Loading summary
Transcript122 lines
- [00:00]
A
I think usually this audience will be more used to hearing about differences in forecasting as like, coming from people who hate each other as opposed to people who are friends with each other.
- [00:12]
B
What, like me and Elliot?
- [00:15]
A
And so I think it will be valuable in this pluralistic environment to be like, we're both forecasters, we're both friends, and we also disagree, and that's okay. And that, like forecasting is not divined by God, forecasting is a way of modeling your understanding of an election.
- [00:34]
B
And, and I've said this many times, I also think there are areas of Zach's forecast where they're just better than us.
- [00:51]
A
Hello and welcome to the GD Politics podcast. I'm Galen Drouke. We're four and a half months out from election day 2026, and forecast models are beginning to come online. What are the chances each chamber of Congress will flip? And what's the chance of continued unified control for Republicans or newly unified congressional control for Democrats? The stakes Will Democrats gain the ability to investigate the Trump administration and muck up any legislative goals? Will Republicans be be able to fill any Supreme Court vacancies until January of 2029? The answers to these questions will define the final two years of Trump's presidency. And today I'm going to dig into their likelihood of happening with two, I think it's fair to say, up and coming election forecasters. Here with me today is Laksha Jain, head of Political Data at the Argument and CEO of Split Ticket. Welcome back to the podcast, Laksha.
- [01:43]
B
Hey, thanks for having me.
- [01:44]
A
Also here with us today for the first time is Zach Danini, head of Data Science at Vote Hub. Welcome, Zach.
- [01:51]
C
Thank you very much for having me on. Looking forward to it.
- [01:53]
A
I'm excited to have you here. We love first timers. I should also say before we get going that Vote Hub's election forecast is currently live, so folks can go check that out. The forecast from Split Ticket and the argument, which is Laksh's forecast, is not live yet, which means that GD Politics podcast listeners are getting a sneak preview of an unpublished forecast. Don't ever say I didn't do anything for you. And Laksha, thank you so much for coming and giving us a sneak preview before it's actually live. I think you guys are planning for sometime midweek, right?
- [02:26]
B
That's correct. We are. We're writing up the release post right now. We're really excited. In the process of making it, I have been consistently surprised at some of the things the model outputted at first before looking in and being like, wait, that actually makes sense. Why are people talking about this race so differently? So it's always kind of fun. I think Nate had said Galen, back when you guys were at FiveThirtyEight, he was like, sometimes the model just surprise me. And then you look into it and it's like, oh, that actually makes a lot of sense.
- [02:52]
A
Well, and it makes you sort of challenge yourself and it makes you go on the record in a way that many pundits never go on the record. And so I think this conversation is going to be very useful. I should also say your forecasts disagree significantly in some key places. And so today we're going to have the opportunity to. To hear different theories of how to model an election. I think there are even places where you guys think that the other person's forecast is better. And so far from the usual, which is like different elections forecasters going to war with each other, I think we're going to hear some disagreement that is also very civil. Is that right? You guys are friends in real life? No.
- [03:36]
B
Yeah, we are.
- [03:37]
C
I mean, so we, we do podcasts together every week.
- [03:40]
A
So yeah, I think I. I think that counts as friendship. Okay, so before we go any further, I just want to give you guys the chance to introduce and talk about your background in election forecasting. Zachary, as the first timer, why don't you go first?
- [03:55]
C
Awesome. Yeah. I'm Zachary Danini. I'm head of data science at vote hub since 2025. Before that, you might know me. I worked at Decision Desk HQ for almost four years and we do a lot of election night coverage at Vote Hub. You might know us for our precinct maps and we launched our midterms forecast about a month ago now.
- [04:15]
A
And this is the first ever election forecast that you're fully helming. You were involved in the forecast at decision desk HQ for 2024 and 2022.
- [04:26]
C
And then of course, the Vote Hubs forecast is a team effort. Right? You know, there's a ton of other people who have put so much, you know, work into this.
- [04:34]
A
Congratulations. We're excited to dig into it. Laksha, you have created a forecast for Split Ticket which is also being used by the argument. A fellow subst folks have heard your voice before, but tell us a little bit about your background in election forecasting.
- [04:50]
B
We made Split Ticket about five years ago. We've been basically become known for making models of all kinds. Folks may know about our war model, which is wins above replacement for candidate quality, but we also forecast elections a lot in 24. We had one of the most accurate forecasts in the nation. We only got eight House seats out of 435 wrong, which was something we were really proud of. Our whole principle is basically just we don't really trust our punditry very much. So we make models to check them.
- [05:17]
A
I want to set the table a little bit, since we're diving back into the world of election forecasts after, I don't know, a year and a half away from them. What do you view as the goal of making a forecast model? Like, why do it in the first place?
- [05:36]
B
I think without a forecast model, I think it was Brit Hume who spoke about how they used to forecast elections before polls as media people. Right. And he said, we would go to rallies, we would see where the energy was high, we would see the ground game. We would talk to a couple of people coming out of rallies and out of polling stations and make our inferences that way. And it strikes me that that is a phenomenally unscientific way to do this, to the point where people would get shocked by certain election results that just diverge from what anyone expected. Results that were just really, like, entirely foreseeable if you knew what to look for. Right. One example I think of here is back in 08, actually, everyone kept thinking that the election was a toss up for a long time. And Nate Silver's forecast was actually like, this is not even close. Like, what are we talking about here? And that, to me, was where the Saga begins. So the 08 election was really the first election I actually followed at all in any way. And I remember people thinking back in September, like, people telling, oh, it's going to be close. Who do you think is going to win? I was 11. I'm like, I don't know. These guys both seem cool to me, but one of them seems old, I guess, to make it short. Right. Like, the way that I see the purpose in an election model is that everyone has different opinions, but the data is the data. What does the data actually say is going to happen? The purpose of a model is to communicate that to people.
- [07:09]
A
I think that's interesting, but people are also going to learn today that you can also have different opinions about the data or which data is most important, which data to include, which to exclude. And so I think it's important to say you ultimately are schooled in the sciences, but there's also an art to creating a forecast as well that we'll get into. Zach, I'm curious if your understanding for the point of an election forecast is Similar to latches.
