Silver Model X

[Lalo Scalf]

Ithink I started listening to the FiveThirtyEight podcast a little after the 2016 election, and started reading written content regularly in 2019 or so. I believe I got more value from the written/podcast content than the model alone because the contextualization helped me better understand the trends and larger political environment.

A couple of weeks ago, I wrote a short post about how you should steel yourself for 2024 because it\u2019s going to be very long election year. But I kind of chickened out: originally, that post was more about my own election-related stress. For the first time since 2008, I\u2019m not under contract with any media entity for election coverage. I have a choice about how to cover the election, or even whether to cover it at all. I still have some decisions to make \u2014 decisions that I\u2019ve been procrastinating on for a while now.

OK, OK, let\u2019s stop with this \u201Ccover the election at all\u201D business. I\u2019ve admitted to myself that I\u2019m probably going to be covering 2024 in some capacity. As much as I might want to be on a beach in Thailand or a vineyard in Bordeaux throughout the duration of the campaign, I know that realistically I\u2019ll be thinking about the election whether I want to or not. That being the case, I\u2019ll probably want to write about it. (Revealed preference: I\u2019ve already been writing about the election more here than I was planning. I promise there\u2019s a bunch of NBA stuff coming soon.)

Some of the anxiety is inspired by the way the primary process is unfolding. Covering \u2014 I don\u2019t know \u2014 a Raphael Warnock vs. Nikki Haley election sounds almost refreshing. However, barring some big surprises, that\u2019s not what we\u2019ll be getting. A Trump-Biden rematch is not quite inevitable, but it\u2019s looking rather likely. It may be that neither primary is really competitive at all. That means there could be another 14 months of a potentially very close election where everybody is on hair-trigger but there also aren\u2019t a lot of new storylines about the campaign from day to day.

Frankly, I can already sense that this is going to lead to a lot of annoying arguments, such as about the reliability of polling and about the role played by election forecasting and horse-race coverage. These discussions, like everything else in an election year, will be filtered through a highly partisan lens. (Any election involving Trump tends to make these issues worse.) It already feels like September 2024, even though it\u2019s September 2023, so I can only imagine what it will feel like a year from now.

I also want to cover the election without being a participant in the story. I don\u2019t want to be an avatar for someone else\u2019s grievances about election coverage. I don\u2019t want to be name-dropped in Nancy Pelosi\u2019s fundraising emails. I don\u2019t want some media blogger to write something like: \u201CYou know who really has a lot on the line tonight? Nate Silver.\u201D Not when the actual stakes of the election are so high.

Although written content is easy enough to scale up or down, the more difficult question is what to do with the election model. (I still retain that intellectual property even after my split with Disney.) I don\u2019t want to make this some philosophical rumination on the journalistic value of probabilistic election forecasts. I do think most of the critiques are ill-considered \u2014 forecasting is one of the fundamental ways to test whether your view of the world is accurate, and journalists ought to be concerned with accuracy \u2014 but let\u2019s leave it at that.

However, the FiveThirtyEight election model was subject to a certain, weird problem: it was way too popular. No seriously. It was too popular, and it tended to swallow everything else in its wake. The election forecasts we ran at FiveThirtyEight were seen by very large numbers of people \u2014 tens of millions of people. In fact, the 2016 forecast was literally the most engaging piece of content on the English-speaking Internet, according to Chartbeat.

When a product reaches that scale, you lose control of the narrative about it. As much as I might have been literally screaming at people to take Trump\u2019s chances seriously in 2016, for instance, far more people saw the numbers than read my careful contextualizations about them.

I don\u2019t think the story is quite so simple as \u201Cpeople don\u2019t understand probability\u201D or \u201Cpeople round 70 percent up to 100\u201D (there\u2019s actually not much empirical evidence for the latter claim1). It\u2019s more just that people interpret these forecasts in a lot of different ways, and those interpretations are mediated by a lot of partisanship and election-related stress. For every one reader who was interested in forecasting as an intellectual exercise \u2014 or who at least appreciated the effort to make horse-race coverage more rigorous \u2014 there were probably 10 who just wanted reassurance that their guy was going to win.

