🕒 The latest on Trump’s approval rating

Updated April 29, 2026

We’ve seen tons of new polls released this week. None of them have been good for Donald Trump. Many firms, including AP-NORC and Reuters/Ipsos, show him at a second term low. In CNBC’s polling, Trump is less popular than at any point in his first and second terms.

Unsurprisingly, this deluge of new results has moved our average. Today, 39 percent of Americans approve of the job Trump is doing, while 57.7 percent disapprove. And yesterday his net approval rating hit a new second term low of -18.8. For historical context, that’s right around where his net approval was at the end of his first term, in the aftermath of January 6th.

Trump’s approval rating on inflation and the cost of living has also fallen to net -40 in our average, blowing past its previous second term low. -EMD, 4/23/26

See also: Generic congressional ballot dashboard and Iran War support dashboard.

Subscribe now

This is the landing page for Silver Bulletin presidential approval ratings. We’ll regularly update the charts below as new polls about Donald Trump’s approval rating come in.1 This page is designed to be viewed on the web rather than in our email client or in the Substack app.

Click here for more information on how the average works. The Silver Bulletin average weights more reliable polls heavily — you can find our latest pollster ratings here.

The topline: So, just how popular is Trump?

Our default average reflects a combination of all polls, whether conducted among adults, registered voters or likely voters. If a pollster releases multiple versions of the same survey, we use the all-adult version of the poll before the registered voter version.2 This is because all Americans have a say in how popular the president is — whether or not they vote.

The polls: What do the surveys say?

Each poll gets an “influence” score based on its pollster rating, its sample size, its recency, and how often a pollster is publishing numbers.3 Sometimes, surveys with mediocre pollster ratings have more weight in the model just because they were conducted very recently or polled more people.

Inevitably, there’s a lot of disagreement from survey to survey, not just because of statistical variation but because pollsters have long had trouble pegging down Trump’s popularity. By clicking on the “adjusted results” tab, you can see how the “house effects” adjustment that corrects for these predictable differences works in our model. You can also click here to download every Trump approval poll in our database — including some additional details not shown in the chart below.

The issues: What do Americans think of Trump’s policies?

Pollsters don’t just ask about Trump’s overall job approval, they also ask about how well he’s dealing with different issues. We’re tracking how the public feels about Trump’s handling of four topics: the economy, immigration, trade and tariffs, and the cost of living.

The deep dive

We also have a few features we’re reserving for paying subscribers:

• How do Trump’s approval ratings compare to every past president since Truman?

• How many voters strongly approve or disapprove of Trump?

• What are Trump’s numbers in only polls of adults or only polls of likely and registered voters?

You can find all of that, plus downloadable data on every president’s approval rating every day since 1945, down below.

Read more

Subscribe now

This is our landing page for polls about the Iran War1, tracking support for U.S. military action in Iran since March 1, the day after the war began.2 We’ll regularly update this average as new polls come in.3 This page is designed to be viewed on the web rather than in our email client or the Substack app.

Click here for more information on how the average works. The Silver Bulletin average weights more reliable polls more heavily — you can find our latest pollster ratings here. Our process for calculating support for the Iran War is similar to that for calculating President Trump’s approval ratings.4

The topline: So, just how popular is the Iran War?

This average includes polls that ask about support for the Iran War, strikes in Iran, or U.S. military involvement in Iran. Compared with tracking approval ratings, questions about war can be less straightforward. Our goal is to limit the analysis to polls that ask neutral questions rather than introduce assumptions about how the war is going or its objectives. (See below for more details on how we make this assessment.) Our default version of the ratings reflects a combination of all polls, whether conducted among adults, registered voters, or likely voters.5

Share

The polls: What do the surveys say?

You can see all the numbers for each poll that feeds into our average below. You can also click here to download all the Iran War numbers.

We exclude questions that:

1. Make presumptions about the reasons for or against the war and/or offer an assessment about whether these objectives have been met (even if respondents are offered arguments on both sides);

2. Ask how the war is going relative to the respondent’s expectations (i.e., “Is the war going too far”?);

3. Ask about President Trump’s approval on Iran, as opposed to support for the overall war operation6, or;

4. Ask about some particular aspect of the war (e.g. targeting nuclear facilities) rather than the overall situation.

Also, in line with longstanding Silver Bulletin policy, surveys that fail to provide the specific question wording or other key details about the poll (dates, sample frame, sample size) can’t be included.

Polls are adjusted for house effects (i.e. persistent differences between a pollster’s findings on the war and the consensus) and are weighted based on each firm’s pollster rating and the poll’s sample size and recency. The “influence” score in the table below shows the relative importance of a poll in our current average given these considerations.7 In practice, recency is often more important than the other factors in our weights.

As a reminder, we update this database about once per day. If a poll seems to be missing, we may not have added it yet, we may be awaiting more information from the pollster, or it may contain question wording that doesn’t meet our standards for this project.

Subscribe now

Subscribe now

This is the landing page for Silver Bulletin’s 2026 generic congressional ballot polling average. We’ll regularly update the charts below as new generic ballot polls come in, and eventually, this average will feed into our 2026 midterm forecast model.1 This page is designed to be viewed on the web rather than in our email client or in the Substack app.

Click here for more information on how the average works. The Silver Bulletin average weights more reliable polls more heavily — you can find our latest pollster ratings here.

Who is favored to win the House in 2026?

Our default average reflects a combination of all polls, whether conducted among adults, registered voters or likely voters. If a pollster releases multiple versions of the same survey, we use the likely voter version before the registered voter version.2 That’s because for this average, we’re interested in people who plan on voting in 2026.

Every generic ballot poll in our database

Each poll gets an “influence” score based on its pollster rating, its sample size, its recency, and how often a pollster is publishing numbers. You can find that in the table below. Sometimes, surveys with mediocre pollster ratings have more weight in the model just because they were conducted very recently or polled more people.3

Inevitably, there’s a lot of disagreement from survey to survey, not just because of statistical variation, but because some polling firms consistently lean toward Democrats or Republicans. By clicking on the “adjusted results” tab, you can see how the “house effects” adjustment that corrects for these predictable differences works in our model. You can also click here to download every generic ballot poll in our database — including some additional details not shown in the chart below.

State benchmarks and every generic ballot poll since 1994

We also have two cool features we’re reserving for paying subscribers:

• Benchmarks in each state. Which party would we expect to be ahead in, say, Georgia or Michigan or Ohio, given the current generic ballot?

• And generic ballot averages going back to 1994.

You can find all of that, plus downloadable generic ballot data for the past 30 years, below.

Read more

Compared to some recent hot topics — like the Iran War and the attempted shooting at the White House Correspondents’ Association Dinner — election administration looks fairly mundane. Some might even call it dull. That’s generally a good thing, because vote counting only becomes the headline after something goes wrong. But one man who does care about election administration is Donald J. Trump. Aside from his longstanding insistence that the 2020 election was “stolen” — which, of course, it wasn’t — the president has put out a steady stream of executive orders and Truths about how Americans vote and how those votes are counted.

Many of the Trump administration’s election-related actions get held up at some point in the legal or legislative process. Still, it can be hard to know what to make of the competing executive orders and bills. Will they all get struck down in court and not change anything? Or will Trump be able to use the presidency to affect — or even steal1 — the midterms? To answer those questions, I thought it would be best to talk to someone who covers election administration for a living: Nathaniel Rakich, the Managing Editor at Votebeat (and Nate’s former colleague at FiveThirtyEight). Here’s the lightly edited transcript from our conversation.

Subscribe now

Eli McKown-Dawson: So before we get started — how have things been at Votebeat?

Nathaniel Rakich: They’ve been great. Elections are such a hot topic, and the administration of elections has been under the microscope to a degree that it really hasn’t been, at least during my political consciousness. These are really momentous things, and certainly, if the worst fears of people on the left come true about ICE at polling places, that’s going to be a tremendously important story. Obviously, who wins the election is the ultimate issue. But the really important thing is just making sure that everyone can vote and that all the votes are counted faithfully.

Trump’s executive orders have mostly been limited by courts

Eli: We’ll get to ICE at polling places eventually, but I’ll start a little simpler. There have been two big executive orders from the Trump administration on elections. The first one was issued about a year ago. Can you give me a brief walk-through of what was in that order and where it stands now?

Nathaniel: The first executive order was a real grab bag, and really was the first salvo in Trump trying to bring the federal government into election administration to an unprecedented degree. The main thrust of it was basically the SAVE Act and the SAVE America Act. It attempted to require people who register to vote to prove their citizenship with a document — like a passport, or a combination of documents, like a birth certificate plus a driver’s license. They tried to change voting machine certifications. Obviously, Trump pushed conspiracy theories about voting machines manipulating votes back in 2020. He’s been all about paper ballots.

The upshot is that the executive order was immediately challenged in court, mostly on the grounds that the Constitution pretty clearly delegates to the states and to Congress the power to run elections. The president does not have a role. So any executive order Trump issues is automatically going to be suspect, and that’s probably a kind word. Courts have consistently struck down, or I should say blocked, because the cases continue to proceed, and Trump could theoretically take it to the Supreme Court, the executive order. None of the major provisions from that order are operative right now, which is why the elections are proceeding normally.

Eli: A year later, we have a second elections-related executive order that was just issued. That one is more specifically about voting by mail, right?

Nathaniel: Yes. The key provision says that the Postal Service will maintain a list of people who are allowed to vote by mail, and it cannot deliver ballots from anybody who is not on this list. The confusing thing about the order is that it doesn’t specify how that list is generated. The order also contains provisions for a couple of other lists. It says that states, 60 days before the election, should send a list of people they expect to vote by mail to the US Postal Service. And then the very next provision in the executive order says the US Postal Service will send to the states a list of people who are allowed to vote by mail.

A lot of people assume that the second list is going to be based on the first list, but the order doesn’t actually say that. It doesn’t say how that second list is going to be generated. It doesn’t say anything about what the Postal Service should do with that first list from the states when they get it. And then there is a third list that appears to be completely separate from all of this. The order directs the Department of Homeland Security to create a list of US citizens above age 18 residing in every state, and to send that list to election officials in those states. Again, it does not require the state election officials to do anything with that list.

There’s been a lot of coverage that talks about this as creating a federal list of registered voters, which of course doesn’t currently exist — states and smaller municipalities like counties are responsible for maintaining their own voter rolls. But that’s not quite what this is, because it doesn’t have any legal teeth. And also, the list is not actually of voters, it is of adult citizens. That’s an important distinction, because the list doesn’t make any attempt to, for instance, weed out convicted felons who in most states can’t vote.

There have been some states, particularly Republican states, that have used federal databases to cross-compare against their voter rolls to check if they have non-citizens on the rolls, and it’s not difficult to imagine that they will use this list for a similar purpose. It’s also been speculated that the Postal Service will use the DHS list to be the intermediary between the two lists. The logic there would be: states send the Postal Service a list of voters that the states say are going to vote by mail; USPS goes through that list, strikes out the people who aren’t citizens according to the DHS list, and then sends the list back to the states, saying these are the people we think are allowed to vote by mail. But again, the order does not specify any of this. The order basically just puts these lists out into the ether.

This executive order is also being challenged in court on much the same grounds as the first one. There’s really no reason to think it’ll meet a different fate. The DHS provisions seem like they could hold up. DHS can make whatever lists it wants, especially since the list isn’t being used explicitly to remove people from voter rolls. But the parts about involving the USPS in mail voting, which is supposed to be the purview of the states, are likely to get struck down, or at least put on hold temporarily. I would be surprised if this order changes anything about the way that the midterms are run.

We are about seven months before the election, and it’s starting to get to that point where people are getting nervous about Trump trying to change rules at the last minute in ways that are perceived to benefit him.

Eli: And how involved is the Postal Service currently in administering elections?

Nathaniel: This is unprecedented. The US Postal Service has never had any other role in determining who will vote by mail. It just does the mailing.

Eli: One thing I was curious about with these lawsuits: is it all Democratic states or have there been Republicans signing on as well?

Nathaniel: It’s been partisan. There are currently four lawsuits. One is by the Democratic Party. Two were by voting rights groups. And then the last one, the most important one, is from the states. The states are obviously the ones who are the aggrieved party in this case, because it’s their powers that are being allegedly usurped by this executive order. That lawsuit was joined by Democratic officials from 23 states plus DC. That said, a couple of Republican secretaries of state have weighed in, neutrally to negatively. They’ve reacted coolly to the order, but none of them have sued over it.

Will Congress help Trump change how elections are administered? Probably not.

Eli: With both of these executive orders, what do you think the purpose was? Trump has for a long time talked about voter fraud in 2020, and non-citizens voting. But you have to assume that the expectation was that these orders would get held up in court. Did the administration expect these orders to do anything in the midterms? Or was it more of a messaging thing?

Nathaniel: That’s the million-dollar question. We don’t know what Trump’s motivation was. It’s hard to read his mind — famously, as the whole country of Iran is currently thinking about. On one hand, there’s a theory that he thought it would genuinely work. Maybe he thinks that he has these powers and thought that the courts would agree with him. Obviously, the federal court system as a whole is pretty conservative. He’s appointed a lot of them. Should it reach the Supreme Court, because of how clear the Constitution is on this issue, I still think they would rule against him, but it’s not a slam dunk. So Trump might have thought this was really going to work.

He could have thought this might not work, but it doesn’t hurt to issue the order and see what happens. And then the cynical interpretation is that he knew these weren’t going to pass court muster, but he’s just trying to stir up discontent around the election and create a premise for saying “Hey, I tried to require proof of citizenship, I tried to clean up mail voting, but the courts didn’t let me. And the election went forward and Democrats won, so it’s fraudulent.” And when the courts rule against any attempts to overturn the results, he can continue to say they stopped him, and delegitimize the election. The only thing I am comfortable saying for sure is that Trump clearly has strong feelings about elections and how they should be run, and feels that they are not run well and that it disadvantages Republicans.

And then it was, all right, Plan B — that was the SAVE Act and the SAVE America Act. If that were to pass Congress, the legal challenge would be a lot harder to make, because the Constitution explicitly says Congress can pass laws governing this process. But obviously the filibuster exists, and so that is unlikely to bear fruit as well. The timing of this executive order was probably not a coincidence, because it was after you had seen the SAVE America Act stall. Trump said you have to pass the SAVE America Act before you put any other bills on my desk. You should blow up the filibuster to do it. He put a lot of pressure on John Thune, and the votes just weren’t there to change the rules. At that point, that effort is dead. And so then Trump is like, all right, well, the ball is back in my court.

Eli: It doesn’t look like that’s going to pass at this point, right?

Nathaniel: Yes. In order for things to get through the Senate, they need to get past the filibuster, and that requires 60 votes. Republicans only have 53 votes. Democrats are never going to allow the bill to get past that threshold. A lot of Republicans were advocating using the talking filibuster.

The issue with that is that it burns potentially weeks of legislative time, and the Senate has other stuff it wants to do. A lot of Republicans didn’t even want to change the rules to do that, much less abolish the filibuster as a whole. Republicans are looking ahead to a potential future in which Democrats control the Senate again, and they would like the filibuster to still be in place at that time. So as long as the filibuster is still in place, there’s just no way for that bill to pass.

There have been talks about squeezing the SAVE America Act into the reconciliation bill somehow, which doesn’t require the 60-vote threshold. But it’s hard to make an argument for why a voting bill would be related to the budget, which is a requirement for reconciliation. So I would say that the odds that the SAVE America Act passes through the Senate are very small.

Eli: Just for the record, the SAVE America Act contains a lot of the same content as that first executive order, yes?