- [07:37]
C
When it comes down to it, the number one thing that's going to power everyone's election forecast this October is polling. Right? But what forecasts are really, really, really good at is defining uncertainty regarding polling.
- [07:50]
A
Right?
- [07:51]
C
And specifically maybe how different polling errors might correlate. So, Laksha talked about 2008. I'll talk about 2016. I'm old enough to remember in that September and October, a lot of forecasts that I don't think were correctly made gave Hillary Clinton a 95 or even 99% chance to beat Trump. But again, I'll go back to Nate Silver's 2016 forecast, right? Where. Which was 70, 30, which people like me would consider to be a very, very, very good forecast, right? Because he's able to take polling data that shows Hillary Clinton favored to win and quantify it with some level of certainty. Say polling isn't perfect. How might it miss? Might polling misses in Wisconsin, Pennsylvania, Michigan be correlated? Which in 2016, they were.
- [08:37]
A
So I'm getting a sense that we have two fivethirtyeight slash Nate Silver acolytes on the call. Is that fair to say?
- [08:45]
C
Uh, you know, I think he kind of founded this industry.
- [08:48]
B
Yeah, I'd say so.
- [08:49]
A
Zach, I do have to ask you, though, because Latcher showed his cards and told us how old he was in 2008. How old were you in 2016? Just so we can make me feel really old on this podcast.
- [09:00]
C
I was 14 in 2016. I'd like to note that I. I stayed. Yeah, no, I'm very young. I. I stayed up the whole night. In 2008, actually, when I was. Was five, I begged my parents to let me watch the 2008 election coverage.
- [09:15]
A
Nate, if you're listening, the kids are truly coming for us. Okay, I think we have filibustered quite enough here. We're going to talk about the actual forecast. So, in the House, vote Hub, Zach, that's your forecast. Gives Democrats a 72% chance of winning a majority, while the argument slash, split ticket. That's you, Laksha, gives Democrats a 90% chance of winning a majority. Now, let's talk about the Senate, which you guys are a little bit closer on. So vote Hub. Again, That's Zach. A 55% chance that Republicans maintain their majority in the chamber. Lacha, split ticket, slash the argument. You give Republicans a 53% chance of maintaining their majority in the chamber. I also, for what it's worth, have some odds from Kalshee that I will pepper in throughout our conversation. But I Think let's lead with the hard work that you all have done, and then we can get in to the prediction markets afterwards. Let's start with the house. Watch out. 90% chance. We just heard Zach mention the 2016 election when some forecasts maybe got carried away with a certain level of certainty of Hillary Clinton's win. I don't know if 90% gets us back into that area. Probably not. But I'm curious, why are those odds so high, and in particular almost 20 percentage points higher than Zach's odds at 72% chance?
- [10:44]
B
There's two reasons. Number one is, and I actually have wrestled with this a lot myself to think about, like, what are we actually comfortable with here? But ultimately, you put out what the model says and you don't try to cherry pick your priors too much. The reality here is that Democrats are up in the generic Ballot by 7 percentage points right now. In the aggregated polls in our survey that we conduct, right. We actually have likely voter screens and we do this very regularly. And the rolling average over the last three months, the Democratic likely voter advantage is at 9 percentage points, which by the way, is less than what the time Siena had it at, even among registered voters. The biggest difference between Zach and I comes in the environment that we are modeling. Zach is modeling an environment that is about D7. I am modeling an environment that is about D +9. If you took Zach's data and you put our polling average on top of it, right? Like, because that's coming from our polls directly at the argument, because we do our own surveys. If you took that, I would bet you that Zach's forecast would shoot up to like 85, 87% for Democrats. And that would be much closer to ours.
- [11:57]
C
Exactly. So right now the Vote Hub generic polling average is, I believe, D 5.4. You know, as. As Laksha said, there are reasons to think, you know, between LV and then also just generally waves tend to move towards the out party as a year goes on in midterms. So between, you know, expected LV screen adjustment, likely voter, likely voter differences, and then also just data moving towards Democrats throughout the year. We predict a D plus 7.1 environment right now. Right. So basically 1.7. Very similar from, from Laksh's, you know, 7 to 9. Ours is like 5.4 to 7.1 right now.
- [12:35]
A
Zachary. That's interesting because that also diverges from other polling averages that are out there. I mean, if you look at Silver Bulletin, for example, Democrats lead in the generic ballot by 7 percentage points. Why is vote hub at sort of 5 and a half percentage point advantage for Democrats right now?
- [12:53]
C
The generic ballot polling has been very weird recently. I think we have a streak of 12 straight, had a streak of 12 straight polls that were D 6 or worse. And I think at some point some of the differences in the generic ballot averages right now are just coming from how sticky they are in the sense that I get the feelings that Vote Hubs is a little bit less sticky because those 12 polls changed it a bit more in the Republicans direction than Silver Bulletins, I'm guessing.
- [13:23]
A
Is it fair to say that your forecasts will likely converge between now and and Election Day? It sounds to me like Lancia is putting more emphasis on where he expects the average to end up. And Zach, maybe you're putting more emphasis on where it is now. I mean, is this like a now cast versus a forecast situation or is that a misinterpretation of what I'm hearing?
- [13:49]
C
Well, I wouldn't say ours is quite a nowcast, right. In the sense our generic ballot pulling average is 5.4 and we have our predicted environment at 7.1. Right. So we're still expecting, you know, a move, I think, Laksha, you can correct me if I'm wrong. A big difference is that your general argument polling right now is just rosier for Dems than a lot of these like, you know, public polls from like, you know, Signal or YouGov or Morning Consult or something.
