So...it's been an annoying trend that I just didn't get. Why did my model keep favoring democrats even though the race is 50-50 or even favoring republicans slightly? When the race was 50-50, Harris would win 55-60% of the simulations in my basic model. When it was 45-55 in Trump's favor, it dropped closer to 50-50 with Harris still coming slightly ahead. Now with the race down to 43-57 in Trump's favor, the simulation is FINALLY starting to favor the republicans a bit. But why? Well, remember how I mentioned working on a massive election simulator to use on election day? I'm still adding metrics to it to measure, but I finally got it working, and what I cobbled together gives me some insight into why. So if I may infodump a bunch of stats from my outcome dashboard a bit...
Yikes. So that's a lot of data to take in. I'll walk you through it.
This version of my simulator takes my simple simulator, and scales it up to 1000 outcomes AT ONCE. This gives me more data to analyze, allowing me to study trends within my model that are impossible to do with the one at a time version.
As we can see, Trump is finally winning in it. In this particular batch, he won 515/1000 outcomes, while Harris won 471. Scaled down to a normal sample of 100, Trump won 52, Harris won 47, and 1 was a tie. Basically that's what this means.
Normally, I'd expect Trump to win 569, Harris to win 431. If it followed the trend model perfectly, that's what would happen. But obviously, like always, the model has the dems slightly overperforming.
Well, lets look at those percentiles, and we start to see a picture emerge. The fact is, the democrats simply have a higher ceiling, and the republicans a lower floor. And while the simulator generally moderates outcomes because each state is treated as its own trial and if one state goes red that should go blue, often another state that is supposed to go red will go blue to compensate. So the flips will cancel each other out somewhat. As we can see, only like 4% of outcomes are actually 319-219 Harris or 226-312 Trump or better. Basically anything like that is outside of the realm of statistical significance. But we do see that Harris generally gets outcomes like that more. Again, why?
Well, by this point, I have a hypothesis. It's been my hypothesis all along and the data seems to be kind of starting to correlate with it. To put it simply, "Blorida" and "Blexas" (blue florida, blue texas). Those states are MASSIVE and if they flip, they can save the democrats from an otherwise certain defeat. Following the trend model, if the dems can't turn tossup and now mild lean red sun belt states and rust belt states, they don't have a hope of turning texas and florida. But this simulator treats each state as its own outcome, with its own random number being assigned to it. And given Florida is expected to flip 6% of the time and Texas 8% of the time, those states can single handedly save the outcome for the democrats. Take this particular outcome from my one at a time model w/ map.
So...this is a weird one. A REALLY weird one. The democrats lose the rust belt. They lose Nevada. They even lose Oregon. They even lose NEW MEXICO! I mean this is a BAD map for democrats. But...they win the sun belt. They win North Carolina, and Georgia, and Arizona....but that only gets them to 256 electoral votes. That puts them over the edge? TEXAS! (also ignore white alaska, that was a bug in the simulator i fixed after taking the screenshot).
That said, go back to my chart above from my new mass simulator. I also developed some stats to measure what state tips us over the edge to 270 in it. For republicans and democrats its different since they start from opposite ends of the electoral college. For them, with the current map, North Carolina, Michigan, Arizona, and Georgia are all big tipping points for dems. PA isn't because it's actually one of the most blue (or alternatively least pink) states. But if you go down, Texas and Florida both save the dems about a collective 6% of the time. This ins't even considering Iowa or Ohio as possibly being a factor either.
Meanwhile, go back to the republican side. For them. They need Pennsylvania, Georgia, and Wisconsin as their common tipping points. But beyond the big 7...there ins't much that saves them. Maybe Minnesota or Oregon, but they don't tend to deliver victory as often, probably because they're smaller.
Take this map for instance where we get red Minnesota but it doesn't tip the scales.
If this were Florida, the GOP would win. 30 electoral votes. 271 for the republican, 267 for the democrats, boom (and yes, I know this simulator doesn't show it the same way the big one does, it's not important for the sake of this discussion). Texas, the GOP would have 281 votes. The GOP would win this outcome. But because Minnesota is only 10 votes, it's not enough to compensate for winning the rest of the rust belt and parts of the sun belt.
And that's where the asymmetry seems to be in my model. Because each state is treated separately as its own trial, Texas and Florida are able to flip the outcomes for the dems about 6% of the time. And factoring in other red states like Ohio and Iowa, we're talking the dems being able to flip outcomes from red states about 8-9% of the time.
On the flip side, republicans only are able to flip outcomes about 5% of the time across 5 different swing states. So, the model is inherently gonna cause the democrats to overperform about 4%, at least in this subset of models. And this is it literally takes almost a half point handicap for the republicans before they start to break even and win in my model. It's just the power of Texas and Florida. And while I wouldnt seriously expect them to flip in a scenario where they become a tipping point, in reality we'd be going into democratic landslide territory before they flip...stranger things have happened. Remember Wisconsin randomly flipping in 2016 despite having a 5% shot? Yeah. So...individual states can have weird overperformances like that at times. And in the above model where there's a general rust belt underperformance, but sun belt overperformance...I could kinda see an outcome like the above happen. It would be weird. It wouldnt be my first guess, but i guess it isn't impossible.
So yeah. I just wanted to discuss that as it's something that's been vexing me ever since I built the thing.
No comments:
Post a Comment