Nathaniel: It is the requirement for proof of citizenship. If you could ask Trump, “If you could change one thing about elections, what would it be?” — I think he would say we need voters, when they register to vote, to provide physical documentary proof of their citizenship. That is the main thrust of SAVE America. It also has a provision to require a photo ID to vote. Even for states that already have photo ID requirements — some accept certain types of IDs, and other states require other types — this would be a uniform requirement and override those. Also, a lot of Democratic-led states do not currently have voter ID laws, and this would implement them.

Eli: How is this looking at the state level? Because certain states, Florida for example, have tried to pass, or have passed quite similar bills. How many states have done that or are going to try before the midterms?

Nathaniel: That’s certainly where the real action is. If a state passes a law requiring documentary proof of citizenship, that is totally kosher — at least, you’d have to come up with a different legal argument to block it. I’m sure somebody will still file a lawsuit. In fact, I think they did in Florida. But there has been more movement on the state level.

Obviously, the blue states aren’t going to do anything. I believe there are four states that have enacted proof of citizenship laws just since the beginning of the year: Florida, South Dakota, Utah, and Mississippi. Going into the year, only three states had this policy: Arizona, New Hampshire, and Wyoming. So that is obviously a big change due to a movement that Trump has started. I should also note that none of these laws take effect until 2027. [Editor’s note: Rakich later clarified that only Florida’s law is delayed until 2027; the others take effect this year. See correction below.2] So that’s an important difference. The SAVE America Act would be effective immediately, and that is a big concern for election administrators. If they somehow abolish the filibuster tomorrow and pass this bill, election administrators would immediately need to start checking citizenship for newly registered voters. They don’t have the funding to do that. They don’t have the time to do that in addition to their regular duties preparing for the election.

Another way in which states have moved in the direction Trump wants is by changing absentee ballot receipt date deadlines. As part of the first executive order in 2025, Trump attempted to require all ballots to be in by Election Day. A lot of states — I believe it’s 14 — allow ballots to count if they arrive after Election Day, as long as they are postmarked by Election Day. That’s why in states like California the count takes a long time, because ballots are still coming in. Trump, in his first executive order, attempted to say no, every ballot is due on Election Day. That part was blocked in court. But there have been [several] Republican-led states that have said we are going to change our law from a postmark deadline to a receipt deadline since Trump entered office.

The other interesting thing is that Trump has said on Truth Social several times that we should get rid of mail ballots entirely. He has also said we should go totally to paper ballots and not use voting machines. Those are two things that even red states are not moving on at all. Vote-by-mail in general — and by that I mean no-excuse absentee voting — is popular. Polls show that even among Republicans, a decent chunk think voting by mail is fine. It’s popular among Republican politicians too. Campaign operatives know that it’s really helpful to be able to get mail ballots to people and track who has voted that way. So there is just no interest among Republicans to get rid of that voting method. And then there is an understanding that getting rid of voting machines would be an extremely costly and logistically daunting, if not impossible, task. So Republican states have quietly ignored that part of Trump’s request.

Eli: It’s also a speed and accuracy thing with voting machines, right?

Nathaniel: Yes. The Trumpy, platonic ideal is that everybody votes on a paper ballot and then those ballots are counted by hand. But a couple of municipalities have decided, in the wake of all this, to hand-count their ballots. It’s just a mess. It takes forever. It takes significantly longer than by machine. It’s less accurate. There’s a reason that we use voting machines to tally ballots.

Can Trump actually rig the midterms?

Eli: Obviously, with these measures we’ve been talking about — citizenship laws, vote-by-mail, absentee ballot receipt dates — there tends to be a partisan split. On one hand, there’s the “you want everyone who is eligible to vote to be able to vote” angle. But people also often talk about the partisan political dimension: is making it harder to vote going to benefit one party? The common understanding used to be that these laws will make it harder for Democrats to vote and advantage Republicans. Now lots of people talk about how that’s flipped, and actually, if you only allow people with a passport to vote, it’s going to be good for Democrats. What do you think of all that?

Nathaniel: It is not clear to me at all that these laws would hurt Democrats. I do think that is Trump’s expectation. He’s basically all but said, if you get rid of mail ballots and have voter ID, then Republicans will never lose another election. But I just don’t think the evidence is there. It’s highly uncertain. I think it’s important to have wide error bars on this.

But there are ways in which it hurts the Democratic coalition — people of color are less likely to have these documents. But those voters, especially Latinos, have been getting more Republican. In addition, older and rural voters are two groups who would be uniquely hurt by proof of citizenship requirements specifically. And then for rural voters, one thing that is underdiscussed about proof of citizenship requirements is that they would functionally eliminate the ability to register to vote by any method other than in person.

In most states, you can register to vote by sending in a form via the mail or do it online. Now, you could send those forms in, but before your registration was finalized you would still have to go to an elections office physically and show them your documents. For people in rural areas who might live a two-hour drive away from their county seat, that would be a real hassle. So it is not at all clear to me that these laws would have the partisan impacts that Trump hopes.

Eli: Whenever one of these executive orders comes out, you get a lot of Democrats who have this very cynical view of what all of this means: that this is Trump trying to take control of elections, and that he’s going to keep pushing. It goes anywhere from having ICE at polling places to straight-up canceling the midterms. As an observer, I tend to discount a lot of the most out-there stuff. How concerned should people be, if at all? Is there anything there to be taken seriously?

Nathaniel: There’s no question that Trump wants to involve the federal government in elections to a higher degree. That’s what the executive orders do. He’s also had the Department of Justice request voter rolls from every state except North Dakota, which doesn’t have voter rolls. DHS has also created or augmented an existing database called SAVE, which is essentially a database of people and their citizenship status. The Trump administration upgraded that system last year to be able to upload voter rolls in bulk. So several Republican-run states have given their data over to DHS to basically spit out a list of people who aren’t citizens.

The issue with that is there are a lot of false positives in the data, because the data that SAVE’s database relies on is unreliable. A state like Texas, for instance, got this data back from the federal government that said “here are all the non-citizens on your rolls’” The Secretary of State’s office told county clerks to get these people off the rolls. But when they went to check, a lot of those people were citizens. It’s important to note that states aren’t allowed to kick people off their rolls without notifying them, or at least giving them a chance to prove their citizenship. So it’s not as bad as it could have been — not mass voter purges — but the federal government is being used to at least try to take people off the voter rolls.

The worst-case scenarios are definitely more outlandish. Trump cannot cancel the midterms, full stop. He could issue an executive order tomorrow that said the midterm elections are hereby canceled, and they would still happen — both because the Constitution does not give the president the right to set election laws, and because the president has no role in administering the election. That is run on the state and local level by people who are a mix of Democrats, Republicans, and nonpartisan election officials, and they are almost uniformly — from all my conversations with them and our reporting — committed to doing the election. They are also all legally obligated to do the election. States have their own laws that they have to follow. So there is no chance that the midterms will be canceled.

The question of federal troops at polling places is a thornier example. That’s where you get into some uncharted waters. It is clearly illegal under federal law and a lot of state laws for armed troops or agents to go to polling places. But we don’t really know what would happen if Trump tried to do something like that. There’s also a chance that, for instance, ICE agents could conduct a bunch of raids in the neighborhood on Election Day — that’s not at polling places, so it conveniently goes around the prohibition. But obviously that’s something that might depress turnout, and could be considered interfering with the election. That is a concern. But the administration has said repeatedly that they are not going to send troops or ICE to polling places.

There was a concern earlier last year about National Guard troops. It’s also obviously worth noting the National Guard is under the control of the governor, and the courts have so far found, when he was deploying them, that if the governor doesn’t accede, then Trump can’t override that. But there are states with Republican governors. Greg Abbott in Texas, if that Senate race is tight, could he allow Trump to use the National Guard? Potentially. But it couldn’t happen in a state like Pennsylvania, where there’s a Democratic governor.

So there are a lot of legal and practical barriers between us and a situation where there are federal troops at polling places. But it is one of the bigger question marks, because even if you think it is unlikely — and I think it is unlikely — it’s not impossible. Obviously, if you’re an election official, you should be preparing for this scenario, even if there’s a 10 percent chance, because it would be a pretty bad scenario. If, for instance, National Guard troops impounded voting machines or ballot boxes, that raises questions about whether those ballots will be counted because the chain of custody is broken, which means you can’t ensure that they haven’t been tampered with.

The other thing I would say is that you can’t have a mass operation to do this. They’re not going to send National Guard troops to every polling place in Texas, because there just aren’t the numbers to do that. But could you have one instance in which a National Guard unit or a group of ICE agents is sent to one specific polling place in a Hispanic-heavy neighborhood and impounds the ballots? That would be bad enough, because it would interfere with the clean administration of the election and potentially throw out ballots of several dozens or hundreds of eligible voters. So it’s something worth being mildly worried about.

Eli: When you say mildly worried, if I had to press you and give you a scale of, say, one — normal election — to ten — canceled elections — where are you on that scale today?

Nathaniel: It’s somewhere in between. I think you should be at a one. That should be the default. Anything above a one is not good. I don’t want to give you a specific number. What I’m trying to communicate to people is that there are unusual things happening in this election with regard to federal interference, and there are potentially things that could happen that would be quite disruptive. But at the same time, it is unlikely that there will be anything on any kind of wide enough scale to reverse an election result. And certainly, the Democrats’ doomsday scenarios of a canceled midterm, or outright invalidating the election, are so far-fetched as to be not worth worrying about. But that doesn’t mean that the specific scenario I described of maybe 200 ballots getting thrown out in Arizona in a key House race — that is worth being concerned about in and of itself, even if it doesn’t signal the death of American democracy.

Eli: You mentioned talking to election officials, and them being almost uniformly committed to running the election normally. What has your experience been talking to them? I’d be curious about Democrats, and especially Republicans.

Nathaniel: Virtually all election officials are really good public servants who’ve been doing this for a long time and really want things to go smoothly. They don’t want any problems at polling places. These are local people. In many cases, they know their voters; they have personal relationships with their communities. Nobody wants to disenfranchise their own voters. So they are preparing for the election the same way that they always do. A lot of them are nervous and are hoping that none of these nightmare scenarios come to pass, but they are preparing for them. For the most part, they’re unfazed. They’d rather not have the executive orders looming over their heads, but they are getting ready to administer the elections the same way they always have.

Eli: The two executive orders are probably going to be held up in court, the SAVE Act probably not going to pass. Is there any chatter about anything else coming down the pipeline between now and the midterms?

Nathaniel: At the signing ceremony of the executive order last week, Trump made a comment that was basically like, “We’re still working on the citizenship issue” — it’s not clear whether he was referring to the SAVE America Act or whether he was saying that they’re going to do another executive order. It is definitely up in the air that there could be another executive order. Executive orders are ultimately going to be a fruitless way to make any changes to the election. But the closer you do it to the election, the more nervous it makes people, and the more potential there is to at least confuse people about the rules.

Eli: What are you going to be paying attention to, election administration-wise, heading into the midterms?

Nathaniel: Certainly making sure we keep an eye on the executive orders and the SAVE America Act. I don’t expect anything to happen there, but you never know. There are two important voting cases in front of the Supreme Court right now. One is about whether ballots can arrive after Election Day. Basically the argument there is that Election Day is Election Day, and any ballots that come in after that have missed a deadline. If the Supreme Court agrees, then those 14 states that still have postmark deadlines would change. That would be a fairly significant change, especially for a state like California. The other case threatens to overturn the Voting Rights Act. At this point it is too late to materially affect the 2026 midterms. But if the Voting Rights Act is indeed overturned, that would lead to widespread, probably nationwide redistricting before 2028 that would eliminate a lot of majority-minority districts, especially in the South, which tend to benefit Democrats.

Otherwise, we’re going to have to see how aggressively Trump tries to interfere with the administration of the election in the fall. And then we’ll see to what extent the results are legally challenged. The 2020 election was challenged on extremely dubious grounds, and it was a concern for democracy, culminating in January 6. But at the same time, the system held. I would expect the same to be the case here. The only concerning scenario is if Democrats do a lot worse than they think they’re going to do, and the House comes down to one or two seats, and then those seats are really tight. Any close election is subject to litigation, and there could be recounts, and the validity of every ballot gets scrutinized. That’s always a concern if the election is super close, but that is such a specific scenario that I don’t think it’s worth worrying too much about. But it’s something I’ll be keeping an eye on.

Subscribe now

This the landing page for the Silver Bulletin average of Elon Musk’s favorability polls, tracking him back to the start of 2024.1 We’ll regularly update this average as new polls come in.2 This page is designed to be viewed on the web rather than in our email client or in the Substack app.

Click here for more information on how the average works. The Silver Bulletin average weights more reliable polls more heavily — you can find our latest pollster ratings here. Our process for calculating Musk’s favorability ratings is similar to that for calculating President Trump’s approval ratings, although we use slightly more conservative settings as Musk is polled less often than Trump.3

The topline: So, just how popular is Elon Musk?

Our default version of the ratings reflects a combination of all polls, whether conducted among adults, registered voters or likely voters. If a pollster releases multiple versions of the same survey, we use the all-adult version of the poll before the registered voter version.4

Share

The polls: What do the surveys say?

You can see all the numbers for each poll that feeds into our average below. You can also click here to download all the Musk favorability numbers — and some additional details not shown in the chart below — for every Musk favorability poll in our database.

This is a free feature, but it requires work to update and maintain. We very much appreciate your support via paid or free subscriptions.

As I wrote after the assassination of Charlie Kirk, we struggle with how to cover this sort of political violence at Silver Bulletin. On the one hand, these stories don’t really play into our expertise. And “takes” in the immediate aftermath of these events are notorious for relying on incomplete or inaccurate information and otherwise “not aging well”. On the other hand, it’s an objectively important political story — it would have felt weird to publish the story we were originally planning for today (a good politics story but not time-sensitive) as though nothing had happened over the weekend. And to some extent, the premise of any good blog is in having access to the author’s thought process while events are still in progress.

So I compromised by talking it out with Eli, who was attending the Substack New Media Party near the White House2 when the WHCD incident went down. You can find the video above. (Note that there are a few minutes of dead space at the end.) As much as I’m trying not to write a long, complicated essay about this, there were basically two points that I kept returning to in our conversation.

People shrugged this off, but it’s not a good sign if this feels “normal”

How one experiences breaking news depends on one’s circumstances. On Saturday night, I was in kind of a zone trying to get our soccer/World Cup model over the finish line. And uncharacteristically, I wasn’t really checking my phone. By the time I learned about the WHCD incident, the danger seemed to have passed.3

I know from friends how harrowing the experience was if you were actually at the Washington Hilton. But as Eli says, otherwise the show pretty much went on. (Including at the Substack party, where guests were literally locked in for some period of time.)

On Twitter, people were continuing their usual conversations about the NBA playoffs and the red and blue button4 and whatever else. Roughly two or three times a year, there’s a political event that stops the world in its tracks, where writing about anything else would seem “tone deaf”. Both the assassination attempt against Trump in Butler, Pennsylvania in 2024 and the Kirk assassination fell into that category. Based on the social media vibes, Saturday night didn’t feel the same way — even though the events unfolded at the literal epicenter of media and political power.

And maybe that’s … not great. I never want to tell people what they should or shouldn’t care about or critique their emotional reactions. But an assassination attempt against the president of the United States is the very definition of a major political story. And the fact that this sort of thing is happening more often is a reason to feel more worried, not more complacent because you narrowly averted disaster. If your next-door neighbor drives home drunk every Friday and has knocked down every other mailbox on your block at some point, your conclusion shouldn’t be that he’s a great driver because he’s avoided worse.

Share

One can debate exactly how close Allen got — he’d checked in as a guest at the Hilton, and his manifesto trolled the Secret Service. But multiple layers of security exist for a reason, and he was stopped. It wasn’t Butler, where a bullet literally hit/grazed5 the president’s ear.