- [14:14]
B
I actually don't know if that's correct because among registered voters, which is what gets put into Nate Silver's averages, we're at D 6 and we've been at D 6 for the last three months. So it's not like our polling is rosy for Democrats. It's that everyone right now is polling registered voters other than Sigma Signal, which is, we should note, a very good pollster, they're also a Republican aligned pollster. So generally the assumptions that these companies tend to bake in is a little bit more conservative if you're a conservative pollster, a little bit more liberal if you're a liberal pollster. There's house effects to adjust for there. Right. But in general, we have had D plus 6 for the last three months among registered voters. It's when you put the likely voter screen on, that's when the Democratic edge goes to D plus 8 or D plus 9. I fully expect that when the election nears our forecast will align more because at that point we are looking at like basically the same thing. Everyone's polling likely voters at that point and we've been pulling likely voters all throughout as well. That's why there is a difference between the two. That's why I wouldn't quite say that like we're rosier for Dems because it's really just everyone can make a likely voter model, just that not many people do right now. And if you just looked at our registered voter numbers, we'd be in the same place as Zach.
- [15:38]
A
Well, Laksha, then let me ask. Most pollsters only switch to a likely voter model around Labor Day or after because, you know, it may take some time to get a sense of who's really tuned in, who's really enthusiastic. Yeah, I think there are, there is an argument that that's no longer the case. We just have a better sense of what the coalitions are. We know that Democrats are now the better educated, more enthusiastic, more reliable voter. And is that kind of what you're relying on or do you think that even in elections past people were too late to apply a likely voter model?
- [16:13]
B
No, not only is that what we are relying on right now, but it's the fact that this is a midterm is the reason why we're doing that. We would never do this in a presidential year. In a presidential year. I actually think there's a good argument for waiting a little bit more. However, that advantage of waiting has gone down with time across the board and we now have a much better idea, as you said, of what coalitions just are across the board, who's likely to turn out based on past voting behavior, based on self reported enthusiasm and whatnot, to the point where I think if we just went with registered voters, we would be losing information on what the forecast is actually meant to tell us, which in my opinion is not what the result would be like right now. And I know this is not what Zach's doing either. It's what would the environment be like in November. And we think we can make a pretty good model of that at the moment, especially because it's a midterm and the flaky people won't vote usually.
- [17:13]
A
Zach, how do you react to that?
- [17:15]
C
Yeah, so I, I think, yeah, so I think one big difference is that so vote Hub, we don't do any polling.
- [17:20]
A
Right.
- [17:20]
C
We use polling done by other people and we aggregate it and do math surrounding it based on, you know, their track record and you know, stuff like that. While the argument vericite is lot, Laksha is doing his own polling and he has his own likely voter screen. So for us, when we're looking at polls, we're looking at, you know, like LV polls that have come out recently often. Right. So for instance, we rate Marquette University Law School as an A pollster. They were one of the best pollsters in 2024. They have a good track record. They had a D1LV poll in late May. I'm guessing, like, even just one data point like that, because we don't get that many high quality LV polls in a couple weeks, is probably having a pretty reasonable effect on our forecast right now. Right. It's definitely, you know, an outlier poll. It's good that pulsers publish outlier polls. You know, hurting is really bad for forecasts. You know, I think Marquette's poll in late April was like D 10. Right. I think because of that, our, you know, Vote Hub's forecast is a little bit more, you know, based on the polls that that come in. I think, you know, Laksh's the argument Verisi poll has a bigger impact on his forecast than ours.
- [18:27]
B
Yeah.
- [18:28]
A
Okay, so there's some differences in terms of what polling you're relying on and how much emphasis you're putting on registered voters versus likely voters. Are there other aspects of your House forecasts that are different? And I already know the answer on one count, which is that, Zach, you include the predictions over at Kalshi in your forecast model, which strikes me as interesting, and a little bit of like the snake eating its own tail, if I can say that. And I want you to defend your decision there. But like, usually the forecast models and the polls are in many ways what inform the prediction markets. And so you get a little bit of this, like the forecasts are informing the prediction markets, but then also incorporating the prediction markets back into them. Like, how do you justify doing that for your forecast?
- [19:26]
C
So the first thing I'll say is in House races, Kalshi is going to be weighted very, very, very low right now. And the reason why is because those markets are like very low volume. So we've done a lot of research in the past into what is prediction market's track record of predicting things. Well, generally, you know, there's there's kind of a certain, you know, volume threshold where if I wanted to right now, I could go put $500 into Texas 35 and swing the Kalshi odds by a lot. And we don't want someone just putting $500 into Texas 35 to be able to touch our forecast very much. So where Kalshi odds are weighted A bit higher right now would be like, you know, Senate races where there's actually, you know, more volume, where you're getting a bigger wisdom of the crowd, because higher volume markets have been more accurate in the past.
- [20:14]
A
But I guess my question is, why do it in the first place? Why incorporate the Kalshi odds into the forecast that Vote Hub defaults to?
- [20:24]
C
Yeah, so we, we have, you know, looked at a lot of different years of prediction markets, and the reality is combining fundamentals and polls with prediction markets has been more accurate in years past. So we, we used predict it a lot. And Kalshi is obviously, you know, a different product, but it's a similar. So our view is kind of like we have our legacy model, which is like the Vote Hub model, but then if we're trying to communicate for some reason to people what we actually think the chance of things happening are we want to try to take all information. Even if I think a lot of people would say including expert ratings and Kalshi are like cheating because it's other people's work. Right. But we just kind of want to create a model that's as accurate as possible. And based on backtesting, we expect our full model to be more accurate than our legacy model to that point.
- [21:11]
A
I've had these conversations with Nate during our model talk episodes where we talked about why include house ratings from Cook Political Report, for example. And he had the same answer, which is ultimately, although this forecast itself may influence some of the race raters, doing so gets you a more accurate forecast. And so I respect that. I do have a question here though, which is using Kalsheet instead of predict it, even though maybe you looked. Your research focused on predict it. There is. There was a research paper out of Vanderbilt that suggested that predict it is the most accurate of the prediction houses or whatever, then Kalsheet, then polymarket. And so I guess, like, how do you navigate which ones you use? Because there is divergence.
- [22:00]
C
Yeah, it's really. It's really hard. Right. I think.
- [22:03]
B
So.
