So, I showed a few friends my model and the general results and they said that the polls were wrong in 2016 so that doesn't mean I'm right.
Sure. But I did get 2016 more right than even Nate Silver, so maybe give me the benefit of the doubt.
Anyway, i know I did a few models already of 2016 back in 2020 just to test my model, but I figured I'd do a more expensive run through of 2016 based on the prediction I made at the time.
Last time I did this, 8 predictions went to Hillary and 2 went to Trump. Let's see how this goes this time.
Trial 1: 311-227 Clinton
Very bullish for clinton outcome.
Trial 2: 296-242 Trump
Random Trump win with AZ going blue. Why does AZ keep wanting to go blue? I have no idea.
Trial 3: 284-254 Trump
Another trump win. I know we're talking 3 trials and Biden was ahead at this point too, but this is happening surprisingly often compared to last time. Unless last time was just a crap sample size.
Trial 4: 298-240 Clinton
A predictable result. Honestly, not sure how the fudge Trump managed to flip Wisconsin in this model.
Trial 5: 278-260 Clinton
A narrow expected win but a win nonetheless
Trial 6: 280-256 Trump
Anyone who in 2016 didn't think trump had a chance on election day was literally delusional.
Trial 7: 295-243 Trump
WHAT IS HAPPENING?! OMFG!
Trial 8: 284-254 Clinton
Ok, back to clinton. Reminder. Last time I did this with this same data, 8 outcomes with Clinton and 2 Trump. Now we're at 4-4.
Trial 9: 302-236 Clinton
Blue Georgia but Red Pennsylvania. See? Statistical quirks do happen.
Trial 10: 306-232 Trump
Weird Trump win except the rust belt goes red and the sun belt goes blue.
Trial 11: 281-257 Clinton
Trial 12: 294-244 Clinton
The weird one where Clinton manages to take South Carolina. I mean it was as likely as Trump taking Wisconsin so I guess it's possible.
Trial 13: 291-247 Clinton
Trial 14: 319-219 Clinton
A very strong one for Clinton.
Trial 15: 290-248 Clinton
Trial 16: 279-259 Trump
Trial 17: 275-263 Clinton
Trial 18: 273-265 Trump
Another weird Trump win where Trump gets the rust belt but Clinton gets the sun belt
Trial 19: 298-270 Clinton
Trial 20: 288-250 Clinton
Strong rust belt showing. I could kinda see Bernie getting this result.
So at this point we're 7 Trump 13 Clinton. Let's keep going.
Trial 21: 316-222 Clinton
We finally did it. We flipped Wisconsin. But then Hillary flipped Arizona and Ohio so it was a net loss.
Trial 22: 307-231 Clinton
This one is nonsensical. Clinton won both congressional districts but lost Maine itself. That's stats for you. Otherwise a strong win for Clinton.
Trial 23: 308-230 Clinton
Clinton wins SC again.
Trial 24: 284-254 Clinton
Trial 25: 274-264 Clinton
Trial 26: 287-251 Clinton
Another weird SC win for Clinton
Trial 27: 306-232 Clinton
Another weird sun belt win for Clinton
Trial 28: 283-255 Clinton
When you win Ohio but somehow lose Pennsylvania...
Trial 29: 272-266 Clinton
Oh hey we finally got the result I predicted.
Trial 30: 274-264 Trump
Trump managed to eek out a lucky win here thanks to flipping Minnesota and New Mexico.
Trial 31: 282-256 Trump
When you win North Carolina somehow but lose Virginia.
Trial 32: 337-201 Clinton
This is an "all your base are belong to us" for Clinton.
Trial 33: 288-250 Clinton
Trial 34: 286-252 Clinton
Wisconsin flips but Clinton still wins. I'll be discussing some weaknesses of this model after this is all said and done.
Trial 35: 338-200 Clinton
Another one where Clinton just runs the whole map just about.
Trial 36: 335-203 Clinton
Trial 37: 274-264 Clinton
Another weird VA/NC flip
Trial 38: 280-258 Trump
Trump wins Wisconsin and Pennsylvania, but loses Florida.
Trial 39: 285-253 Clinton
Trial 40: 315-223 Clinton
So...first, the results. They were exactly the same as my 2024 predictions in net. Except this time it was 30 for Clinton, 10 for Trump. 75% chance of a Clinton win.
This model also NEVER got the right result. Not even close. Why? Because of the trend model. THere are two models I use to look at elections. One is a model of a "wave", in which the states are related and if one state goes a certain way, other similar states will too. I actually find this to be the most accurate in net. My 56% chance for Clinton prediction in 2016? That was wave. And it turned out accurate. Trump won almost all of the states he was extended, AND Michigan, Wisconsin, and Pennsylvania.