Still, one theme in studying low-probability events is that near-misses are informative. A candidate who loses an election by 100 votes in what would have been a huge upset — well, for forecasting future elections, that tells you basically the same thing as if she’d won by 100 votes.

Saturday night could easily have been much worse. So, obviously, could Butler. So could January 6 but for the bravery of the Capitol Police. The lesson we should take from these events shouldn’t be “usually, everything turns out OK”. It should be that the nation is constantly flirting with disaster.

Rationalization, minimization and denialism

Here are a couple of points I’m specifically not making:

• I don’t particularly care whether the left or the right is “worse” at stochastic political violence in America in 2026. It isn’t a race you want to win. Or if it is a race, it’s a race to the bottom. The fact that I can rattle off from memory a half-dozen recent violent incidents with Republican or right-coded targets and another half-dozen with Democratic or left-coded targets … in either case, that’s way too many.

• I’m not into language policing. Lots of common idioms involve gun metaphors or other violent imagery (“the CFO had a target on his back”). Actually, most people understand metaphors and have a lot of experience distinguishing figurative from literal speech. I’m also not much into tone policing. I don’t think people should feel any obligation to express outrage or “thoughts and prayers” especially when it isn’t sincere. You don’t have to say something about every terrible thing.

Rather, what I object to are substantive rationalizations of political violence. Few “respectable” people do this outright. Instead, as Eli and I discussed, it usually comes in the form of “yeah, buts”. One classic example of a “yeah, but” is U.S. Senator Chris Murphy after Luigi Mangione’s murder of a health care CEO:

Listen, I’m never going to condone violence. I don’t here. I spent my entire life fighting gun violence because I know that the devastation is enormous. ... What I see happening in this country, though, is a real visceral anger that the outrage at Brian Thompson’s death or the outrage at the death of any powerful person isn’t matched by the anger over the thousands of people who die often anonymous deaths every single day in this country at the hands of a healthcare industry that mostly doesn’t give a s--- about people and only cares about profits.

Let me be precise here: I definitely don’t think that Murphy is condoning or endorsing Brian Thompson’s murder. But he’s providing a rationalization for Luigi. And his preamble doesn’t change that much. If you tell your waiter “no offense, but this is the worst meal I’ve ever had, and your service stunk too”, the waiter is within his rights to take offense. If you’re yeah-butting, the “yeah” doesn’t give you immunity for what happens after the “but”.

The other main technique is denial. I wouldn’t say it’s super pervasive, but I’ve been surprised at the number of semi-respectable Democratic-leaning commentators who have flirted with the idea that Butler was a “false flag” or that Saturday night was. Another form of denial is to misinform your audience about the identity of the shooter or his motivations: sure this was bad but our side didn’t do it. I was very disappointed by Heather Cox Richardson, for example, when she implied that Tyler Robinson, Charlie Kirk’s assassin, was a right-wing MAGA Groyper, and did nothing to correct the record when it became clear that he obviously wasn’t.

Somewhere in between denial and rationalization is minimization. The most common form of this is the other side is worse or just as bad. Alternatively, you can minimize the larger argument by foregrounding irrelevant details. There are just so many times when I’ll see someone trying to score some narrow political point in the wake of a tragedy or near-tragedy. They probably think they’re being persuasive, but it’s a reliable sign that they’re not worth listening to the next time around.

Yesterday, Virginia voters approved a referendum that would give Democrats partisan control of the state’s redistricting process, removing it from a bipartisan commission. Barring court challenges, the new districts will take effect for this November’s midterms, with Democrats favored in 10 of Virginia’s 11 districts, creating the possibility for a 4-seat gain over the current 6-5 delegation. (Although Democrats might have picked up one seat anyway, given what’s likely to be a favorable environment.)

Some of the districts are pretty jagged, as is typically the case when parties get to draw their own maps:

The new map is not totally without downside risk for Democrats. Five of the state’s districts were carried by Kamala Harris by single digits, leaving Dem incumbents vulnerable in a Republican wave year or should Virginia revert to being more purple.

But the reward is worth it. Axios estimates that Democrats have actually gained ground from the mid-decade redistricting war that began in Texas. And Kyle Kondik calculates that the map that will be contested in November is almost perfectly fair.

Subscribe now

There are still court decisions left in a few states, and our model will provide a more precise version of this calculation once it launches — it wouldn’t shock me if the map winds up slightly favoring Democrats.1 Still, if Democrats win the popular vote for the House by any margin (and they probably will) they’re very likely to take control of the chamber.

Nobody likes I-told-you-so’s, but … well … here’s what I wrote in August:

That column pointed out that if both parties took a truly maximalist approach to redistricting, it might actually work to Democrats’ benefit. Although Republicans have a gubernatorial/legislative trifecta in more states, Democrats’ trifectas are in more populous states. So once Abigail Spanberger flipped Virginia’s governorship in November, Democrats inherited trifecta control over a larger number of Congressional seats:

Another part of my theory was that Democrats tend to turn out in bigger numbers than Republicans for any sort of special election like the one in Virginia yesterday, which makes the party more nimble if districts are constantly being rejiggered.

The Virginia margin was underwhelming, but the result was still impressive

But here’s some grist for the Nate haters: in Virginia, part of my “I told you so” deserves a demerit. As Patrick Ruffini points out, turnout was actually higher yesterday in red-leaning counties where voters are understandably concerned about losing their representation in Congress to gerrymandered districts.

It wasn’t enough to carry the day, though: “yes” won by 3 points. (Results aren’t official yet.)

One could argue that the margin was a little underwhelming. Harris won Virginia by 6 points in 2024 in a mediocre electoral environment for Democrats; “yes” won by 3 with Trump’s approval rating at 38.8 percent and falling. There were a few voters like Virginia resident and friend-of-the-newsletter Matt Glassman who voted “no” on principle.

Indeed, yesterday’s result may come to represent the upper bound for how far Democrats are willing to push things. In a slightly more purple state or in a slightly more red-leaning political environment, the referendum might have failed. And Spanberger’s downside is protected because governors in Virginia are limited to one consecutive term. Change any of those factors, and I’m not sure you’re going to see governors pushing for this. I wouldn’t expect Michigan’s Jocelyn Benson to campaign for partisan districts in that state should she and Democrats win a trifecta there later this year, for instance.

On the other hand, this is a pretty damned impressive display of partisan coordination. Less than a decade ago, polls showed Democrats opposed to gerrymandering by 2:1 margins — note that Republicans were opposed to it by almost as wide a margin. And one of Barack Obama’s and Eric Holder’s big post-presidency projects was to fight for “fair maps”, which used to mean opposing gerrymandering and partisan control over districting.

Obama urged Virginians to vote yes, however.

Is this hypocritical? Yeah, kind of. But I’d like to play a tune from the world’s tiniest violin for Republicans who are crying foul about all of this. Gerrymandering goes back to the early days of the republic, and mid-decade redistricting is basically a Republican invention upheld by Republican-friendly courts. The recent lineage goes back to Texas — but actually Texas in 2003 under Rick Perry and Tom DeLay rather than last year.

Democrats solved for the equilibrium

Personally, I think having fewer competitive seats and less representative districts is pernicious to representative government. But I don’t think something like Virginia is a particularly close call.

It’s completely obvious that you can’t have unilateral disarmament, and some measure of credit should go to Gavin Newsom (not usually my favorite politician) who kicked things off by fighting back instead of just complaining.

Democrats in Virginia understood what was at stake. The county-by-county margins were almost perfectly aligned with the 2024 presidential vote:

The closest thing to an exception was in the Northern Virginia counties shown in dark blue. These are home to the canonical “high-information” voters, with many residents who work for the federal government or government contractors. These counties were generally above both trendlines, meaning that in some suburban DC counties like Fairfax and Loudoun, the yes margin actually beat Harris’s over Trump last year.

Yes, consuming lots of political news often correlates with being highly partisan, but this is one case where naked partisanship helped steer Democrats in a more strategically optimal direction.

Perhaps because my degree was in economics, because I play a lot of games — OK, really just poker, but I play a lot of poker — and because I’ve spent a lot of time covering highly competitive environments like sports, my brain tends to want to “solve for the equilibrium”. If one “player” in the game makes a move, I assume that the other party will reciprocate by employing roughly their most self-interested strategy according to the tenets of game theory.

That doesn’t always yield flawless predictions about political behavior. If Democrats were perfectly rational, they’d probably have nominated someone other than Harris (although she was a big improvement over Joe Biden). And it’s hard to understand what Trump has gotten politically or otherwise from starting a war with Iran.

But when the incentives are simple and straightforward, voters usually behave rationally. And a democratic system — even a competitive authoritarian system like Hungary — is often a resilient way to express those preferences.

Don’t like this? Ask Congress for help.

In game theory terms, redistricting is a classic prisoner’s dilemma. The rational move is to “defect” and screw over your “partner” unless you have some means of coordination or cooperation. What most people don’t understand about the prisoner’s dilemma, though, is that it’s not a zero-sum game: the prisoners would be better off if they could trust one another. But if they can’t, it’s every man for himself.

I think it goes without saying that even if Republicans “started it”, we’re locked into a race-to-the-bottom now given extraordinarily high levels of partisanship. In this cycle’s redistricting wars, there have been a few pockets of resistance: Indiana for Republicans and Maryland for Democrats. And there are a few voters like Glassman who will take a principled stance.

But not many. If the parties can’t trust one another, Texas and Virginia are what you’re going to get until and unless there is some enforceable means of coordination.

Share

And that means regulation. It’s not going to come from the Supreme Court, or at least anything resembling SCOTUS’s current construction. In 2019, with Ruth Bader Ginsburg still on the court, the Supreme Court ruled 5-4 in the Rucho vs. Common Cause decision that control of districting was a “political question” — in other words, up to Congress.

I don’t think it’s completely crazy to think that bipartisan legislation on redistricting could someday pass Congress. In the abstract, gerrymandering is still quite unpopular with voters. It might grow more unpopular when voters in Texas and Virginia realize in November that their longstanding member of Congress has been displaced because districts have been changed and they’re now represented by someone in a completely different part of the state.

Furthermore, it ought to be unpopular with current elected officials, since they’re often unseated by aggressive redistricting.

But it’s tough. You need a system that 1) both parties agree is roughly fair, 2) that will withstand legal scrutiny and 3) doesn’t have too many loopholes.

It’s honestly the kind of problem that some sort of bipartisan good government group might want to get to work on, understanding that partisanship will probably have to retreat from its current levels before some sort of solution is actionable. And while I’m more optimistic than most pundits about the future of American democracy, I’m not asking to volunteer.

This is SBSQ #31, the Greg Maddux edition, named in honor of one of the most cerebral athletes of all time. So even though the headline item is about politics, of course we’re also going to take a nerdy question about baseball. You can leave questions for the Magic Johnson edition (SBSQ #32) in the comments below. But for today, let’s answer these three:

• Why aren’t Democrats further ahead on the generic ballot?

• Trump’s second term and the forest for the trees, a.k.a. what happened to the 113 Trump predictions follow-up article?

• The game theory of baseball’s new automated strike zone

But before we get to the questions, I have an announcement to make:

More models coming soon!

I’ve teased at this in the past, but I can now confirm we’re going to have a Silver Bulletin World Cup model. Not only that, but we’ve also developed a soccer rating system, PELE1, which will serve as the basis for those projections.

This isn’t my first soccer rodeo. I first designed a soccer model, Soccer Power Index, for ESPN in 2010, and then we had SPI 2.0 at FiveThirtyEight, one of our most popular products. But I wanted to start fresh on this. There’s more and better data available now. And I’ve become more experienced at building sports models, having figured out some insights while working on COOPER and ELWAY recently. Furthermore, AI coding tools tend to help with the parts of the job I’m mediocre at (i.e., programming) in a way that frees up bandwidth for thinking about model architecture — international soccer is quite tricky for a number of reasons2. Though I think the LLMs are a long way from being able to build a good model on their own, look for a post about that soon.

In any event, I finished PELE last night. (About a week or so ahead of schedule; that doesn’t usually happen.) We’re going to take a couple of weeks and work on the graphics and presentation before we launch it sometime in early-mid May. And then we’ll publish the World Cup model in late May or early June and update it every day after matches are played. The World Cup itself starts on June 11. We’ll keep PELE running even after the World Cup is done — the international soccer “season” never truly ends. Most of the code stack can also be incorporated into club soccer projections, but that’s (much) further down the line.

The World Cup stuff being on track also means I’m feeling less stressed about getting our midterms model launched reasonably soon also. I don’t think it’s prudent to launch it before early June. Candidates aren’t locked in yet in some key races like Maine and Texas. That we could work around, but districts aren’t even locked yet, either. Long story short: look for the midterms model at some point in June or July.3

Subscribe now

Still, I will need to steal some time here and there to finish up these models. So posting volume may be slightly lighter in the weeks between now and roughly Memorial Day. (As it also was for the past week or so, frankly.) We’ll make it up to you with what we hope is a very busy June through November. Onto the questions.

Why aren’t Democrats further ahead on the generic ballot?

Alec Schneider asks:

SBSQ #31, two polling questions: 1) You’ve mentioned Dems will receive a boost when pollers switch from RV to LV screens however Trump approval among LV is higher than RV according to the SB Trump Approval Dashboard, could you square these? 2) What do you make of Trump approval cratering while there is little movement in the SB Generic Ballot? Is it just that a couple highly rated polls from before or early in the Iran conflict prior to Trump approval decline (TIPP and Noble) are buoying the R top line number?

Let’s take a quick snapshot of where things stand, Alec. Trump’s approval rating actually again hit a new second-term low in our tracking today, at 39.3 percent. But Democrats are ahead by “only” 5.7 points in our generic ballot average. I put “only” in scare quotes because that’s probably enough for Democrats to take the House anyway, and make the Senate competitive. But Democrats will probably need a couple more points than that to give themselves even or better odds in the upper chamber. (Or at least, that’s what I assume our model will say. Sometimes it surprises me.)

Trump is the second-most-unpopular president out of the past nine midterms: he’s actually more unpopular now than he was in 2018, when Democrats gained 40 House seats.4 But at this point in 2018, Democrats led by 7.6 points, not 5.7: not a huge difference, but exactly the sort of difference that might tip the balance of the Senate. Moreover, the only president more unpopular than Trump 2.0 was George W. Bush in 2006, and Democrats had a whopping 11.4-point lead as of Apr. 21 that year.

So let’s put this data on a scatterplot. Is this year some sort of outlier?

Read more

If you’re an avid listener of the Bill Simmons podcast1, you might have heard a recent bit he did with Zach Lowe, where they pondered whether this was the NBA’s weirdest season ever. For what it’s worth, excluding global pandemics, that claim has a lot of merit. The Clippers started the season near the bottom of the league amidst allegations they paid Kawhi Leonard 28 million dollars under the table; a gambling scandal led to the arrest of a coach and a player on the third day of the season; Bam Adebayo scored 83 points in a game; and fully a third of the league was literally trying to lose games.

It’s hard to say whether the weirdness has undermined the NBA’s effort to rebrand the league in a forward direction. With the league moving most of its games to NBC and Amazon Prime, the presentation has focused less on its legacy stars — Steph, Durant, LeBron — and more on its insurgent young core. The guy everyone wants to watch in the playoffs is Victor Wembanyama, not LeBron. But the shift has been backed by an influx of talent. I spent a lot of time analyzing this while building out our NBA draft model, PRISM, and it was hard to miss just how much talent has entered the league over the last few years.