- [22:03]
C
Something that we really like about Kalshi is when what we're putting in our model just isn't the number you see on Kalshi. Because that wouldn't make sense. What we can do is we can pull the order book data and the activity, and the combination of being able to look at that granular data with the fact that Kalshi is higher volume, we think lets us create a number to put into our forecast from Kalshi that is very, very predictive. Based on our testing and moves in a way we really like compared to trades. So Kalshi, bigger volume, more markets basically. And then it has some downsides in our view, which is that there aren't as many limits, right? So one person can influence markets more and we try to come up with an algorithm to stop that from happening.
- [22:49]
A
Laksha, I assume you're not using the prediction markets to influence your forecast.
- [22:56]
B
We're not. The thing is that for me, I understand and I respect the reasons to use that. To me, here's the reason why expert race ratings differ from predicted or cauchy or whatever. Experts see a ton of internal polling, like a ton, right? So for the House, for example, there is a serious and substantial benefit to using like the Cook Political Report or whatever when it comes to predicting House races. And the reason for that is Cook sees internals from both sides and they know how to calibrate that and shift appropriately in the rating. And basically if you ask what's in a rating in the Cook Political Reports case, it is basically a ton of internal polling is in that rating. Kalshi to me doesn't quite have that. The people placing bets on Kalshi are making informed guesses. And there is value in that, to be clear, like there is value in knowing what happens there. Because let's just say that like a massive scandal happens tomorrow and polls can't price that in properly. But market odds crash. No poll is going to tell you that if you don't have enough time for polling to come in, right? That's the time at which like market odds are really useful. So market odds pick up some alpha. But I'm not convinced that that alpha is high enough to warrant being placed into a model. When you consider the counterpoint, which is that markets are influenced by models, models are influenced by markets. Not in the same way of the Cook Political Report, which generally has its own ratings and it doesn't get influenced much by like FiveThirtyEight or whatever.
- [24:39]
A
Okay, so let me give you an example of where you provide different value in your different forecasts. If there is a very late breaking October surprise, say just say for example that the director of the FBI reopens an investigation into one of the presidential candidates just days before election day. I'm sure that this would never happen, but let's just say that it did. If the odds on Cowshi shifted significantly as a result, then Zach's model would move accordingly more quickly than Lacha, your model would, is that correct?
- [25:16]
B
That's right. That's right, yeah.
- [25:18]
A
However, who knows like that is up for historians to debate whether or not that really changed the outcome of that election or what have you. Because of course, I am talking about 2016 here for any listeners who are young enough that they didn't know that reference. But there's also the possibility that Kelshi significantly overreacts. I mean, we've certainly seen this before with like the. I mean, just look at the Texas Senate primaries, like, all of a sudden, after John Cornyn slightly overperforms his polls in the first round, like you have, the predict market's going to like an 80% chance that he's going to be the nominee. And of course, we all know that he's not at this point. And so is it fair to say maybe that. Zach, your forecast is liable to be more volatile than Laksha's. Yours? Yeah, both in combination of the prediction markets and responding more quickly to new polling coming in, as you suggested earlier.
- [26:14]
C
So I think that is true. So I do want to add another, you know, theoretical benefit of prediction markets. So something that we've always struggled with or haven't done in the industry of forecasting elections is early voting data. And early voting data, I think a lot of people see as potentially you need to analyze it in a smart way, because a lot of people don't analyze it in a smart way and it can be very misleading. But what we saw in Arizona and Nevada in 2024, early voting data was extremely powerful, which was that in early voting in Arizona and Nevada, low propensity Republicans were turning out more than low propensity Democrats. High propensity Democrats were turning out more than high propensity Republicans. And basically, for those who don't get into the early voting data. Tea leaves. Right. When your low propensity voters are showing signs of turning out better, even when the overall early electorate is blue, that was a really good sign for Donald Trump. Right. So markets reacted very well to that in the same way that my internal priors did in October 2024, especially Arizona. By the time I was going into election night, the forecast I had helped create at DDHQ had the race at tilt R, lean R, kind of depending on what your ranges are. But our priors were probably a 90% chance Trump won Arizona. We just didn't really see a path for Harris in the last week before Election Day. And prediction markets didn't move quite as strong as our priors. We had, like, very strong priors from that early voting data about Arizona, but they still moved a lot in the right direction. And we didn't really have a good way to incorporate that data into our forecast otherwise.
- [27:52]
B
I actually want to echo that and say this is one of those wonky things that I think people won't necessarily think about too much right now. But one of the best signs for how a party is going to do in the election night itself and like, you know, when the results come in is you look to see how their low propensity voters are turning out in the voter file. Now, this is not perfect because you cannot model persuasion using turnout data. Right. But it does give you an indication to see, like, okay, who is likely to beat their electorate. New York City was a really good example in the mayoral race that Zoran Mamdani was actually likely to win that race. Prediction markets capture that. Poll polls did not. One of the things we saw was that Mamdani's electorate was just way younger and way bluer than what any poll was capturing. This is the type of thing that early vote tells you. Prediction markets can pick up on that. They can't necessarily do it as well, but they can pick up on that a little bit just because people are watching them.
- [28:57]
A
Okay, this is where me as like the old 35 year old on the call has to step in and say, that sounds all well and good until coalitions actually begin to change. Because during your careers we have been in a paradigm that has been set by the 2016 election. I mean, we were headed there before 2016, but has been set in many ways by the 2016 election. And Donald Trump has been a feature in basically all of the elections that you have watched and forecasted and what have you. But for example, if you were paying a lot of attention to low propensity voters in the 2016 election, you would find that a lot of low propensity Obama voters were showing up and voting. And if you assumed that, oh, that's a good thing for the Democrats, you would have been completely wrong because those people voted for Donald Trump. And so I do worry here and maybe the. Because ultimately we're talking about whether or not to incorporate the prediction markets into a forecast. So maybe the prediction markets are smart enough to figure that out. But I do sort of, when we get into this, this kind of deciphering the tea leaves of early vote, I worry about us like all over, relying on current dynamics and not being prepared for the next big change that could be coming in 2028 or beyond.
