or even close. I don’t know what else to call a 3% chance for that anything but laughable.
I think people overstate their ability to know what could happen. Current polls have Biden up 8-10 points in NY - yes, it would take a huge shift for him to lose, but I’m sure there have been elections where polls were off or shifted more than 10 points in 4 months. Whether the odds of that are 1%, 3%, or 5% is basically just guessing.
Just looking quickly at past presidential elections - Carter was up by like 30 on Ford in July 1976, before only winning by 2 points. Sure, polls may be better now and there may be less swing voters, but I don’t think it’s absolutely crazy to say that something could happen that shifts the election 10 points. It is funny to see that what supposedly caused the race to tighten so much was Carter telling Playboy he fancied women other than his wife.
1 Like
You are saying that the uncertainty or variance baked into the model must be off by orders of magnitude because Biden being up ~10 pts in NY polling combined with winning by 20 last time combined with no federal GOP candidate winning in 30 years should equal like 99.x% win frequency.
I think it’s probably healthy to not assume so much stability and certainty into an area with such a small historical sample size.
As a mathematician, don’t you have to start with legitimate variables? We all know that if Trump had a stroke and joined a Marxist cult he wouldn’t lose a single MAGA voter
Three dead girls and 2 live boys?
Here’s the breakdown of the district. These margins are more than just low turnout/engaged voter differences. This is a district in Appalachia, not a purple suburb. Apparently Kripchek is a fantastic candidate, good speaker, very personable, and he’s running in November.
1 Like