However, this other model, one where I treat each state as its own independent trial, has advantages. Wave models rarely predict the results PERFECTLY. There is some randomness involved, and randomness can defy trends. This wave model did predict some of the things that happened. There was weakness in the rust belt. Clinton had a 32% chance of losing PA, and a 20% chance of losing MI. And those came up quite often for Trump. Wisconsin was harder, at a 5% chance for Trump to flip it, but despite that, it did come up once in a while.
However, this model, using complete randomness, if anything, underestimated the underdog's chances. Why? Because while a random model like this might find a state or two flipping, on the whole it's hard to flip electoral outcomes through random chance like this. In no prediction did all three rust belt states Clinton was supposed to win but didn't go for Trump. 1-2 of them often did, but not all 3. And because of that, it was harder for Trump to win in this model.
Heck, i'll tell you exactly why Biden (30% chance) and Trump (46% chance) both ONLY won in 25% of my models here. It's because in 2024, a lot of the states dont have strong polling either way. A lot of them are +2 or +3 for trump, and Biden CAN flip them. +2-3 only means around 69-77% chance for that person to win it. Flips can happen fairly often.
meanwhile, while the trend had a much softer win for clinton, with New Hampshire being the deciding state, all Clinton needed was one flip to her side to win. And because Trump needed several consecutive flips in the 80%+ range to win a lot of the time, he couldn't get them. I mean, you can win ONE OR TWO states like that statistically. But...theoretically, so could clinton.
So you would see some models where yes, Trump wins Pennsylvania, or Michigan, or Virginia, or even Minnesota or Wisconsin. But then all Clinton needs to do is flip Florida, or North Carolina, or Iowa, or Ohio, or Arizona.
I mean, when you really think about it, while Clinton only had a 56% model by trend, a lot of pro Trump states like Florida, Nevada, and North Carolina were WELL within striking distance. And Clinton could just take a couple of them and win. Florida was 29 electoral votes. North Carolina 15.
Meanwhile after New Hampshire, what did Trump have? The next was Pennsylvania. A 68-32 Clinton state. And Trump COULD take that. But losing Florida or North Carolina would neutralize that. And beyond that, Trump is just facing an increasingly difficult statistical wall. And while a few states are within range, sure...uh....Clinton can just take more Trump states the other way. Colorado was in the Clinton camp as firmly as Iowa was for Trump. Ohio and Michigan were statistical opposites. Georgia and Virginia weren't much different statistically. South Carolina was right there with Wisconsin.
So, let's face it, this model, if anything, underestimated Trump's chances. Because in practice, politics isn't all unrelated, and each state ISN'T an unrelated random trial. Each state is somewhat related to the others. The national environment can determine the direction of the election as a whole in terms of trends. As can states that are similar in nature. Nevada, Georgia, and Arizona all seemed connected in 2020. Wisconsin, Pennsylvania, and Michigan all shared a similar fate. Now, are they always all going to go the same way? no, if the results are really razor thin, you could see one go one way and another go the other. But they are related.
This model did not treat them as such. I mean, I could roll a 10 on Wisconsin, making it go blue. A 57 on Michigan making it go blue. And then on PA I'd get a 26 and it would go red. I could then roll a 85 on Ohio making it go blue.
I mean, we saw some stuff like that. Virginia goes red and NC goes blue. Ohio goes blue and PA goes red. In practice, that's nonsensical. That's never gonna happen. I mean, we even had one trial where Clinton won both congressional districts in Maine and then lost Maine. Because I treated the two as separate. If Clinton won both congressional districts, she definitely won the state. But because of the weakness of this model, Trump managed to win the state.
Both models have value though. I mean, I think it's clear that these models show what CAN happen. And they did show things like the rust belt flipping. And they also showed some of Clinton's sun belt strategy actually working and states like Arizona, North Carolina, and Georgia flipping blue. It can happen. Statistically, all of these things are possible.
So in a sense, this exercise shows a rough range of results that are possible. It doesn't necessarily get it completely right, but it shows things that CAN happen through statistical flukes and the like.
And honestly? I think it does offer an important lesson. Anyone who actually thinks that Clinton's win was inevitable on election day in 2016 was either reading old polls, was flat out delusional, or a flat out propagandist. I know Maddow and the like on MSNBC were like "here's how Hillary's gonna win" and their prediction was basically...trial 35 or something. It COULD happen. But there was no reason to think it would. The actual results to be predicted were close, I had a 272-266 Clinton win personally. That was the most likely statistical outcome.I dont think anyone could've predicted how much the rust belt would shift red, but let's be honest, it wasn't impossible. Even if this model makes all 3 states that flipped red not all flipping red at the same time.
So yeah. Interesting results.
All things considered, this should give you some ideas of how 2024 CAN go. In my previous prediction, I had Biden winning a few times as an underdog. Could a trend exacerbate those models in his favor? Sure. I dont think that any of the predictions this model spits out will be 100% right. But at least one of them is bound to be close. And yes, Trump is predicted to win. But anything can happen. Just like anything could happen and did happen in 2016.
No comments:
Post a Comment