- [30:10]
B
It's a very valid concern. And one thing I would say is that in 2016, in the polls, we actually saw that Trump was gaining a lot with those types of voters. You cannot look at early vote data just based on historical partisanship allegiance. You have to look at it in conjunction with what you see today and combine it with the polling. Right? Like if, for example, in 24, we saw a lot of low propensity voters turning out. Right. What you actually saw was that those low propensity voters were turning out. If you only looked at historical data, you would think that Democrats were doing really, really well. You would actually think that like Arizona would be pretty close. In fact, if you only looked at historical data, however, if you knew, like we did from the polls that Kamala Harris is losing a lot of points with low propensity voters, then you'd figure out, hey, this, this is actually not really good for Democrats. This is all getting a little bit too abstract at this point. But the point is just that you can't really replace human intelligence as easily as people think. But here's why I don't incorporate prediction markets. Let me talk you through the Virginia Attorney General race.
- [31:23]
C
Okay?
- [31:23]
B
Jay Jones vs Jason Miaris in that race, Miariz went ahead in the prediction markets after that scandal, J. Jones won by six. A fundamental model would have had J. Jones winning the entire time. People very badly mispriced the impact of that scandal. Just as these markets can be good in predicting and figuring out impacts like massive scandals. Like let's say tomorrow if a scandal comes out where a Democratic nominee sexually assaults someone. Right. Like, or Republican nominees convicted of like child abuse or whatever. Right. We're not going to have that in our model. Prediction markets would. So there are advantages and disadvantages to having it. I settle on the front of don't have it, Let the polls pick it up. Take the polls into your data and let that come with what may. That's just the way I see it. I see other arguments for having it differently and I respect them just to
- [32:19]
C
get a little bit more hard data. So I agree with what Laksha says that there are some races where you look at it and you say, ooh, the prediction markets handled that race badly. But from 2016 to 2022, if you just look at polls in statewide races versus prediction markets, we find that you can get about 0.4% closer using a prediction market based model instead of a poll based model. I'm not saying, oh wow, prediction markets are obviously way better because of this. It's a small difference error by model. But we do, like prediction markets, have a reasonably promising track record. In a sample of about 70 to 80 recent statewide races where they very slightly outperform the polls in creating a margin based model. So it was enough for us to say at least of the two models on our site. We'll offer one with it.
- [33:12]
A
Okay, well 2026 will be another data point in determining how well these prediction markets are doing and we can all meet up back here after the talk about it. Just to put a button on that for the House again, vote hub 72% chance for Dems split ticket slash the argument 90% chance for Dems and just to say it, Kelshi 78% chance for Dems so a bit of a choose your own adventure here. GD Politics is powered by you, the listeners. If you enjoy the show, the data driven analysis, the genuine curiosity and lack of partisan bs, and yes, sometimes the silliness too, please consider becoming a paid subscriber@gdpolitics.com paid subscribers help make the podcast possible, and they also get twice the number of episodes and access to the paid subscriber chat where you can send me whatever questions you've got. Independent political media only works if the people who value it support it. So if GD Politics is part of your week, head to GDPolitics.com and become a paid subscriber. That's GDPolitics.com I so appreciate it.
- [34:18]
C
Thank you.
- [34:20]
A
Let's talk about the Senate where everyone is a whole lot closer. So vote hub, 55% chance that Republicans retain control. The argument slash split ticket 53% chance that Republicans retain control. And then over at KALSHEE It's a 56% chance that Republicans retain control. Like I mentioned though, there is a lot of difference between your two forecasts. When you look at Underneath the Hood. I'm going to cite just a couple so that folks can get excited for what we're going to talk about before we really dig into it all. So if you look at Georgia, Vote Hub, 87% chance that Ossoff wins that race, keeps his seat in the Senate. You look at split ticket, a 98% chance that Ossoff wins his race. Laksha. We're going to make you get into that. Okay, but first before before we, before we do Maine, also some some distinctions there. Vote Hub 62% chance for Graham Platner. Over at Split Ticket, 74% chance for Graham Platner. Let me mention one other one. I mentioned a couple areas where split ticket has the Democrat doing better than the Republican. But if you look at Iowa, for example, Vote hub has a 69% chance for Ashley Hinson against Democrat Josh turek. Split ticket, 81% chance for Ashley Hinson against Josh Turek. So is part of the distinction that Laksha, you're relying more on just the partisanship of the state, the fundamentals, so to speak?
- [35:52]
B
Yes and no. We are relying a lot on the partisanship of the state. But one of the big things with Georgia, I'm going to start with Georgia because that is a really bold call, right? Like saying Ossoff at 98% and we're
- [36:04]
A
in Hillary in 2016 sort of territory there.
- [36:07]
B
I, I completely stand by that. I, I just want people to understand two things. This is a state that was Biden plus 0.5. It is Trump plus 2. Okay. This is a state that is not significantly more Republican than the nation as a whole. It is also a state where the Democrat is running as an incumbent. You do not therefore expect like a significant divergence between like, oh, it's way red or down ballot? No, not really. Because Ossoff is also an incumbent. He is also a pretty good candidate. Like we saw that in 2020 with Purdue. So the combination of those things suggests to us that like, Ossoff should win reelection pretty handily. I am genuinely in this point. I am in like Obama, McCain territory for that race. Like, I do not see why everyone is saying that ra race is going to be close. I don't think it will be. And none of the data that I find suggests that none of the statewide polling data, none of our polling data, none of the fundamentals suggest this is going to be close.
- [37:14]
A
What's the polling average there?
- [37:16]
B
Our polls point to a baseline D +6 environment before you even incorporate candidate quality and incumbency in. I think the polling average right now is kind of in flux because we don't know who's going to win between Collins and Dooley, but both of them would lose handily. I think it's just Collins probably does two or three points worse than Dooley, but I think Dooley loses by 10 or by 7 and Collins loses by 10, I think right now. Right, Zach?
- [37:42]
C
Yeah. So that, that sounds very, very, very solid. Seven and ten. I think I, I want to be clear that I'm not one of the people arguing that. I think this is like a competitive race. Right. So Our forecast is D 7.5 for Ossoff right now. And as Laksha said, we don't know the nominee yet, and we think there'd be a pretty big difference in how strong they are. And then we have ossoff at an 87% chance to win.
- [38:05]
A
Right.
