So, FiftyPlusOne (previously FiveThirtyEight) released their midterm forecast and it's looking....pretty reasonable. I largely concur with it in broad ways, although I may differ on details. They give the democrats a 56% chance of winning the senate, a mere point off from my 55% forecast (and theirs was 55% earlier). On actual senate predictions, we vary, they tend to think Michigan has a much higher chance of going blue, whereas they seem to think Nebraska is more likely to go red. Their probability for each race seems far more....likely than my forecast. I tend to be more likely to reach extreme outcomes more easily (like 97% certainty, whereas they'll be at 85%), but these are all methodology issues.
As discussed in 2024, my model is more simple. It's polling only (outside of placeholder data where I estimate the outcome based on previous cycles), and merely represents a snapshot of where the race IS, not where it may be several months from now. Meanwhile, 538/50+1 likes to do things like do long term forecasts based on fundamentals. I guess if you do a fundamentals model, yeah, youll be more likely to make michigan blue and nebraska red. If anything, that might even be more correct. I just let my model model. Sometimes the data picks up on things that seem counterintuitive, but end up happening, like the blue state weakness in 2024 in states like new jersey, illinois, and new york. So who knows? And I guess you gotta temper certainty when the election is months out and technically anything can happen. Im fine with adjusting things on the fly, but yeah these guys just do things differently.
My own house model is currently more fundamentals based given the sheer lack of polling. They come up with 231 house seats and an 87% chance for democrats. Again, willing to split the difference is this is a long term forecast and not a snapshot. My current model is like 233 with a 97% chance. It was 231 with a 96% chance this morning but RCP updated their GCV polling average and plugging it into my model boosts dems chances.
I will say that with the lower probability, they account for a higher range of outcomes than mine. In the senate, they seem to think that democrats will get between 47-55 seats, with their median outcome at 51. My own model agrees with 51 being the median (technically 51-52 due to iowa being a tossup), but and my range of outcomes is 46-54. Not a huge difference but it is weird seeing up to 55. Do they really think states like south carolina, montana, or florida can flip? Well, they have a higher probability than my model so I guess they do.
THe house is where this really gets funky. They seem to have outcomes ranging from like 211 on the low end for democrats to around 253 on the high end. My own model has a 95% confidence interval between 220-244 currently. Even if we account for the sudden 1 point shift we're still talking 217-244. This isnt to say outliers outside of these ranges cant happen, but I tend to try to stick to around 95-96% confidence (two tailed). Still looking at their charts, it's the same. If we include extreme outliers they got senate outcomes with 58 dem senate seats and 280something house seats. No way that's gonna happen. My model doesnt even account for extreme possibilities like that, since once we get beyond 99.8-9% or so I'm just like "let's just say it's 100%."
Again, interesting to see the methodology differences here. I'll stick to my own model out of pride but yeah. We'll see. I find my model tracks pretty well in practice. Aint perfect, but i aint expecting it to be. Again, it's the AK47 of election models. Simple, easily accessible, it just works. Might not be the most accurate or sophisticated in the world but it gets the job done. Sophisticated models are a double edged sword, the additional complexity allows more accuracy in theory, but given the raw noise of probability, it can backfire. We'll see who is more accurate on election day. At least it's better than their 2024 "50% no matter what" forecast they seemed to have.
Speaking of 2024, let's look at their final forecast.
So, they had it 50-50, and the deciding factor was polling error. That's about where I was. Funny how I can have a simple model, they have a complex one, and they end up in the same place more or less. Goes to show how even the best models are gonna have error and inaccuracies, same as mine. On the state level they seemed to do what my model did, with MI and WI leaning left, but they also though PA leaned slightly left too. So they technically had harris winning. I know it's being really nitpicky, given a 0.6% difference between theirs and my model, but mine was actually ever so slightly more accurate in that regard (although I was personally even more bullish toward harris where i thought the error would go in her favor). Still, they knew that polling was gonna be inaccurate, and estimated it being off by 3.8 points in either direction. And yeah....that would lead to either 226-312 as we saw.....or 319-219, depending the direction it went in.
In the house, they came to roughly the same conclusion I did. Slight dem advantage but virtually a dead heat. In the senate they gave the republicans a 92% shot whereas my model was at 87%....and yeah. Very minor differences. Again. Really goes to show how my model punches above its weight. Only slight differences from theirs as of election day.
So interesting model, let's see how ours compare in 2026.
No comments:
Post a Comment