This year’s Rookie of the Year race has been one of the best arguments for that explosion of talent — and, fittingly, one of the weirder ones we’ve seen in a while. Kon Knueppel and Cooper Flagg, former Duke teammates and roommates, traded the #1 spot on NBA.com’s Kia Rookie Ladder all season long. Flagg, the No. 1 overall pick (and the namesake for our college basketball model) has the highlights — 51 points against Orlando, a 42-point game in Utah before he turned 19, the kind of explosive, load-bearing performances that typically lock up the award. Knueppel, taken three picks later, has been the quieter storm: he leads the league in three-pointers made while shooting 43 percent from deep, and played a leading role in Charlotte’s leap from 19 to 43 wins. Flagg is ahead in counting stats per game across the board, but Knueppel has the efficiency and wins.

Subscribe now

Is the stats case even close?

There have been attempts to flatten Knueppel's case into some version of "he shoots threes really well," which ignores the off-ball gravity, passing, defensive IQ, and the way he's made Charlotte's entire offense function differently. Flagg’s case is also bigger than “he’s dominating the box score on a bad team” and people should be accounting for poorer context — the chaotic situation in Dallas — which makes it harder for him to produce at the same efficiency as Knueppel.

This framing is not so uncommon in awards discussions — the player with higher counting stats against the player with higher efficiency — and this type of discourse tends to get fans riled up, which is probably why mainstream outlets love to play into it. Analytical types tend to frame basketball around impact, which cuts straight through the noise of trying to calibrate all of these stats. While I don’t think ranking players is as cut-and-dried as just taking some aggregation of their impact metrics, there isn’t really an argument for Flagg from a pure stats point of view. On a per-possession level, he lagged far behind Knueppel.

Voters aren’t answering the same question

Of course, there is an argument against advanced metrics here that isn’t just luddite pessimism — stats like EPM, part of a family of RAPM-type impact stats, don’t always do the best job of distinguishing between a player’s impact and the difficulty of their role. In fact, just the opposite: RAPM and its descendants measure how team-level point differential shifts when a player is on vs. off the court and implicitly tend to reward being on good teams. That’s genuinely useful, but it tends to measure how good a player is at the thing they’re currently being asked to do — as opposed to either how rare the thing they’re being asked to do is, or what they might be able to do in another context.

Knueppel’s role in Charlotte — moving without the ball, spacing the floor, catching and shooting threes within a functional offensive system — is considered an easier adjustment in the NBA. He averages under two dribbles per touch — whereas Flagg’s role in Dallas, which involves a lot of primary creation, high-usage shot generation, and running the offense as an 18-year-old on a team with no other reliable engine, structurally suppresses the efficiency numbers these models care about.

For what it’s worth, Rookie of the Year voters traditionally do care about how much of an advantage creator you are. Knueppel would have the lowest on-ball percentage of any Rookie of the Year since Karl-Anthony Towns won the award in 2016.

So, yeah, EPM can tell you that Knueppel's minutes are technically more productive than Flagg's, but that doesn't settle the vote the way some analytics people want it to. The strongest version of the Flagg case isn't really about counting stats or highlights. Rather, it's a reframing of what the award is measuring — that Rookie of the Year should go to the player who has adjusted to the NBA best, and under that lens, Flagg's case sharpens: he's 19 years old, carrying a bad roster as its primary creator, producing against elite defenders scheming against him every night, and doing it at a level only LeBron, Luka, and a handful of other all-time teenagers have matched.

Still, that framing is building in some implicit credit for potential — Knueppel would be a very valuable player with or without rookie status, while we can’t definitively say that about Flagg this year. And even if we adjust for playtype difficulty, Kon is still ahead of Flagg in efficiency.2

But some Rookie of the Year voters aren’t necessarily even answering the question of which rookie was the best player so much as taking their mandate to mean “who will be the best player?” — and under that criterion, Flagg takes the cake.

Advanced stats like eWINS are admittedly not for everyone. They expose some inconvenient truths: rookies usually stink on defense, and their shiny, high-volume box score production is often paired with middling (at best) efficiency. The arrangement is a win-win — the player gets seasoning and sneaker deals, while their teams remain comfortably within the lottery — so long as you don’t ask too many questions of the data. And most voters don’t. Jae’Sean Tate, for example, had the highest eWINS in 2020 and got zero first-place ROTY votes — most of his value came on the end of the floor voters aren’t paying much attention to.

This year’s race only halfway matches that template. Kon and Cooper have been roughly similar defensively by most advanced metrics — and both are rated as about league average, which is unusual and promising for a rookie. Yet Flagg carries the reputation of an elite defender because without enough signal in the box score, the media is defaulting to their priors about his long-term defensive projection.

There’s also a more subtle issue: when voters treat ROTY as a forward-looking judgment rather than a full-season retrospective, they naturally time-decay the regular season. Victor Wembanyama was clearly better than Chet Holmgren by the end of 2022-23, and voters treated that midseason leap as license to wipe away his early-season struggles.

All of this would be forgivable — maybe even commendable — if voters were actually good at picking the best future player. The problem is they aren’t: take a look at that table above. Tyreke Evans over Steph Curry? Or even Kyrie over Kawhi? ROTY voters have often systematically underrated the lower-drafted guy.

The prediction markets won't sit still

But what makes the race genuinely weird is a string of late reversals in the conventional wisdom — and an unexpected twist involving former ROTY Luka Doncic. When I first planned on writing about Knueppel, I was confident he’d more or less locked the award up, and as of a few weeks ago, the prediction markets agreed, giving him a 94 percent chance at the trophy. Then Flagg put together back-to-back statement games, and Knueppel’s odds cratered to as low as 21 percent.

Silver Bulletin has covered prediction markets extensively in the past.3 They’re often very smart, and they’re certainly not easy to beat. But as Nate wrote about the election of the new pope last year — Leo XIV had been trading at only 1 percent odds — they aren’t necessarily at their best when trying to anticipate the behavior of a cloistered group of insiders. Polymarket’s ROTY contracts have millions of dollars in volume, which don’t exactly approach something like its presidential election markets, but are also not the kind of market you’d expect to flip 50 points in a day.

The first thing worth flagging is that we actually do have information about how voters feel. Only a few days before Flagg’s big weekend earlier this month, ESPN conducted a straw poll of 100 media members, some of whom are in the pool that actually votes on the award. Knueppel received 80 first-place votes to Flagg’s 20. So we have a prior suggesting as of just two weeks ago, Knueppel had a decent lead over Flagg — a lead hard to eclipse even with a full week of regular-season dominance.

Tim Bontemps, the reporter who conducts the straw poll, said something interesting on the Hoop Collective podcast the other day — that an 80-20 split in a room of 100 voters doesn’t mean each individual voter is 80 percent sure. It’s more like a collection of 60-40 decisions that mostly broke the same way. That’s a good point, but let’s analyze it. The lower bound on what 80-20 means is that every single Knueppel voter was close to 51-49 and just barely tipped his way. The upper bound is that all 80 were completely certain. The vote count alone can’t distinguish between those worlds, and Bontemps is right that we shouldn’t confuse margin of victory with depth of conviction.

But we can actually test what happens between the straw poll and the real thing, because Bontemps has been running the MVP version of this poll since 2017. In seven seasons of MVP straw polls, the leader has won the actual award every single time. In 2021-22, Nikola Jokic’s straw poll lead of 62-29 widened slightly to 65-26 in the real vote. In 2020-21, his 89 percent straw poll share held almost exactly at 90 percent. The two most recent seasons did narrow slightly — Jokic went from 85 percent to 80 percent in 2023-24, and SGA went from 77 percent to 71 percent in 2024-25 — but neither came close to flipping.

The bull case for Flagg probably comes from 2022-23. That year, Joel Embiid led the straw poll, but it was essentially a tie: he actually had two fewer first-place votes than Jokic. When the actual ballots came in, Embiid won 73-15 in first-place votes. Despite the close straw poll, voters didn’t scatter — rather, they broke hard in the same direction.

Steve Aschburner, the NBA.com writer who runs the Kia Rookie Ladder and is himself a voter, published his final ballot on April 8th with Knueppel at No. 1, and directly addressed the weekend performances that swung the odds: the 96-point two-game stretch swayed oddsmakers, he wrote, but didn’t eclipse Knueppel’s body of work. Historically Aschburner’s Rookie of the Year ladder has correlated strongly with the actual vote share — his one recent miss being Evan Mobley over Scottie Barnes in 2022, which was decided by 5 first-place votes, the narrowest under the current format. Bill Simmons has also rallied around Knueppel. Two voters, neither of whom are necessarily advanced-stats truthers, have come out with their vote, and they’re sticking with Kon.

That raises another issue: Could there be some form of insider trading? It’s obviously not an out-of-bounds question. Still, the scope for it might be limited on this market — no single voter’s private knowledge of their own ballot moves the needle much in a 100-person electorate, and any conspiracy between media members to collude on voting would be difficult, especially before the ballots come out. Building sharp models in awards markets is also difficult — these are one-off events with small voting bodies, so there’s no clean way to build a model for an event where the criteria can shift with narratives and storylines.

The market has corrected multiple times since the initial flip, with Knueppel briefly reclaiming favorite status before Flagg snatched it back over the last few days. Kon didn’t help his case in the Hornets’ dramatic play-in win on Tuesday — he had one of the worst games of his career, capped by a late-game benching. Despite being the NBA’s leader in threes as a rookie, he was absent from the floor over the last seven minutes, even when the Hornets desperately needed a three to send them into overtime.

And yes, even though the play-in game isn’t technically part of the regular season — statistically, it exists in purgatory, since it isn’t part of the postseason either — it’s going to bleed into voter perception. Because of Luka Doncic’s minimum games appeal, ballots were delayed, and Knueppel’s performance on Tuesday might empower some of those wishy-washy 51-49 Kon voters to go with Flagg, even if that’s technically not within the criteria of the award. The market certainly believes that voters will be swayed — following Kon’s game, Flagg is now inching toward being almost a 3:1 favorite.

The prediction market feedback loop

It would be easier to map out this race if close Rookie of the Year races happened more often. Over the last twenty years there have only been two nail-biters — Scottie Barnes over Evan Mobley in 2022, which came down to a 15-point margin and is still the closest vote since the current format started, and Evans over Curry in 2010.

Everything else has been a blowout, or close enough to one that the discourse around it never really mattered. Wemby was unanimous, KAT was unanimous, Lillard was unanimous, and even the races that felt close at the time, like Ben Simmons and Donovan Mitchell in 2018, turned out not to be. Perhaps the closest precedent is 2007-08, when Kevin Durant took home the trophy despite significantly worse advanced stats than Al Horford, who helped lead the Hawks to the playoffs as an 8th seed. That would seem like bad news for Knueppel, but the voting body is different now, in that they’re more receptive to the use of advanced stats than in 2008.

So, historical analogue or not, when people look at the current market and say it's behaving weirdly, part of what they're picking up on is that we don't really have a reference class for what a close ROTY race is supposed to look like, especially between two players with such different profiles. Some of what looks like noise is just what happens when a binary outcome gets close enough that small inputs start to matter: every new game has more leverage in a way that doesn’t seem rational on the surface.

Of course, that’s really just another way of saying buzz is dominating a race that, on the merits, might not be that close. And maybe this is where the whole thing gets a little recursive. The “buzz” isn’t just a reflection of the race, it’s also an input into it. On one side you have a player who’s had a great rookie year, and on the other, a guy who looks like a future superstar, even if he was an outright negative earlier in the season. Those are different kinds of cases, and voters trying to be rigorous about their ballot are still going to get nudged by which story feels more compelling in a given moment. A 30-point Flagg game in March probably carries different weight than a 30-point Kon game because of what it implies about the next 5, 10 or 20 years.

Prediction markets potentially feed into the same loop. They’re aggregating information, but in a race like this, most of what they’re aggregating is the buzz itself — which podcasts said what, which highlights went viral. You know, the vibes. It’s not that prediction markets are broken per se, it’s that there’s just not much independent information for them to work with — so they end up measuring the zeitgeist and handing it back with the authority of a price.

Now, do I think Kon is still a good bet at 28 percent? I think so, because the vote trackers show a fairly close race, and the priors for him are pretty strong. Still, the ballot delay gives recency bias more room to breathe than it usually gets, and if Kon has another rough night tonight it’s going to get harder to shrug off his doubters. Two weeks ago I would’ve told you this race was over. While the underlying evidence from an 82-game season hasn’t changed very much, if Flagg is holding the trophy in a few weeks, I will no longer be surprised.

It was a good, free-flowing conversation, but we wound up spending the first half of the chat on AI, including the recent violent attacks at the home of Sam Altman, the OpenAI CEO. So we didn’t get into quite as much detail about the mechanics of the Twitter algorithm itself; today’s newsletter is intended to remedy that.

Twitter was always a better news hub than “social media” platform

I’ve written a lot about Twitter and its competitors at Silver Bulletin (see here, here, here and here) and I’d like to think I have a pretty nuanced critique of the platform. Even so, it can be easy to get one’s timelines scrambled about exactly which changes were made when.

The point of last week’s post was that relatively subtle changes to the algorithm can make huge differences to the end result. On a case-by-case basis, I think many of the changes that Elon Musk and X’s head of product, Nikita Bier, have made to the algorithm are defensible or even good.1

The cumulative effects, however, have been both to promote slop and to greatly reduce the amount of traffic to publishers like the New York Times who mostly use the platform to link to offsite articles. Some of this is because the Twitter algorithm has, at times, punished tweets that contain offsite links through a variety of implicit and explicit methods. (The extent to which it still does this is disputed and is constantly changing.)

Probably the more important factor, though, is that X now very strongly defaults people to the algorithmic “For You” feed instead of a chronological feed of the accounts users signed up to follow. That’s why the NYT feed often only gets around 70,000 views on a tweet even though it has 53.2 million followers. It’s not (just) that fewer people are clicking on tweets with links. Rather, it’s that users are mostly seeing the algorithmic feed, and the @nytimes account doesn’t tend to have a lot of “voice” or to trigger the other “engagement” signals that the algorithm rewards — particularly not replies, which are valued much more heavily than “likes”. It’s also not clear how much passively following an account is taken as a positive signal, if at all, even though this reflects an explicit preference on the part of the user.

This greatly reduces Twitter’s value as a news aggregator. It can be almost useless for following, for example, developments in Iran, where you might mostly be interested in reporting and reliably sourced hard news (and not necessarily interested in commentary, especially from non-experts). And this is a shame, because Twitter used to be an absolutely unparalleled way to “monitor the situation”, essentially the social media equivalent of London Heathrow, with departure points to everywhere else that you might want to visit.

Subscribe now

What happened within the platform — an anonymous account would yell at a reporter like Maggie Haberman, and she might even respond back — might be a value-add to some users and annoyance to others. Of course, once you're on Twitter posting links and promoting your outside work, you might get swept up in it. (I’m a prime example.) But it wasn’t necessarily the reason you were there or the most differentiating feature, any more than you go to Heathrow to visit the BA Lounge.

Did it all start going downhill in March 2016?

However, the switch to the algorithmic feed didn’t happen under Elon. Instead, the algorithmic feed became the default in March 2016, part of a series of efforts to increase “time on site” and make it more attractive to advertisers. (Twitter has never been a particularly good business despite its once-very-high “mindshare” and capacity to serve as a hub to “drive the conversation”.)

The timing is interesting in that early 2016 was about the time when Twitter became a much less pleasant place for anyone like me who tends to have sharp elbows and abhor groupthink. This was also an inflection point for certain progressive-coded pieties, a.k.a. wokeness. Twitter was probably not the main reason that progressive and even neutral spaces suddenly became much more woke2, though I suspect it contributed to it — and that changes to the platform, including the addition of the quote tweet in 2015, further catalyzed some of this behavior.

Here’s the thing, though. The content that people engage with in the form of things like comments and quote tweets isn’t necessarily the content they find the most valuable — especially not in the long run where it’s exhausting if the volume is always turned up to 11.