- [38:05]
C
That's still like very high. I would say the one key thing that pulls us down a bit on Ossoff compared to the arguments forecast is the fact that we believe more competitive races in 2024 shifted left less towards Republicans, and in 2026 they will shift less towards Democrats. So there's this idea right, where we're going from a pretty good year for Republicans to what we expect to be a quite blue year for Democrats. So we maybe think the nation as a whole will shift around 9 points towards Democrats. But our math has Georgia only shifting 6 points towards Democrats because we believe the 2024 presidential election dynamics had Harris outrun expectations in the seven competitive presidential states. Like, we did some analysis on kind of neighboring precincts across state lines and we found that, that we believe the competitive state dynamic in 2024 influenced Georgia to be about 2 points bluer than just what you'd expect based on the baseline partisanship. So a way to think about it. Right. Is I think Trump won Georgia by two and Texas by 14. Right. I think that's about right. We actually think Texas is only about 10 points more red than Georgia, not 12. Even in 2024, before thinking about changes and the turnout dynamics basically buffed Harris in these key states.
- [39:31]
A
Yeah, so that was actually one of the arguments in the 2024 Post Game Analysis was that because the races were significantly closer or like Harris did better in the seven competitive states than she did nationwide, that actually the Harris campaign was effective in part. And so you're saying like, because Harris and Trump didn't vigorously contest Texas, we don't know what Texas really would have looked in 2024 compared to Georgia.
- [40:01]
C
Exactly. If it was competitive. And I think, you know, to play devil's advocate, you could see, Zachary, this is overfitting. We just saw this in 2024. We haven't really seen this dynamic in almost any other recent election where you can argue that there was this much of a difference between a state if it was competitive and then a hypothetical where it wasn't.
- [40:21]
B
So that's kind of why we did it that way though. Right. And we have state specific data that shows it. And in our state specific data, like Georgia is genuinely also swinging to the left. Like it's. If you. So if Zach's theory was correct, which by the way, I have a lot of time for it because I think there is reason to believe this was the case in like Wisconsin. And we are, we are seeing in some places that it's true. But look, if Zach's theory was fully correct, Georgia wouldn't be swinging to the left in statewide polling by as much as it is right now. If Zach's theory was true right then Democrats would be gaining way, way, way, way more in California than the polling currently suggests that they are. I think forecasting trends is really difficult to do and I think it's very difficult to forecast a latent partisanship of a state where people aren't voting. I'm not saying you can't do it. I am saying that that it comes with a significant risk of overfitting. Let's talk about Texas. Okay so Texas is a place where we have Democrats gaining 13 percentage points from 24. Wisconsin is a place where we have them only gave gaining seven. So Wisconsin we have as D plus six and Texas we have is R plus one. That's just like what it was in 2020. Right? Now you look at Maine. Maine didn't swing very much against the Democratic Party in 2024. However, we still have Democrats gaining a lot in areas like Maine.
- [41:53]
A
And you're saying that that's just because that's what the data currently shows and you're not relying on readjusting for 2024 Dynamics?
- [42:02]
B
That's right. Like I just think the trends of states are super unpredictable. I'll give you another example. Pennsylvania has one of the biggest Democratic swings in our entire data set. Said Pennsylvania barely swung in 24. It barely swung. It swung two and a half percentage points.
- [42:18]
C
Yeah.
- [42:19]
B
If you try to forecast the trends of the state you will often be wrong. More often than not is what our back testing showed. That's why we didn't do it. But I respect when people do it because maybe they're just smarter than I am.
- [42:29]
C
I, I couldn't do it so just on like Pennsylvania. I think no Senate race in Shapiro gov just could add to the vibes of Pennsylvania shifting way more blue in 26 and even in congressional races. I think the counterargument I would have is the kind of state specific polling is all well and good if it works but it is just kind of hard to poll at the regional level get enough sample size and some MRPs, some models with post stratification where you're pulling a lot of different people and then trying to very accurately pull what you think subgeographies will look like based on that it has a solid track record in the past. And I think it would be really cool if the argument Veracite's state level estimates were completely accurate in 2026. I think at that point I'd be like, wow, that's really cool. This is truly state of the art.
- [43:27]
A
We're getting a little bit nerdy for a second, so I want to make sure we're explaining what we're talking about to our uninitiated listeners, which you're getting at another methodological difference. Laksha, you include MrP. Zachary, you do not. Is that correct?
- [43:44]
C
Yes, that's right.
- [43:46]
A
Lecture. Since you're the one doing it, can you explain what MRP is and why you're doing it?
- [43:51]
B
Yeah, I'm actually gonna not bother too much with the technical terminology and I'll explain what the model does to readers, because I found that if I start explaining technical terminology, everyone loses me. Here's what we are doing. And it's not exactly classic MRP for the stats nerds, but whatever. We have a lot of data. We have about 11,000 responses in our data set at this point. We have a lot of data on how voters all across the nation are behaving. Okay. Every single poll, we get like about anywhere between 1500 to 3000 responses, and we have data on where they responded from.
- [44:26]
A
And this is polling that you're doing with Verisite. So it's original polling of the idea that you're relying on some of your own original data.
- [44:34]
B
Right. And this is what the Times Sienna does too, in a lot of ways. Right. To be clear, like the Time Sienna uses their own polling data to inform their models and so forth. So we're doing that here too. We basically collect data at a demographic level by region, state and nationality, obviously. Right. So the US Than the Deep south than Texas, for example, we have data on how demographics behave per region, per state, and nationally.
- [45:05]
A
So you can say just to. Just to really flesh this out. So you can say, okay, this is how Latino voters are moving in South Florida, or I guess maybe Florida generally. Are you breaking it down even further than state, or are you not state. So you're leaving.
- [45:21]
B
Yeah.
- [45:21]
A
Which we've seen in the past. You can say, like, Latino voters in Florida are moving this many points to the left, whereas Latino voters in Minnesota are maybe like, not as swingy or something like that, or more swingy or what have you. That's what you're saying.