Impressions of the model can also be affected by media coverage readers see elsewhere, or by things they hear about the forecast from other people. Or, people may misattribute results they see from a different model to mine.2 In years like 2016 and 2022, for instance, the FiveThirtyEight model often got lumped in with other coverage even though it actually bucked it. (Our model was much more bullish on Trump than the consensus in 2016 and somewhat more bullish on Democrats than the consensus in 2022.) Indeed, there\u2019s a whole narrative about the FiveThirtyEight election models among political partisans that\u2019s completely detached from their long, well-documented track record of strong calibration and accuracy. It\u2019s hard to escape the depressing conclusion that many news consumers aren\u2019t that interested in accuracy at all.

I have a book coming out next year about gambling and risk that I\u2019m super excited about \u2014 it doesn\u2019t really have anything to do with elections at all, and is only tangentially about politics. I\u2019ve found it extremely refreshing to cover a new beat. I want to cover the election in the context of those other things I\u2019m doing. So the question is whether I can find a way to orient the model in the right way to the right audience \u2014 without it completely taking over my life. Otherwise, it might not be a good risk to take. Alternatives might involve taking the model private, selling it, or even just taking the year off from modeling and approaching the election with a little more arm\u2019s-length detachment. Some possibilities I\u2019ve considered:

I\u2019ve also had interesting conversations with outside parties about aiming for a different type of customer. At FiveThirtyEight, for better or worse, we were catering to the politics junkie audience. Leaning more into the business/tech/econ crowd or even the forecasting/gambling crowd might be a better fit for the model.

I\u2019ve had other conversations about doing a deliberately low-fi version of the forecast. Maybe I\u2019d hire a couple of people, and instead of updating the numbers 24/7, we\u2019d publish them once or twice a week. Maybe we\u2019d round probabilities to the nearest 5 percent. Maybe we\u2019d have some simple, attractive graphics, but nothing too immersive. Maybe we could cultivate a healthier relationship between readers and the forecast rather than having people click the refresh button 10 times a day.

Finally \u2014 and I have to admit this is an option I\u2019m finding increasingly attractive \u2014 I\u2019ve also had some discussions about taking the model private, whether selling it outright or licensing it for the cycle. People use election forecasts to hedge financial risk and to make other sorts of bets. One could even argue that this is the better use case for the model. Whether Trump\u2019s chances are 30 percent or 50 percent or 70 percent might not affect your decision-making that much as a voter. But it\u2019s definitely relevant if you have to make some calculated financial decisions based on the outcome.

If you\u2019re interested in discussing any of these possibilities, you can email me at NATE DOT SILVER DOT MEDIA AT GMAIL DOT COM. Thanks for reading, and back to our regularly-scheduled content soon.

To sign up so that you can be one of the first people to see the new forecast, just use the link below. Polling averages are free for all readers, while forecasts, probabilities, and additional detail on model inputs require a paid subscription.

\uD83D\uDD52 Last update: 2 p.m, Sunday, July 21: Joe Biden is dropping out, so this will be the final version of our Biden vs. Trump forecast. We\u2019ll turn the model back on in a week or so once we\u2019ve gotten enough polling on Kamala Harris and/or other prospective Democratic nominees.

Well, honestly, we don\u2019t know \u2014 but we can give you our best probabilistic guess. This is the landing page for the 2024 Silver Bulletin presidential election forecast. It will always contain the most recent data from the model.1

The model is the direct descendant of the f/k/a FiveThirtyEight election forecast2 and the methodology is largely the same, other than removing COVID-19 provisions introduced for 2020. (Other changes are documented here.3) If you\u2019ve received this post by email or are viewing it on the Substack app, we strongly recommend that you instead use the web version for interactive charts. Charts in the app don\u2019t update when we add new data, so switch over to your browser to see our most up-to-date forecast.

The Silver Bulletin polling averages are a little fancy. They adjust for whether polls are conducted among registered or likely voters, the presence or absence of RFK Jr., and house effects. They weight more reliable polls more heavily. And they use national polls to make inferences about state polls and vice versa. It requires a few extra CPU cycles \u2014 but the reward is a more stable average that doesn\u2019t get psyched out by outliers.

3a8082e126