Opening up the Silver (Bulletin) curtain

I recognize that Silver Bulletin is an unusual newsletter in many ways. We cover a few topics intensively — elections, sports, and sometimes tech/media/AI stuff — when most other newsletters are either broader or narrower. We have a mix of very data-driven stuff and other work that’s more take-y. We have a slower publishing cadence but with longer, more detailed articles than most of our competitors. And in contrast to most other Substacks, only about half of our overall views come from people reading in email; we’re really using Substack as a hybrid of a newsletter-delivery system and a web publishing platform. Fortunately, enough of you seem to like it — or at least like enough of it — that it’s a pretty good business.

Share

Nonetheless, we do have a lot of fairly reliable data on Silver Bulletin’s engagement metrics — something hard to come by for anybody else’s stats — so let me share some of it with you.

Substack basically gives you seven different metrics to play with:

• Likes

• Comments

• Shares (within the Substack ecosystem)

• Total views (web pageviews + views in email + views on the Substack App)

• Email open rate

• Paid subscriptions

• Free subscriptions

Let’s take everything Silver Bulletin has published in the past six months and see how they compare. I originally ran these numbers back to the start of 2026, but on second thought, I went ahead and included everything since Oct. 14, 2025 to introduce a little more variety in the form of things like the Nov. 2025 elections, the endgame of the government shutdown, some posts that touched a nerve with libs rather than conservatives3, the launch of our NFL model ELWAY4, plus a few things that were just plain weird.

The first five metrics are all fairly strongly correlated with one another, especially likes and shares, which almost exactly match. (The correlation coefficients shown here run on a scale from -1 to +1 from perfectly inversely correlated to perfectly correlated.) The one partial exception is email open rates, which may reflect the fact that people who read us in their email browsers tend not to engage in these other behaviors that are easier to do on the web or in the app.5

But how do these measures correlate with the number of paid subscriptions generated? Not very well at all.

If you look at every Silver Bulletin post over the past six months, the correlation between “engagement” metrics and paid signups is literally almost zero or, in some cases, even slightly negative. Now, I certainly wouldn’t say that paid subscriptions are the only thing we care about: if that were true, we’d probably paywall a (slightly6) higher share of posts than we do. I’d even like to think some of the things we publish are in the public interest in various ways, and those are almost always free. But paid subscriptions are unique in that they very literally require a higher commitment than any of these other things; it’s never been particularly easy to get people to pay for content or to go through the mechanics of punching in their credit card information.

To be fair, there are several confounders here. Our models7 often do disproportionately well as a driver of paid subscriptions and can also produce a disproportionate number of pageviews since they’re updated repeatedly. There’s inherently some trade-off between satisfying your existing customers (as measured by things like open rates) and expanding your reach to new ones. Most importantly, the correlations do turn positive — although they remain low — if you disaggregate between paid and free posts. Even though comments are always limited to paid subscribers at Silver Bulletin, free subscribers are also less likely to like or share posts that are paywalled.8

But the impulse to like a story and to pay for a story don’t really come from the same place. Liking and sharing posts, especially on Substack where likes are public, are almost definitionally a social behavior. Posts that get the most “engagement” on Silver Bulletin are often posts like this one or this one that contain a relatively simple hypothesis that is revealed by the headline/subject line and are broadly agreeable to a portion of the Silver Bulletin audience.9 But even though free posts can induce more paid subscriptions than you might assume, these high-engagement posts often do just OK relative to others.

In contrast, highly detailed deep-dives like our NBA draft model often do quite well for producing paid signups, even if it’s pretty lonely in the comments section. Many Silver Bulletin readers aren’t interested in things like our women’s March Madness predictions at all, but those who are might be quite intensively interested, enough that they’ll pay for distinctive work.

Now, to complete the story, free subscriptions do look much like the other metrics:

Still, it’s interesting that comments — equivalent to the replies that the X algorithm values most — are quite inferior to likes, shares and views even as predictors of free subscriptions. As you’re probably aware, here are some common templates you might encounter in Silver Bulletin comments:

• There’s a typo in the first paragraph

• Nate is an idiot

• The other commenter is an idiot

• WHY DOES ELWAY HATE THE CHICAGO BEARS?!?!

Don’t get me wrong: the first category is genuinely constructive and helpful. Still, you sometimes get more comments because you 1) screwed something up; 2) touched a nerve that you didn’t necessarily intend to touch; or 3) “disappointed” (i.e., pissed off) a portion of your user base because your take was half-baked. All of those are negative signifiers that a dumb algorithm might mistake for a positive one.

I was going to say that there’s a simple lesson here — don’t chase engagement. But it’s a little more subtle than that: don’t chase engagement for its own sake. Even if it sends some positive signal on average, it doesn’t always, and it deteriorates when you over-optimize for it. (See also Goodhart’s Law: “When a measure becomes a target, it ceases to be a good measure.”)

Especially for web content, many of your best customers don’t “engage” with it in any way other than in the most important way: by reading it. (And hopefully subscribing before long.) At Silver Bulletin, for every 1000 people who read a post, only about one likes, shares or comments on it. A platform that doesn’t value this “silent majority” of customers can become an algorithmic hellscape before long.

A few weeks after Donald Trump’s second presidential win, I took the train up from London (where I was living at the time) to Oxford to attend a conference on polls and forecasts of the 2024 election. Most of the attendees were pollsters or academics, but I also watched presentations from Aaru and Electric Twin, two companies that do what is interchangeably called synthetic sampling, silicon sampling, or creating synthetic audiences. Sans startup jargon, that means they use large language models (LLMs) to simulate responses to public opinion polls by having AI agents take on the role of survey respondents.

I had already heard of Aaru thanks to some articles with eye-catching headlines like “No people, no problem: AI chatbots predict elections better than humans” in the months leading up to Election Day. The guys behind the company were making some big, some might even say far-fetched claims, such as: “within two years, we will simulate the entire globe — from the way crops are grown in Ukraine to how that impacts production of oil in Iraq, trade through the strait of Malacca, and elections for the mayor of Baltimore.” When Semafor asked Aaru’s cofounders — Cameron Fink and Ned Koh — about my boss, they said “we respect all those who came before us.” Nate (as he so often does) shared his thoughts on Twitter:

Fink and Koh were relatively good-natured about this back-and-forth when we spoke at Oxford. They even offered to mail me one of the t-shirts featuring Nate’s quote they apparently had made. I never took them up on the offer, which I now somewhat regret.

These synthetic sampling companies fell off my radar for a while, but they do still exist. In fact, Aaru recently received a $1 billion valuation. Is what they’re doing anywhere close to the most important frontier in AI development? Not by a longshot, especially when Anthropic just developed a model so adept at exploiting software vulnerabilities that it’s only being released to 40 companies.

Still, silicon sampling is increasingly finding its way into public polling. Axios reported in March that “a majority of people trust their own doctors and nurses” based on findings from Aaru — without mentioning that the “people” in that sentence were actually LLMs. Around the same time, the Public Sentiment Institute “boosted” their online sample of 373 real survey respondents with 114 AI agents.1 (Spoiler alert: even the co-founder of Electric Twin doesn’t think that’s a particularly defensible approach.) Polling companies like Qualtrics and Ipsos are also developing synthetic data panels.

So, what should we make of these … “polls”? Let’s get one thing out of the way: whatever they are, they’re not polls in the way that term is usually defined.

Subscribe now

You can’t replace polls with AI

On one hand, using LLMs to essentially make up fake survey respondents sort of sounds like the dumbest idea ever, one that will only imperfectly replicate real polls while introducing all sorts of biases. On the other hand, with LLMs improving at a remarkable, perhaps even alarming rate, maybe I’m a dinosaur at the ripe old age of 24 because I still want to rely on polls that talk to actual people.

I’m not going to argue that synthetic samples are completely useless. In fact, as I’ll return to later, there is evidence that some techniques can replicate topline survey results quickly and cheaply. But the marketing from certain companies can be slightly optimistic. “No traditional poll will exist by the time the next general election occurs,” said Fink in 2024. We’re just 206 days away from the midterms, and based on the fact that I still have to collect a bunch of polls every day, I’d say he should have run that prediction by a sample of AI agents before the interview.2

To see why synthetic samples can’t replace polls, here’s a quick primer on how they work. The simplest version of these models involves taking a LLM (like ChatGPT or Claude), giving it a demographic profile (e.g., a white, college-educated woman who lives in Utah and makes $70k a year), and then asking it to respond to a survey question. You repeat that process a few thousand times using different demographic profiles and end up with a sample of synthetic survey responses.

The actual models used by private companies are more sophisticated than this, usually because they incorporate more demographic characteristics for each agent and provide them with extra information. Aaru, for example, feeds agents a diet of news and information they’d be likely to consume, while Electric Twin incorporates their customers’ proprietary data about the audience they’re trying to replicate. The way Ben Warner, the co-founder of Electric Twin, explained it to me was “we have a large amount of data on… for instance, 5,000 people. Can we make an accurate prediction of how they would respond to another question?”

Still, without any reference to cost, speed, or accuracy, it should be obvious why synthetic samples can’t replace polls. Polling is fundamentally a data collection process. We might use surveys to make predictions by feeding them into election forecasts, but the main purpose of a poll isn’t prediction, it’s gathering new data about what people think and how they feel. Silicon sampling, on the other hand, produces no new data. It’s simply a model: you input LLM training data, demographic prompts, and a bunch of other information, and it spits out a prediction for what a poll would say.

We love models here, but models aren’t polls. That difference is an important philosophical sticking point for most pollsters I talk to. “I think politics should stay away from [synthetic sampling], because we’re trying to… represent the voice of the people,” said Natalie Jackson, a vice president at GQR Insights. Democratic pollster John Hagner told me “I think I’m just incredibly skeptical of this idea. I don’t think it’s research. At that point, you’re asking the machine to tell you what you already believe.” Hagner has seen some presentations of early synthetic sampling experiments, but so far, “if it’s being used in a campaign, people are keeping it incredibly quiet.”3

But Eli, I hear you saying, aren’t polls themselves increasingly governed by modeling decisions? Indeed they are: pollsters’ choices on which sampling method to use, how to define their likely voter models, and how to weight their samples can and do lead to dramatic differences in the results they publish. Aaru even referenced these limitations in the methodology statement included with that maternal mortality “poll” — although I’m using the term “methodology statement” loosely here, because it doesn’t really explain how the model works at all.

We can ignore the (frankly preposterous) implication that synthetic sampling isn’t subject to a separate set of biases. The important point is that there’s still a meaningful difference between using weighting and other statistical techniques on actual polling data and using a model to predict what a poll would say. The latter is far closer to election forecasts or techniques like MRP — potentially useful models, but not a replacement for polls.4

To be fair, other synthetic sampling companies are perfectly happy with the distinction between polls and models. Warner compared polling and synthetic sampling to different tools in a toolbox. “The mistake I think we make is we think that these new tools should either work in exactly the same way or somehow replace these old tools,” he said. “Rather than thinking of it as, okay, so we’ve always had the hammer, we’ve always had the screwdriver, now we’ve got a saw. But don’t use a saw to try [to] do the job of a hammer.”

Are these models any good?

If synthetic surveys are just a new type of model, the next obvious question is whether the models are at least accurate. The answer to that question depends very much on who you ask.

On one end of the spectrum, you have the maximalist argument that synthetic sampling is better and more accurate than actual polls. “It’s an incredibly challenging problem to go to someone and say ‘hey, we’re going to be more accurate at predicting human behavior than you, even when you talk to your customers directly,” Koh recently told CNBC. In his view, synthetic sampling isn’t a saw to polling’s hammer, it’s “magic.”

There’s certainly evidence that synthetic samples can replicate certain survey toplines. But if Aaru does have any examples of their approach outperforming the polls, they’re keeping those to themselves.5 Aaru’s 2024 election model, for example, had Kamala Harris leading in Michigan, Nevada, Pennsylvania, and Wisconsin on November 4th. And although they’ve since taken down their forecast page, they gave Harris a 50.5 percent chance of winning the race on November 2nd.6

After the election, Fink told Semafor he was happy enough with those results because they were “within margin of error,” a term that is meaningless when applied to a “sample” of AI agents. And of course, Aaru says their models have improved since 2024, so supposedly now they’d be more accurate than the polls? Still, their stronger argument is on cost: “We are significantly faster and cheaper than traditional polling, and still more accurate,” said Fink. The first two claims there are undeniably true, but the third brings us to the opposite end of the spectrum.

Both Jackson and Hagner are skeptical that these models are reliable for anything far beyond replicating common survey toplines. “I just… don’t think the machines are what we want when we’re looking for nuanced views. My example on this is people in Arizona and Nevada in 2024 who voted for Trump and voted for expanding abortion in their states on ballot initiatives,” said Jackson. Hagner identified a similar issue: “the reports that have come through at the meetings that I’ve been at are that the early experiments on this, they cannot get respondents to be as racist or sexist or, frankly, as negative as human respondents.”

Academic research mostly agrees on this point. While there are some papers that show promising results when using LLMs to replicate polling data, most show that LLMs suffer from various quirks like producing too few “don’t know” responses and can seriously overpredict the favorability of politicians like Donald Trump and Kamala Harris. They also seem to struggle with too little variation between demographic subgroups, so the difference in predicted opinion between Democrats and Republicans, for example, is too small.

When I asked Warner about these studies, his response to these papers was that just because academics can’t get synthetic sampling to work doesn’t mean that the technique doesn’t work in general. “Actually, the argument is, okay, yours does not [work]. That does not mean […] for this complex set of machinery, which uses a lot of investment, a lot of time, a lot of money, you can’t get it to work.”

Cards on the table, I’m somewhat sympathetic to this argument because academics aren’t exactly great at making election forecasts. Usually, the people with skin in the game are the most accurate. Warner’s argument is that the approach Electric Twin takes — which includes, for example, making multiple predictions for each synthetic respondent using different models and prompts and subsequently averaging those to get a final prediction like a sort of ensemble forecast — produces better results than the simpler academic models.

Warner shared a comparison between his method and the method from a recent academic paper with me, and Electric Twin was indeed able to get more accurate replication. But even still, he acknowledged that synthetic sampling “is not a crystal ball.” “If you asked me, do I think using other data sources will be more accurate than asking somebody who they will vote for, I would probably say no. But if you asked me ‘would your system be useful for our turnout modeling today?’ I would say yes.”

But for better or worse, it looks like the method is already getting more popular in the market research world. Most of the clients Aaru talks about these days are businesses like EY and McDonald’s.

That’s not to say AI won’t pop up in other parts of the political polling process. Pollsters are already using it to code open-ended survey responses, and some firms, like YouGov, are testing using LLMs to ask survey respondents questions.

More worryingly, one danger to actual polls is that AI agents can be used to infiltrate online surveys. Most online polls use various checks to prevent that from happening, but there’s conflicting evidence on how effective those filters are and how prevalent AI agents currently are in online panels. If those agents ever become impossible to detect, it might spell the end of online polling, but the solution isn’t to replace all of your respondents with ChatGPT.

I’m not super eager to talk about military strategy as opposed to the political implications of the war with Iran. But as someone with some expertise in strategy and game theory generally, I suppose I’d like to make a quick point here.