- [45:36]
B
Precisely. And what we gain by doing that is basically we know the demographic breakdown of every single state. We can use voter file data to assemble a likely voter electorate per state, and we have. We have polling data to show how every single demographic we can model how every single demographic should vote on a statewide basis by blending, based on sample size, the state level estimate, the regional level estimate and the national level estimate. When we do that, we can say like, okay, voters in Florida, Hispanic voters in Florida projected a swing four points to the left. White voters in Florida projected a swing six and a half points to the left. Black voters in Florida projected a swing eight points points to the left. And you see what percentage of the electorate is each of these little groups. When you have that, you can say like, okay, Hispanics are what? Like, I'm going to make up numbers here. Like Hispanic voters, like 35% of Florida. Black voters are 20% of Florida. White voters are 45% of Florida. So 45 times 06 plus 35 times 04 plus 20 times 0.08 is equal to. That's the swing away that we get from the last presidential election.
- [46:51]
A
I mean, why do all of that as opposed to just looking at a poll?
- [46:56]
B
Because we don't have that many polls. We just don't have that many polls of statewide basis. Right.
- [47:01]
A
So is this aspect of your forecast going to diminish between now and election Day, when maybe we are getting more polls?
- [47:09]
B
Good question. It depends. You're not going to get that many polls. For example, like high quality polls of Florida. They just don't. A high quality Florida pollster, find me one, please, I beg of you, find me a high quality Florida pollster.
- [47:23]
A
But let's say in Maine or Texas or Iowa, this would be a smaller feature of your forecast because if we, you know, expect to get a bunch of polls.
- [47:33]
B
Right? Because we have a lot of data in that case for that state. Right. We have a lot more data in Texas at that point. Like by November, we'll have high quality pollsters in Texas. So at that point we can start relying on the polling averages more. But the MRP component will never fully go away. It will always be there because that is also polling data that we have that we've collected ourselves and that we are basically calibrating so we can basically figure out, okay, how do we project each of these voters to behave in these states?
- [48:08]
A
An important piece here is that like we've had issues in the past where pollsters are able to access a particular type of white voter or a particular type of even white Republican or even a particular type of Latino voter. And the people who don't respond to polls are unique in their own way and they still end up having an impact on the election even if they're not picked up in Polls. Do you worry at all about that, Laksha?
- [48:31]
B
All the time. We have a fundamentals level that we actually blend in. It's not just a pure mrp. Right. Like our model consists of fundamentals blended in with the mrp. So to some degree when you're like, wait, this doesn't make sense, like, why is that the case? That acts as a regularizer to drag it back towards what the fundamentals say. It's not too heavy, but it's there. Right now it's about 15% of our forecast.
- [48:54]
A
Zachary, you don't do this, defend yourself. I mean, in many ways he's defending himself for doing this because this is maybe a less traditional way of building the forecast. Zach, you are a little more traditional in this respect, if maybe a little less traditional in the Kalshi respect.
- [49:11]
C
Yeah. So I think, first of all, I think it's really, really cool. Like I'm always super interested in, like, you know what Laksha's MRP model says. It takes a lot of resources and time and effort to do this and I like greatly, greatly appreciate it that someone's doing it. So, like, for instance, like, I'm always like, you know, very excited to see Locksha's forecast and how it updates it is like really, really hard to do well this stuff well though, in the sense that. So I'll just talk about a different survey in 2024, which is called the CCES, like the Cooperative Election Study from Harvard University. And I love this thing. To be clear, it is the best kind of survey of voters. And for those who don't know what the CES is, it asks people like a bajillion questions about themselves like, are you a veteran? Did you get robbed in the past 12 months? Or what's your race? Did you get laid off from your job in the past 12 months? And then who you voted for? And the issue is surveying people is really, really hard in the modern day. So I'll give one example of where the CES in 2024 failed. And honestly, any pre election survey I saw in 2024 failed, which is it really understated how big the Asian shift was across the country. The Asian shift we find by looking at precinct data was probably about twice as big as the Cesar. It was much bigger than a lot of like, you know, people expected pre election. It was something I just wasn't prepared for on election night, like because of this, like if you're indexing really highly to what, like you think the survey will say in a particular district or state Even for Senate races, it could lead you astray if what you're finding like isn't completely correct.
- [50:52]
B
It thousand percent could. But I'm going to counter here that this is a problem with polling in general. Any forecast that uses polls is going to be susceptible to this. We're not using like. So for us we're way more careful with like how we try to represent different demographics. Right. We're also very, very sensitive to like small sample sizes and we don't let for example, like something with only like 70 responses play a significant part in our model. Like that would be crazy, right? So when we don't have data, we don't try to pretend. We just go with a lot of data when we have it. And I do understand still that this does not solve the representation problem, which is that, okay, cool, let's say you only go with Hispanic with like modeling Hispanic voters per state when you have like at least 300 or 400 respondents. But that doesn't really change the fact that are you picking up the right responses? There's two things here. Firstly, in a midterm this is much more solid than a presidential. I would be way more careful about using this for a presidential because is low propensity. Voters are really hard to pick up in polling. In midterms this is less of a problem. The second thing is that there is a lot of post stratification you can still do when you're polling to make sure that you actually represent these voters properly. So we've weight bipartisanship as well as by demographic. So we want to make sure we have the right amount of Republican Hispanics, the right amount of like Democratic Hispanics and so forth in our surveys. And so this is not, this is not just what the CCES does either. This goes one level deeper and tries to make partisanship level representations of each subgroup so you never end up crazily like swinging one way or the other.
- [52:39]
A
So Latcha, it sounds like this is a pretty interesting experiment you have going on here. I mean, what I'm really liking about this conversation in general is that you're both trying lots of different things. And then we're going to get results in the fall and we can sort of meet back up here and discuss how everyone did and what lessons you lear learned from the forecast that you've put together in the run up to election day. I want to, for the sake of drama, just maybe run through a couple other areas where your forecasts diverge. And I want to hear from you if there are any Places where you're like, actually, according to my priors internally or whatever, I think your forecast might be more on the money.
- [53:21]
B
I actually think Zach's forecast in Michigan is just strictly better than ours. The reason is this. We have Michigan, I believe at D 8. We have a rated it likely Democratic. That is indeed what the fundamentals would say with a generic Democratic nominee against a generic Republican. We have reason to believe based on statewide polling from Michigan. Right. Early statewide polling, which is not weighted very heavily in our forecast. Obviously that like, that is probably a little bit bullish on Democrats. Like the best poll we've gotten for Democrats is Haley Stevens leading by seven. And I'm out here saying in our model that the median case is Democrats winning by eight. I would be willing to take the under on that. What's happening is just our Michigan specific data does show a leftward swing, but really it's just that Michigan is a Trump +1 state. And the year we're modeling is D +9, which is a 10 to 11 point swing left from 2024. We just have, you know, the fundamentals point to. And if you, you. The fundamentals just point to this being a pretty Democratic race. There's obviously, this is just one case where I think Zach's forecast is better.
- [54:31]
A
Vohub has a 68% chance of the Democrat winning. You have a 94% chance of the Democrat winning in this case. So you would defer to vote Hub in that case. Lacha.
- [54:44]
B
I would, I would place that, that race in my own prior as a 75% chance of winning. I will say if Abdul El Sayed is a nominee and then we get started putting in his statewide polls because there is no nominee yet. Right. If he's a nominee and we start putting in statewide polls, I expect that to converge very rapidly towards Zach, but we're not there yet.
- [55:02]
A
Yeah. Okay. Zach, are there any parts of the Senate forecast that you look at and say I, I take Laksha's model.
- [55:10]
C
Yeah, absolutely.
- [55:11]
A
And.
- [55:12]
C
And I also think that this particular case also filters down to the House forecast in that area where I think one place the MRP is working really, really well right now in the argument Locksmith model model is Florida. So Florida. I think there's a couple things here. First of all, I think that Moody is a good candidate, which we have baked into both of our forecasts, which is why there's less of a shift here. But also I think the MRP is working really well in the sense that it seems like Florida Hispanics and Potentially Florida white voters, which are more older, are going to shift more towards Republicans in 2026. Just looking at how, how Hispanic Florida is, how it voted in 2020 and how it voted in 2024, you might think that Florida Senate is really winnable, but I don't think that that's the case. And then this filters down into Florida's House races where I, I think that I don't know the numbers off the top of my head, but I, I think locks numbers in Florida 14, Florida 25, Florida 22 are probably more accurate than compared to my priors.
- [56:19]
A
Okay, so Vote Hub gives Ashley Moody, the Republican, an 85% chance of winning the Florida Senate race, whereas split ticket slash the argument gives Moody a 92% chance of winning that race. Just to shout out a couple other areas that are interesting or I mean, one other main area we mentioned Maine already. That's a race that we've been talking a lot about, about recently. So maybe we don't have to delve back into that. But in Ohio, Vote hub has a 52% chance for Republican John Husted, while Laksha at split ticket you have a 66% chance for Sherrod Brown. Do you want to defend that, Laksha?
- [57:01]
B
Yeah, I absolutely would. I think the statewide polling you've gotten from Ohio, by the way, probably aligns with my forecast here because we are seeing a pretty big state specific swing in Ohio. Also, we're seeing shockingly weak numbers for the Republicans in the gubernatorial race in the generic ballot. I want to point out that Ohio we actually only have shared winning by like2. Some of the polls that have come out have him winning by like 8, which I do not buy. But I actually think here the fundamentals point to when you incorporate Brown's candidate quality and the fact that appointed incumbents generally don't get an incumbency boost, the fundamentals actually point to a slight Brown victory rather than a definition feet. But again, this is all really hard stuff and you know, I'm willing to be proven wrong here.
- [57:47]
C
I think that's, that's fair. At some point though, Ohio's estate won by Trump by I believe about 12 points in 2024. And I think something that pushes our forecast towards Republicans in Ohio, which I've talked about before, is we believe that the more competitive turnout dynamics in Ohio will help John Houston here. Right. So it's this idea that, you know, there might be big differential turnout favoring Democrats across a lot of the country, but the kind of likely voter advantage that Democrats get in Ohio is, is lesser because of the, you know, big spending by both sides. People know that it's a, you know, competitive Senate race. It's very important. My gut says that Sherrod Brown is favored. I think my gut is probably lower than 66% chance for him, though.
- [58:35]
A
All right, so I think we're going to wrap things up for today, and I think it would be great to get you guys back on and talk more about this as we go through the cycle. I do want to mention just before we close out that over at Vote Hub, you have a 45% chance that Dems win control of both chambers of Congress, a 27% chance that Republicans retain control of both, any 28% chance that Republicans win the Senate Senate and Democrats win the House. To round out the forecast in total, I'm curious, as we just close out here, what big questions are you going to be watching for the rest of the cycle in terms of, you know, how the news interacts with your forecast?
- [59:17]
C
Yeah, I think the thing I'm most looking forward to is polling in Maine and Michigan. In, you know, post Labor Day, there's a lot of speculation about, you know, how strong Democratic candidates will end up being in those two races is a good way. You can guess based on a bunch of different things like past electoral track record and experience and stuff like that, which Democrats don't grade especially strongly on. But in 2020, I remember when people said Raphael Warnock and John Ossoff, who didn't have exceptionally experience politically or strong electoral track records, and people said they might not be the strongest nominees. Now, six years later, we're sitting here and Laksh and I are both talking about John Ossoff as being extremely, extremely strong. Strong. So we'll get, you know, a lot of those questions are answered on election night, but we should at least get a better idea about where those races are going in October.
- [60:10]
A
All right, well, we're going to leave things there for today. Thank you so much for joining me.
- [60:14]
C
Thank you very much for having me on.
- [60:15]
B
Thanks for having me, guys.
- [60:17]
A
My name is Galen Druk. Remember to become a subscriber to this podcast@gdpolitics.com and wherever you get your podcasts. Paid subscribers get about twice the number of episodes. You can also join in our paid subscriber chat and pass along questions for us to discuss on the show. And you ensure that we can keep making a podcast that prioritizes curiosity, rigor, and a sense of humor. Also, be a friend of the POD and go give us a five star rating wherever you listen to podcasts, maybe even tell a friend about us. Thanks for listening, and we'll see you soon.