Over the past few days, President Trump had been making some extremely escalatory threats toward Iran. On Easter Sunday at 8 a.m., he wrote the following on Truth Social:

And then on Tuesday morning, Trump doubled down, claiming that “a whole civilization will die tonight” if Iran didn’t make some effort to reopen the Strait of Hormuz by 8 p.m. last night:

I didn’t want to put this idea out into the ether with a glib tweet or something until we got through last night, but this was, in my view, a not-so-thinly veiled nuclear threat. It was a little bit ambiguous, but nuclear threats usually are. And it wasn’t exactly that subtle. I mean, Trump literally said “a whole civilization will die tonight” and that this was “one of the most important moments in the long and complex history of the World”. Whether it was a credible threat was another matter. But I clearly wasn’t the only one who was thinking about the potentially existential stakes. Here was Garrett Graff, from the appropriately named newsletter Doomsday Scenario:

Is the chance that tomorrow at 8 pm ET Donald Trump launches nuclear weapons against Iran zero? Definitely not — and, regardless of whether that’s a one-tenth-of-one-percent chance or two percent or eight percent, anything more than zero is too high. I’d personally put the chance that Donald Trump uses a nuclear weapon against Iran at some point in the three percent range — which is a stunningly high number, given the history of nuclear weapons and the presidency.

I have no idea whether Graff’s estimate of a 3 percent chance of a nuclear strike against Iran was in the right ballpark or even how you’d go about calibrating such a forecast. But I don’t think Graff’s take was “alarmist”. There were really three scenarios: 1) Iran would back down; 2) Iran wouldn’t back down, but Trump would chicken out anyway; 3) Iran wouldn’t back down, and Trump would follow through on his threat.

I’m not sure whether we just experienced Scenario #1 or Scenario #2. But even if we had wound up in Scenario #3, a massive assault through conventional weapons was probably more likely than the literal nuclear option. And maybe if Trump had ordered a nuclear attack — or even something short of that — people in the chain of command would have regarded it as an unlawful order, refused to carry it out, and we’d be in some sort of constitutional crisis.

Even though we have pulled back from the brink — a two-week ceasefire was announced 90 minutes before the deadline — nuclear escalation had entered the realm of the thinkable, and that’s bad enough. I’ll tell you this much: I wasn’t eager to make plans last night.

I have a lot of respect for the intelligence and integrity of people who work in the military (not one of my more lib-coded characteristics). But Pete Hegseth has been firing senior commanders. He’s notoriously erratic and unqualified to lead the Department of Defense.

You shouldn’t feel too relieved

Sometimes, there’s an element of “heads I win, tails you lose” in media coverage of President Trump. He makes some sort of threat that every reasonable person thinks is a bad idea, and then he backs down. So he’ll be (appropriately) criticized if he follows through, but also criticized for chickening out.

Usually, it’s OK to feel some sense of relief when the worst consequences are avoided.

My point is this: I don’t think this is one of those times when you should feel particularly relieved. In fact, I think it’s pretty fucking irresponsible to either celebrate Trump’s strategic genius or troll him for TACO’ing.

Yes, a ceasefire is much, much, much better than the alternatives. No, I don’t know how long it will hold. Nor do I know how much Trump backed down as compared to Iran.

But this certainly isn’t some sort of 13-dimensional genius move. It reflects an unchecked Commander-in-Chief who was erratic on his best days, who is 79 years old, who was boxed into a corner, who has sycophantic advisors who are mostly too afraid to challenge him, and who once bragged that he could maintain approval from his base even if he murdered someone.

States with nuclear weapons have the ability to make a lot of bluffs. The expected value of a 3 percent chance of an infinitely bad outcome is still negative infinity. That’s why many countries, including, of course, Iran, have pursued the bomb.

And it probably does help Trump in some ways that he could be plausibly considered an irrational actor under “Madman Theory”, the term popularized by Richard Nixon. I doubt that a President Romney or President Haley or President Biden or President Harris or President Ocasio-Cortez would have used the same rhetoric, but if they had, it would have read differently.

There were, however, various ways this could have gone very badly, or still could go very badly the next time the “game” is played, and inevitably will go badly if we repeat the scenario often enough:

1. Even if he wasn’t eager to launch a nuclear attack, Trump could feel like he’d be humiliated or would lose credibility if he didn’t follow through. And having a reputation for chickening out might make those impulses worse.

2. Iran could have concluded the threat was empty — with the same overconfidence I’m seeing elsewhere tonight — when it wasn’t.

3. Even if Iran took the threat somewhat seriously, we don’t have a lot of evidence on how real human beings behave in scenarios like these. But there’s a pretty rich literature on how humans react very negatively to coercion and ultimatums and often seek revenge. Iran is also a literally decapitated regime — it’s not clear exactly who’s in charge — and might figure it didn’t have a hell of a lot to lose. And who knows about any political or religious motivations.

4. Even if nobody wanted further escalation, there’s inherently some risk in the fog of war. (See also: Thomas Schelling’s “Threat That Leaves Something To Chance”.) Some American sub-commander could have gotten his signals crossed, misinterpreted an order, or gone YOLO. Iran could have launched some sort of preemptive attack against American interests or against Israel (also a nuclear-armed state), and somebody might have panicked.

I’m not in the mood to get cute about this, so I’ll save the requisite poker analogy for the footnotes.1 But you don’t need to be some sort of game theory maven to appreciate that the small chance of a catastrophic outcome can outweigh the overwhelmingly likely but less than 100 percent chance of a trivial gain. And you need to understand that there can be some dangerous self-reinforcement from making “plays” like this because they almost always work. If you’re tempted to show the bluff and sit there with a smug smile on your face — “haha, suckers!” — that’s even worse because people will take you less seriously the next time around.

You do not want to play this game repeatedly. Nuclear weapons haven’t been around that long, and since Nagasaki, the world has been closer to the brink of nuclear weapons being used in combat than most people realize.

Subscribe now

We’re really doing double duty. See here and here for our most recent coverage of Iran. I’m also working on another story about the market’s reaction and the so-called TACO trade. But our NCAA tournament coverage is probably what Silver Bulletin is best known for outside of politics, and we’re not about to neglect it.

We launched the men’s version of our new COOPER ratings earlier this week. Today, it’s the women’s turn. As you’ll see, there are some much bigger discrepancies: for instance, the top women’s team — UConn, as usual — has an Elo rating of 2686 versus 2265 for the Duke men. For more about how all of this works, see the methodology page.

Our NCAA tournament forecasts will follow once the brackets are announced on Sunday. (Those should get out Sunday night for the men, but I’m guessing we’ll need until Monday for the women.) By default, new models like COOPER and our NCAA forecasts — whether sports or politics — are sent via email to all subscribers. However, if you don’t want to receive non-model emails about sports, here’s how to manage that in your inbox settings. For that matter, we also have some readers this time of year who only care about sports. If you’re one of them, you can opt out of politics-related newsletters at the same link — though you will get an email about our midterm election forecast once that launches just in case you’re interested. —NS 3/12/26

Subscribe now

This is the homepage for the women’s version of COOPER, Silver Bulletin’s new NCAA basketball rating system. This page is best viewed on the web rather than in your Substack app.

Named in honor of Cooper Flagg and Cynthia Cooper, COOPER accounts for wins and losses, margin of victory, team tempo1, preseason polls and conference strength. Ratings partly carry over from season to season, but COOPER uses an “impact factor” that weights results from recent games, close matchups, conference games, and NCAA tournament games more heavily.

COOPER represents an evolution of the SBCB ratings that we used in 2025, which in turn are a derivation of Elo ratings. Like SBCB and Elo, COOPER ratings are Bayesian in the sense that ratings are adjusted on an ongoing basis as new games are played. Basically, we compare actual game results against COOPER’s forecast for each game. If a team beats COOPER’s expectations, its ratings improve. For more on how COOPER works, see our methodology page. Although the men’s and women’s ratings are highly similar, there is some customization for the women’s game, such as to reflect greater team continuity from season to season.

Here are our current top 16 teams:

The rest of this page contains:

• COOPER ratings for all 363 women’s NCAA teams, tracking their high and low points over the course of the season;

• Offensive and defensive ratings for each team, along with strength of schedule ratings and custom home court factors;

• An alternative version of COOPER that only uses objective data and isn’t influenced by human polls;

• Extensive historical data in the form of season-ending COOPER ratings for every women’s basketball team since 2002-03, and;

• A spreadsheet showing how to translate COOPER ratings to win probabilities and projected margins of victory (i.e. point spreads).

These features are a benefit for paying subscribers. And of course, subscribers will also get access to our men’s and women’s NCAA tournament forecasts once those launch.

COOPER ratings for 2025-26

Here are COOPER’s current Elo ratings for all 363 women’s D1 teams, showing their highs and lows on the year, the “prior” we had for each team at the start of the 2025-26 season based largely on its ranking in human polls, and how its COOPER rating was affected by its most recent game.

Read more

Subscribe now

This is the homepage for COOPER, Silver Bulletin’s new NCAA basketball rating system. Named in honor of Cooper Flagg and Cynthia Cooper, it accounts for wins and losses, margin of victory, team tempo1, preseason polls and conference strength. Ratings partly carry over from season to season, but COOPER uses an “impact factor” that weights results from recent games, close matchups, conference games, and NCAA tournament games more heavily.

COOPER represents an evolution of the SBCB ratings that we used in 2025, which in turn are a derivation of Elo ratings. Like SBCB and Elo, COOPER ratings are Bayesian in the sense that ratings are adjusted on an ongoing basis as new games are played. Basically, we compare actual game results against COOPER’s forecast for each game. If a team beats COOPER’s expectations, its ratings improve. For more on how COOPER works, see our methodology page.

Here are our current top 16 teams:

The rest of this page contains:

• COOPER ratings for all 365 men’s NCAA teams, tracking their high and low points over the course of the season;

• Offensive and defensive ratings for each team, along with strength of schedule ratings and custom home court factors;

• An alternative version of COOPER that only uses objective data and isn’t influenced by human polls;

• Extensive historical data in the form of season-ending COOPER ratings for every men’s basketball team since 1949-50, and;

• A spreadsheet showing how to translate COOPER ratings to win probabilities and projected margins of victory (i.e. point spreads).

These features are a benefit for paying subscribers. And of course, subscribers will also get access to our men’s and women’s NCAA tournament forecasts once those launch.

COOPER ratings for 2025-26

Here are COOPER’s current Elo ratings for all 365 men’s D1 teams, showing their highs and lows on the year, the “prior” we had for each team at the start of the 2025-26 season based largely on its ranking in human polls, and how its COOPER rating was affected by its most recent game.

Read more

We mostly talked Trump and Iran. But my post on the decaying state of social media yesterday really touched a nerve over on X, so we discussed that at the top too.

It’s always telling when people…

Read more

Substack Live on Trump/Iran Monday at noon

I needed a break today from writing about Iran, Trump, and/or basketball, so let’s talk about another of my favorite topics: social media.

However, I will be conducting a Substack Live with Galen Druke at 12 noon tomorrow (Monday), where we’ll discuss what increasingly looks like both a foreign policy and a political crisis for the White House. You can join us on the Substack App in real time, or I’ll send the video out to subscribers later.

I’m also taking questions for SBSQ #31, which I’m hoping to get to sooner rather than later.

FiveThirtyEight (re)launched under Disney/ESPN in March 2014. In retrospect, our timing wasn’t great. Despite various missteps, we were trying to deliver a differentiated, engaging, high-quality product. But this was probably the single period since the web rose to prominence when the quality signal was least well-rewarded.1

Publishers believed that maximizing “reach”, particularly as measured by the number of monthly unique visitors, was the key to success. It didn’t matter how long someone stayed as long as they visited your site, for any length of time, at least once during the month. And no platform provided more opportunities to trawl the ocean depths for potential customers than Facebook, which was then close to the peak of its influence. Entire businesses like Upworthy were premised on gaming the News Feed algorithm, often through peculiar clickbait headlines such as “You Won’t Believe What Happened Next”.

You will believe what happened next: it didn’t work. The whole period was like the Underwear Gnomes meme come to life. Phase 1: Collect lots of low-quality traffic from Facebook. Phase 2: ???. Phase 3: Pivot to video.

Subscribe now

It didn’t help that Facebook was constantly tinkering with News Feed, and grossly exaggerating metrics like average time spent watching videos. But more fundamentally, it was locked into a zero-sum, adversarial relationship with publishers. Facebook wanted readers to stay within its walled garden, to spend as much time as possible on Facebook. Publishers, meanwhile, regarded Facebook as the equivalent of the Port Authority Bus Terminal: a miserable, liminal space where you’d hopefully spend as little time as possible before booking a one-way ticket out of town.

FiveThirtyEight eventually matured to the point where it got quite a bit of web traffic — our 2016 election forecast was literally the most “engaged-with” piece of content on the English-language Internet that year as tracked by Chartbeat.2 But we never got much traffic from Facebook. One of the major reasons for this, I later learned from talking with media friends who were much more successful at this stuff, was that Facebook tended to reward emotional sentiment in headlines: surprise and delight on the one hand, or outrage on the other hand. And that was pretty much exactly the opposite of our editorial goals at FiveThirtyEight. Instead, we wanted to encourage cool, analytical, nuanced reactions to things that people are ordinarily quite passionate about, such as electoral politics.

But every now and then, a FiveThirtyEight article would “go viral” on News Feed. When this happened, it seemed literally almost completely random. It wasn’t even particularly well-correlated with the headline. And it was inversely correlated with the depth and quality of the article. But here’s the thing: that “viral” traffic was almost worthless. A lot of it quite literally consisted of people who visited the site for 5 to 30 seconds, read a paragraph or two, and never returned.

Twitter, then also near the peak of its influence, was a comparatively better platform for us, often producing more overall engagement for FiveThirtyEight, even though Twitter was always much smaller. Back in the mid-2010s, Twitter rewarded newsworthiness, subject-matter expertise, and a certain kind of nerdy and snarky but relatively cerebral argumentativeness, all things we were pretty good at. Furthermore, the pre-Elon versions of Twitter were always surprisingly happy to let you direct traffic off of their platform.

By the late 2010s, Twitter would “evolve” in a direction that was more partisan, less pluralistic, and dominated by quote-tweets and dunks. For lack of a better term, it also became much more woke. The enforcement of groupthink was rigid, not unlike what Bluesky has become today. There was always a “main character,” someone who was the subject of the pile-on or the struggle session of the day. Your goal was supposedly never to become the main character. I spectacularly failed at this. Without exaggeration, I was probably a “trending topic” on Twitter (also something you never wanted to be) more often than all but a half-dozen or so other writers and journalists.

So, between my bad experiences with mid-2010s Facebook and late-2010s Twitter, I am elated in some ways that social media has become less and less relevant to media business strategy. For Silver Bulletin, social media is a rounding error. According to Substack’s internal dashboard, the share of Silver Bulletin traffic attributable to external social channels has consistently fallen to the point where it was just 0.7 percent of overall site views in March. And yet, the aggregate number of views of Silver Bulletin content increased by 40 percent from the first three months of 2025 to the first three months of 2026.3

Before you make too much of that data, or ask why we even bother to post to Twitter at all (in fact, we’ve started a Silver Bulletin X account just for shits and giggles) note that there are a number of caveats. Social media traffic is hard to measure. There’s direct traffic in the form of “dark social” where the source of origin is lost. There’s also indirect traffic in the form of the overall amount of buzz you might attribute to an article. Furthermore, that Substack data ironically isn’t counting traffic from Substack-specific social media channels. Substack increasingly considers itself a social media company (among other things), and we have quite a lot of readers who follow one or more Silver Bulletin-related accounts on Substack Notes but who opt out of email delivery of our newsletters. What’s more, the traffic you do get from Twitter and other social media channels tends to convert to subscriptions at relatively high rates. So we do very much appreciate it when you share Silver Bulletin articles. I should probably get more in the habit of including share buttons in articles like this one:

Share

Nonetheless, I feel confident in asserting that social media is a secondary source of business for us and is trending toward being a tertiary one — and that this is probably also true for most other publishers. That’s very different from a decade ago, when Facebook was considered the Golden Goose.

The content that gets “engagement” on Twitter is mostly complete crap

And yet, while Facebook is now almost completely irrelevant to the political discourse, that isn’t quite true for Twitter. Google search traffic in the U.S. for the precise term “twitter” is down quite a lot, but that’s not fair to X because the platform now has a new name. Broader traffic for search topics related to Twitter/X is also down, by more than half relative to the peak in late 2012. But the recent decline has been more gradual: about 20 percent as compared to two years ago. That seems to track with other third-party data showing a slow-but-steady decline in Twitter engagement, though nobody can be quite sure since X is no longer a public company.

But what does that remaining traffic consist of? I recently came across a bubble chart depicting the Twitter accounts that had received the most “engagement” in February 2026. It was depressing: most of the top accounts were extremely low-quality and highly partisan. I hadn’t even heard of many of them and only follow a handful of the top accounts. So I tracked down the original data myself and, with help from Claude, made my own improved version of the chart. Here, voilà, are the Twitter accounts with the most engagement so far in 2026:

It’s not hard to notice that Twitter has become extremely right-leaning. But I’d argue there’s an equally important trend: the top accounts are of incredibly low quality. Elon, with the algorithmic boost he built in for himself, is at the eye of the storm, of course. But “Catturd” literally gets far more engagement than the New York Times, for instance. There are a lot of tweets like this one, from former hunky actor turned minor conservative media star Kevin Sorbo, which is approaching a million views:

I’m admittedly a fierce defender of New York City, where I’ve lived since 2009. But anybody who knows the least thing about New York knows that 1975 was one of the lowest points in the city’s history, literally the year of “FORD TO CITY: DROP DEAD” because the city was on the brink of bankruptcy. This was the era of Taxi Driver, the Son of Sam and The Bronx is Burning, the beginning of a long period of elevated crime rates. Setting some B-roll to nostalgic music doesn’t erase that.

Without really wanting to comment on individual accounts — there are some exceptions — the liberal-leaning accounts that remain prominent on Twitter aren’t much better. They’re partisan and combative, sometimes peddling misinformation. They’re almost like a dark-mirror-world, Waluigi version of the conservative “influencers”, crafted in Elon’s jaded image of what liberals are like. It’s no coincidence that one of the most successful ones is the Gavin Newsom Press Office account, which literally mimics President Trump’s style in a sometimes funny, sometimes cringeworthy way.

The strange ecosystems of social media

Having lived through several eras of social media and web publishing business models, I tend to think of them as ecological systems. There are founder effects, predators and prey, and a lot of different survival strategies, often including mimicry. Most of all, there are selection effects. Some “species” are particularly fit for the peculiarities of the ecosystem and the economic incentives it produces, and within six months to a year, they tend to crowd out all others.

During the Peak Facebook Era, there was sometimes an air of faux scientific precision about the efforts to manipulate News Feed. I have no doubt that some of the people engaged in Moneyball-like algorithmic optimization tactics during this period were smart about it. But it’s important to emphasize that these ecosystems often reflect the “rules of the game” and the quirks of the algorithm rather than deeper truths about human nature or what people really want to read. The hack-ish strategies are often highly fragile and don’t survive changes in the environment. Few of the businesses that were considered hot shit during the mid-2010s are thriving today.

Certainly, I’m subject to my own biases: in thinking, for instance, that early/mid-2010s Twitter or today’s Substack are comparatively healthy ecosystems because they happen to work well for my business model. I certainly don’t think Substack is above critique, for instance. It also produces its share of mimicry. There were probably too many anti-woke newsletters during the early 2020s. But today, the Substack politics leaderboard is dominated by Resistance Lib publications of varying quality. To be fair, we still have roughly two years and nine months to go in Trump’s second term, which is an eternity in the media business. And what thrives on Substack often reflects a correction from oversights by executives in the mainstream media, who often chase away their best customers in an effort to fight the last war or to pursue their own political or ideological objectives.

Nonetheless, today’s social media ecosystem is particularly unhealthy, I’d argue. I’ve written about Bluesky previously, and I’m not sure there’s much more to say about it: it’s small, it’s continuing to get smaller, and last month its CEO announced she was stepping down. But at least when you look at the leaderboard of top Bluesky accounts, it’s fairly predictable: mostly reasonably prominent left-progressives, though almost nobody who is going to challenge Democratic Party orthodoxy from either the center or the left. The problems with Bluesky are less about the individuals who find success on the platform and more about the high school cafeteria behavior that follows when you put them all in a small, confined space together.

X, despite its much larger overall reach, also feels increasingly siloed. There are rarely consensus “main characters” anymore, and although I still do get dunked on more than your average bear, I usually discover this only when it’s force-fed to me by the “For You” tab; there’s often no sign of the pile-on from the 1,736 accounts that I follow. On the flip side, even when a tweet seems to generate a lot of favorable buzz on Twitter — increasingly rare — it’s at best a weak signal for predicting the metrics we really care about, namely engagement on Silver Bulletin itself and especially new subscriptions.4

And “siloed” is on a good day: at other times, Twitter feels like a ghost town. It’s still useful for some topics: the AI discourse on the platform is often relatively robust, for instance. But for something like the war in Iran, it’s next to useless. Links to external websites are substantially punished, and none of the workarounds are particularly helpful. So the tangible rewards from still having 3 million followers can be surprisingly marginal. However, my account is hardly alone in this regard. The New York Times has 53 million followers, and yet its tweets often produce only a few hundred likes, retweets, and replies even when they reveal urgent, breaking news.5

As with Facebook before it, the suppression of external links and the desire to maximize engaged minutes on X and turn it into an “everything app” is short-sighted. X is neither a particularly pleasant place to be, nor a prestigious one. To his credit, Elon now does offer some degree of monetization to verified accounts, although in my experience (my account was grandfathered into the program by virtue of having more than 1 million followers), it’s relatively modest, perhaps paying for a dinner out every other month.

If you’re capable of building up a truly valuable audience, however, there are much better ways to monetize it — including Substack, where you keep about 86 percent6 of all gross revenues and maintain control of your email list and your voice. Without the ability to consistently drive traffic outside the platform, the people who select into X as their main thing are inherently kind of bush league.

There’s a principle in ecology known as the island effect. I’m sure I’m not getting this precisely right, but that basic idea is that when there’s a lack of competition in an isolated environment, strange things tend to happen. There are all sorts of weird mutations that might not be survivable in a more competitive environment that can actually become fitness advantages on an island. Big animals tend to get smaller, and small animals tend to get bigger — like the Komodo dragon, whose range is limited to a few isolated islands in Indonesia.

So this is what Twitter looks like, basically, only with Catturd and the Gavin Newsom Press Office accounts locked in combat instead of a couple of (cute?) lizards:

Thank you for reading Silver Bulletin, which is also an island unto itself — but on its best days, we hope, an island of sanity. Back to Trump and Iran tomorrow.

Subscribe now

Sure, it might seem like the world’s gone a little crazy. But at least we’ll be getting a dose of March Madness. The NCAA tournament is perhaps our very favorite event here at Silver Bulletin.1 In fact, the tournament really needs no introduction. I’ve been running versions of these projections since 2011, almost ten years before I’d even heard of Substack.

But pretty much every part of the model has changed at some point. Instead of compiling other people’s projections, we now mostly rely on our own in the form of our new COOPER ratings — although we blend them with Ken Pomeroy’s ratings (because we think Pomeroy is pretty much the best, present company excluded.) We’ve also upgraded our technology stack. Believe it or not, until last year, part of the work for our NCAA model was still being done in Microsoft Excel.2 Now we’ve built the code to run proper simulations — 100,000 at a time — which enables more sophisticated handling of injuries (and there are a lot of them this year), among other things. This isn’t an “AI-driven” model quite yet, though Claude and other AI tools are increasingly helping us smooth out some of the code's rough patches and turn these numbers around more quickly.

Still, March Madness is eternal. If nothing else, there will be upsets: we’ve calculated that the odds of having a perfect bracket are on the order of 1 in 10 quintillion. I guess I’ll just have to concede that we’re living in a simulation if it happens.

How our NCAA forecasts work

• This is what we call a landing page, meaning that both the numbers and the text will change. We strongly recommend checking out the web version with this link, which provides for better rendering of charts than the Substack app or email.

• We’ll update these numbers once per day after tournament games are played. In the interest of underpromising and hopefully overdelivering, sometimes the update will need to wait until the morning, but we’re hoping to get it out in the evening most of the time.

• The ratings give 5/8ths of the weight to COOPER and 3/8ths to Pomeroy. There are lots of other systems out there, but basically, that’s the mix I’d use if I were betting on the games myself.

• They also account for injuries, travel distance, and — once the tournament is underway — how teams have performed so far in the tourney relative to the model’s expectations. (Early-round over- or underperformance is often fairly predictive of what happens later on.) For more details about how all of this works behind the scenes, see our methodology page.

• Women’s projections are now ready also and can be found here.

Let’s start with the overall leaders:

What you’ll find on the rest of this page — and throughout the tourney

The rest of this page contains:

• Region-by-region projections of each team’s chances of advancing to each round, plus my first thoughts on each region. Analysis of the East region is free for everyone, while the rest is a bonus for paid subscribers.

• Odds for forthcoming games (win probabilities, point spreads, totals, etc.)

• A comparison of the ratings systems used in the projections and the impact of injuries.

• And a spreadsheet version of these projections. We’re hoping to archive these so you’ll be able to see how our odds have evolved throughout the tournament.

Our tour begins in the East:

Next up, the South, who will face the East winner in the Final Four.

Read more

Subscribe now

Personally, I think it’s pretty great that the women’s NCAA tournament has become a huge deal. Who’s going to object to double the March Madness? Or another opportunity to build a cool model for Silver Bulletin subscribers? But I need to remind you that there are significant differences between the men’s and women’s games, and those are especially pertinent this year.

Every year, when we publish the men’s forecast, we urge you to remember that there will be a lot of upsets. The #1 seeds in our men’s bracket this year — Arizona, Duke, Michigan and Florida — have a combined 57 percent chance to win the national title. We put a lot of work into calibrating these numbers properly and that’s in the same ballpark as the historical rate, though #1 seeds have been on a hot streak recently.

On the women’s side, conversely, we have the #1 seeds — UConn, South Carolina, Texas and UCLA — with a combined 93 percent chance of winning the hardware this year:

That directionally matches the historical data — since the tournament began in 1982, only three teams below a #2 seed have won the women’s tourney (all of them #3s). But the pattern wasn’t as extreme in our forecasts last year, largely because UConn (whom we had as the 2nd-best overall team) was inexplicably seeded as a #2. This year, there’s a much bigger cliff between exactly the first four seeds and everyone else, with the partial exception of #2 seed LSU, which would be plenty good enough to earn a 1-seed in most years.

Everything runs through the Huskies this year, though, because UConn is undefeated again. Take a look at their composite rating of 2684. These ratings are literally designed to be on the same scale as Elo ratings that were originally derived for chess. So that UConn rating is basically equivalent to Magnus Carlsen’s ~2800 Elo. And the early rounds of the tournament for teams like UConn usually resemble what you’d get if Carlsen were to face a random club player: something has to go very wrong for them to lose. While upsets can occur in the opening rounds — let’s not forget Harvard-Stanford in 1998 — they’re considerably less likely than for the men, especially considering that the top seeds host home games in the first two rounds.

While there are some differences between our men’s and women’s COOPER ratings — notably, that women’s programs have more carryover from year to year, in part because women can’t join the WNBA until age 221 — this dominance for the top teams mostly emerges organically from the data. The top women’s teams win a lot, often almost doubling their opponents’ scores in the opening rounds of the tournament.

You are likely to see some great games in this year’s tourney, however. Based on our historical COOPER ratings, the other #1 seeds this year — UCLA, Texas and South Carolina — are roughly as good as typical overall #1s from past tournaments. So the Huskies will be challenged, even if it’s not until the Final Four. There’s also a great potential Elite Matchup looming between UCLA and LSU.

Let’s run through the numbers. The rest of this page contains:

• Region-by-region projections of each team’s chances of advancing to each round, plus my quick commentary.

• Odds for forthcoming games (win probabilities, point spreads, totals, etc.) customized to the women’s game.

• A comparison of the ratings systems used in the projections — we blend COOPER with ratings from HerHoopStats — and how we adjust them for injuries.

• And a spreadsheet version of these projections.

Let’s start in UConn’s region, which is clumsily named2 “Fort Worth 1” after the site of the regional final and the Huskies’ #1 overall seed.

Next, let’s go to the Sacramento 4 region, which the winner of Fort Worth 1 will play in the national semifinal in Phoenix:

Read more

Don’t get me wrong: I’m excited about the NBA playoffs. But my interest in the NBA regular season has been flagging in a way that it hasn’t in a long time. There’s a simple reason: it’s the tanking. A full third of the league — five teams in each conference — basically gave up on the season at some point between October and February. The identities of the 10 playoff and play-in teams in each conference were practically locked in a few weeks ago — and they’re literally locked in now.

The NBA is acutely aware of the issue, though it’s unclear whether the league considers it a real problem or just a PR issue. (Here’s why it’s an actual problem: about half the games on any given night “feature” a tanking team. I’m not about to pay $200 a ticket to see a team that isn’t even trying to win play the Knicks at MSG.) And the solutions it has proposed are mostly tinkering around the edges with the current rules, full of the same kinks and quirks that will be exploited by future Sam Prestis and Daryl Moreys.

Maybe that’s because it’s not quite clear what the league wants. Losing isn’t the same thing as tanking, and perhaps it’s fine to reward the former but not the latter. Up to a point, you might want to help teams for sustained losing as opposed to a one-year fluke like this year’s Indiana Pacers, but maybe not if they’re perpetual basement-dwellers. Other things equal, you probably don’t want the same team to get a #1 or #2 pick several times in a short period or for a potential contender to luck into a top four pick. Meanwhile, the myriad of pick swaps and trade conditions creates weird cliffs in the system and perverse incentives. And while there’s a lot of focus on losing teams, it’s also not clear you want winning teams to be able to trade draft capital seven years out when ownership and management may turn over anyway once it’s time to pay the piper.

So what if I told you there’s an alternative that could accomplish all of the following?

• Introduce explicit penalties for tanking or for repeatedly failing to reach even the play-in game.

• Substantially simplify trade rules.

• Flatten the lottery odds to the extent desired by the league, especially toward the top.

• Protect teams against themselves by limiting the amount of future capital they can trade.

• Eliminate cliffs in the system, such as when a team would have to give away its pick if it climbs above a given position in the standings.

• Reduce the element of luck: bad teams would still be rewarded with better capital, but they wouldn’t be quite as subject to the literal bounces of the lottery balls.

• Reduce the chance of already good teams landing top picks.

• Put some curbs on the same team repeatedly winning top picks.

• Provide teams with more flexibility to suit team needs, such as carrying over capital from season to season or acquiring depth rather than a single highly touted prospect who isn’t a good fit.

• Usually guarantee a team the opportunity to call its shots if there’s a player it really likes in the rough vicinity of its board.

• And make “draft night” — now auction night — even more strategic and fun.

Sounds pretty good, right? But it does require one leap of faith. As I’ve been teasing at, we’ll be getting rid of the “draft” per se. No, I’m not advocating for all players to simply become free agents, although there are worse ideas. I don’t think it’s inherently a bad idea to “redistribute the wealth”. But every fantasy football nerd’s favorite solution, an auction rather than a draft, provides for a lot more flexibility to tweak the knobs to the league’s desired levels.1

Subscribe now

True, you could accomplish some of these things within a draft format. But once I started down this path, the auction route proved far more flexible. For instance, you can penalize teams in a proportionate way for tanking by deducting a little bit of capital instead of either giving them a slap on the wrist or employing a “nuclear option” like taking away draft picks entirely.

The basics of ARC

The gist of the system is this: the draft would be replaced with an auction-type system, where teams bid on eligible players each June with Acquisition Rights Capital or ARC. I thought about calling this something cheeky like “DraftBucks”, but a) there’s not really a draft any longer, and b) I wanted to go with the sort of technical term that could plausibly appear in the CBA. ARC, of course, can be traded, but the mechanisms for doing this are simpler than under the current draft format. A unit of ARC can alternatively be designated with an arc symbol ⏜.

Needless to say, the devil is in the details for any proposal like this. There are 21 rules governing the use of ARC. You could tinker with most of them without really messing with the overall spirit of the system, however.

Let’s start with the two most important rules:

1. At the end of each season, teams receive between 25 and 100 ARC based on their order of finish. As under the current system, playoff teams are slotted 1-16, (though the two NBA finalists would be ranked first and second2) followed by non-playoff “lottery” teams.

2. There are no more swaps or conditional picks. Instead, all trades involving future asset rights are made with ARC. All trades or player acquisitions involving ARC must be made in whole units; ARC cannot be fractionalized.

Rule #2 is, I hope, relatively self-explanatory. ARC is fungible, so there’s no benefit in trading, say, Miami Heat ARC for Minnesota Timberwolves ARC. So, instead of “the Rockets acquired Giannis Antetokounmpo for Reed Sheppard, Jabari Smith Jr. and [insert long and convoluted description of picks and swaps]” the transaction wire would read as “the Rockets acquired Giannis Antetokounmpo for Reed Sheppard, Jabari Smith Jr. and 50 ARC”. There’s actually a lot less to keep track of.

As for the allocation, here’s what I had in mind:

There are some subtleties here. Under the current system, the steepest part of the curve, in terms of the expected value gain from dropping an additional position in the standings, is roughly between the 20th and 23rd positions. For instance, the 20th-ranked team has just a 9.4 percent chance of lucking into a top-four pick; that nearly triples to 26.3 percent for the 23rd-ranked team. Teams in this region can usually play a little bit and have a plausible chance of at least making the play-in round. But the NBA’s current setup strongly deters them from competing; our ARC system would reduce the marginal gain from losing in this region of the curve by roughly half. Instead, under ARC, the steepest part of the curve is for teams that have made playoffs anyway; it’s flatter toward the top and the bottom. And the ARC allocation for the three worst teams would be completely flat, similar to how the league awards the same 14 percent chance of receiving the #1 overall pick to the bottom three teams now.3

Here’s how I estimate what each draft slot is worth. The metric we use to evaluate player outcomes in our new NBA draft model, PRISM, is based on Estimated Plus-Minus (EPM). We tallied the number of wins a player produced above replacement level in his first seven NBA seasons (when he’s most likely to be on some sort of cost-controlled contract with his original team4), rounding below-replacement-level seasons up to zero. I then drew a smooth logarithmic curve over this data based on a mix of the average WAR and the median WAR historically associated with each draft slot. Overall, we calculate that ~55 players in each draft class5 produce enough expected wins to be worth a guaranteed contract, which is a close empirical match for how NBA teams behave in reality.6

Under our system, there are 1860 ARC distributed to teams in any given season. (As we’ll get to later, teams can have their ARC allocation penalized under some circumstances, but any “taxes” the league collects are redistributed into the system.) If we divide ⏜1860 by the expected value above replacement level associated with each draft slot, we come up with the following:

As you can see, draft value becomes nonlinear toward the top: the #1 overall pick is worth about twice as much as the #7 pick. Having looked at a lot of other “draft charts”, I’m comfortable that’s somewhere in the right ballpark.7 But value tends to flatten out once you get out of the lottery. The first half of the second round can produce tangible value — sometimes even a lot of value (Nikola Jokic was drafted at #41). But the back half of the second round literally zeroes out; teams are usually indifferent toward keeping the players they select at all.

The chart does hint at one important feature of the system: there’s a ⏜100 cap on how much a team can bid on any given player. If several teams make the same bid, then the league goes back to the ping-pong balls to determine who gets him. So there is still some element of luck: typically, in any given draft, the top three picks will be worth more than the max 100 ARC allocation. You can’t just stockpile ARC and guarantee yourself a Wemby or a Cooper Flagg, in other words.

We’ll return to some of these details later, but first let me run one more comparison between ARC and the existing rules. This chart shows the expected value of each draft slot under the current system based on the lottery odds for first round picks and compares it to each slot’s ARC allocation under our new model. Importantly, the expected value estimates also include second-round picks, which provide a sneaky benefit to the worst teams since the first half of the second round is worth a lot more than the back end. (Under ARC, there’s no more second round because there’s no more draft; teams just keep bidding on players until certain criteria are exhausted.)

So yes, we’re taking a little bit from the poor and giving it to the rest of the league. Finishing in one of the bottom eight positions in the standings is now worth slightly less, especially in the bottom three. But it isn’t that big a shift, and the largest benefits convey to the league’s middle class.

Plus, even if we’re giving them less capital, we’re probably reducing variance for the worst teams. A typical draft features three players who are worth a ⏜100 max bid, and three teams each season who are guaranteed to receive 100 ARC. Surely, some other teams will acquire the capital for a ⏜100 bid via trade (or because they have some ARC left over from previous seasons; there’s a mechanism for that). But generally, teams will have more control over their own destiny. If there’s a player they really like — say, Steph Curry in 2009 — they can grab him, provided he’s not a max player. There’s also a rule (#10) that prevents the same team from repeatedly winning “mix bid” auctions over a multi-year period.

Share

The auction process

3. During the auction, teams take turns nominating players in inverse order of finish. They must bid at least ⏜1 on any player they nominate. A team may pass on its nomination and still bid on players nominated by other teams, but they can’t nominate players once their turn comes up again in the rotation once they pass.

4. When a player is nominated, all teams have 7 minutes to submit simultaneous bids (reduced to 5 minutes on the second day of the auction8). All bids are then revealed, and the player is awarded to the highest bidder.

5. The maximum bid is ⏜100.

6. If 2 or more teams tie for the highest bid, the winning team is chosen by lottery among the tied bidders.

7. Teams may acquire a maximum of 4 players through ARC in any given season.

8. ARC budgets and player trades are frozen 72 hours before the auction. A team cannot select a player and immediately trade him, though in line with the NBA’s current rules, he can be traded 30 days after he signs a contract.

9. The draft ends when all remaining teams have passed on their nominations, have acquired 4 players, or have exhausted their ARC budget. There is no more “second round” under ARC, although teams may sign rookies who aren’t selected through the system as free agents.

The most important rules are #5 and #6, but they’re the ones we’ve already covered. By design, the top three picks are usually worth more than the maximum bid of ⏜100; the average #1 overall pick has a value of about ⏜140, for instance. So in most halfway decent draft classes — maybe not the Zaccharie Risacher draft — there will be multiple max bids for the first few players. Teams will face challenging decisions as to whether to acquire enough ARC to make a max bid; what’s interesting is that for every additional team that plans to bid the max, the odds of winding up with the player decline.9

I’m less hung up on some of these other rules. There is probably some slight tactical advantage to having the right to nominate a player, but it likely isn’t much.10 We might see some unconventional strategies, too, like a team nominating the equivalent of a second-round pick and hoping to get him for a few ARC while holding their fire for a bigger bid down the line; there’s nothing that says the best players have to be nominated first.11 The action would be frenetic on auction night, with new information revealed at every turn. So frenetic, in fact, that as per Rule #8, I think it’s just too chaotic to allow teams to acquire ARC on draft night. Instead, front offices would lock in their budgets and have three days to make their best-laid plans.

Multi-year planning

10. A team may win a player with a maximum ⏜100 bid at most 2 times in any 3-year window and 3 times in any 5-year window.12

11. Teams must spend at least ⏜25 in any 3-year window. It’s use it, or lose it. If a team doesn’t meet this requirement, any outstanding ARC is deducted from its account.13

12. Teams can go into ARC debt, provided they’re guaranteed to meet the ⏜25 spend per three years requirement under Rule #11.

13. Unused ARC rolls over to the next season, but with an escalating tax schedule. Teams pay a 10 percent tax rate on the first ⏜10 of ARC savings, a 20 percent tax rate on the next ⏜10 and so forth.14

14. All ARC subtracted via fines or taxes will be reallocated evenly to other teams.15 Thus, the total amount of ARC in the system is preserved.

The rollover provision under Rule #13 is intended more for “spare change” than for teams to stash large amounts of ARC. Joseph’s work for PRISM found that there’s meaningful and somewhat predictable variation in the strength of different draft classes. Still, the taxes get punitive if you try to retain more than about ⏜20 or ⏜30. And even if you can’t find a good fit for your ARC in the auction, you can always trade it.

Rules #11 and #12 are probably the more important constraints. Under Rule #1, each team is guaranteed to pick up at least ⏜25 every season. Under Rule #11, they can borrow against this and go into ARC debt, but they’re still required under Rule #11 to have enough left over to spend at least ⏜25 in any given three-year span.

On the surface, this is more flexible than the current Stepien Rule, which requires teams to make a first-round pick every other season. Under Rules #11 and #12, a team has to make the equivalent of a late-first-round pick only once every three seasons instead.16 However, under ARC, a team won’t really be able to trade draft capital more than two years in advance17, whereas draft picks can be traded up to seven years in advance under current rules. And it’s the picks in the out years that tend to be most valuable, since acquirers hope that the contending club will play through its competitive cycle and crash out to the lottery by the end of the window.

This is a big change. A lot of NBA tanking discourse takes what you might call a “demand-side approach”, i.e. by focusing on the fact that teams have a lot of desire to acquire high future picks to the point where they frequently tank. What can be neglected is the supply side: there’s a lot of opportunity to acquire such picks because teams can trade draft capital seven years out and they often take an incredibly short-term focus. Reducing the potential for long-term, franchise-compromising moves would protect contending teams (like the Phoenix Suns when they acquired Kevin Durant) from themselves. But it would also decrease the amount of draft capital available to rebuilding teams from trading their own star players. With less draft capital available, they might have to “settle” for acquiring player talent in return, which would keep them more competitive. Or they might be more inclined to keep their own players since there won’t be as many too-good-to-refuse offers.18

Rule #10, meanwhile, is intended to slightly reduce the amount of luck under our system. A team can acquire a ⏜100 player — basically, a guy who is a legitimate top 3 pick — at most two times in any three seasons or three times in any five seasons. If you’re going to spread the wealth around to losing teams, you might as well spread it a bit more evenly.19

Anti-tanking measures

15. If a team has missed the playoffs and play-in game for more than 3 consecutive years, it’s subject to a 20 percent tax on its annual ARC allocation.

16. The league will deduct ARC from teams for tanking or failing to uphold the competitive integrity of the league. If the league suspects a high likelihood of tanking, it will publicly issue a team a warning that it’s been placed under heightened scrutiny. The maximum allowable fine without a warning is ⏜5. Once a team is under heightened scrutiny, the maximum allowable fine is ⏜20 for first-time offenders and unlimited for repeat offenses, with the clock resetting after seven seasons.

Two-thirds of the 30 NBA teams make either the playoffs or the play-in tournament every season. It just isn’t that hard to do. At the end of the 2025-26 season, the only ongoing three year play-in/playoff droughts will belong to the Brooklyn Nets, Utah Jazz and Washington Wizards.

So I have no trouble punishing these teams under Rule #15. More than three years in the wilderness is when season-ticket holders start to give up hope. Would Jaren Jackson Jr., Trae Young or Anthony Davis have been shut down if the Jazz or Wizards could reset their clocks by sneaking into the play-in game? Maybe not. Under this rule, a team with a 3-year drought would actually acquire more ARC if it made the play-in than if it finished with the 5th-worst record.

Rule #16 is likely to be more controversial because it’s more subjective. But the league makes subjective judgments all the time, such as when disciplining players. It’s also fined teams like the Jazz and the Mavericks before for tanking, although the amount of the fines is trivial for billionaire owners. Rule #16 would set up a two-tiered system so teams weren’t totally blindsided; taking away a few ARC and putting them under heightened scrutiny (so for instance, they’d need to see the receipts for purported injuries) would be the warning shot. Repeated, blatant tanking once teams were on probation would be punished more harshly.

Every half-sentient NBA fan (and absolutely every person betting on the games) knows exactly which teams are tanking; it can’t be that hard for the league to police this explicitly, whether the determinations are made through a Potter Stewart standard (“I know it when I see it”) or some sort of algorithm. All you really want is a deterrent: the expected value of making dubious, tank-tastic moves will be lessened if you might face real consequences for doing so.20

Player contracts

17. All players acquired through the ARC system must be offered guaranteed contracts, with salaries tied formulaically to the winning bid and the number of maximum bids.21

18. Teams retain NBA rights to players for 4 seasons once they’re chosen through the ARC process. The ARC spent on these players is deducted from a team’s account when the player is selected, whether or not the player ultimately signs a contract. Players may decline the initial contract and negotiate a higher salary after one year, subject to certain constraints.22 But they are not eligible to sign with teams who don’t have their ARC rights until the 4-year window expires. The rights to such players may be traded, however.

We’re getting into some of the less important mechanics, so I’ll be quicker here. A player selected during the auction must be offered a guaranteed contract. Currently, first-round picks get guaranteed contracts and second-round picks don’t, but about 20 of the 30 second-round picks wind up getting full-fledged NBA deals anyway. That implies that about 50 players will typically be selected in any given ARC auction. Teams can sign unselected players later, just like they can offer contracts to undrafted free agents now.

However, under current rules, teams sometimes do something else with their late picks: choose a “draft-and-stash” international player in the hopes of bringing him into the NBA after he’s picked up a couple more years of seasoning abroad. Rule #18 is intended as a replacement for that process, while also giving international players some leverage to negotiate for a higher rookie salary if their development goes well. It also plays a more subtle role in creating another sort of “asset class” that teams can move. Since there are no more second-round picks under ARC — essentially, the spare pennies under the current system — teams could trade the rights to draft-and-stash players if they want to “round up” their capital to have the right goods to make a trade.

Transition to the new system

19. Before the start of the first season under the ARC process, each team gets an additional ⏜10 above and beyond their standings-based allocation under Rule #1.

20. Expansion teams start with 100 ARC.

21. Teams may trade ARC in advance of the full transition to ARC provided they meet Rules #10 and #11. Alternatively, by mutual agreement with the Players’ Association, the league may seek to accelerate the transition to ARC by “cashing out” existing draft-pick trade obligations to their equivalent ARC values.

Rule #19 creates a bit more liquidity in the system, but I don’t think you need or want to have too much liquidity.23 Teams can already go into ARC debt, and there’s already going to be a lot of demand for winning teams to move their ⏜25 or ⏜30 allocation to losing teams so they’ll have the capital to make a max bid.

Expansion to Las Vegas and Seattle is probable but not quite inevitable. If you’re going to ask for $8 billion (!) for a franchise in mid-tier markets, I say go ahead and give them a full 100 ARC for their first auction.

And yes, one big critique you might make of ARC is that it’s hard to have a seamless transition when teams have already traded draft picks as far out as 2032 (!). But the current system isn’t working. The alternatives proposed by the league are uninspired and might create as many problems as they solve. It’s time to rethink things from the ground up.

The last time I did a check-in on President Trump’s popularity in November, I noted that the decline in his approval ratings had been remarkably linear. There have been peaks and valleys, sure, which you can correlate with particular news events. But the recovery after every bounce-back has been incomplete, leaving some permanent damage behind.

So what’s happened since then? Well, let’s just go ahead and update that chart.

Obviously, the trend is still down. To be fair, his numbers were relatively flat between November and March.1 But Trump just hit a new low in our tracking. For the first time in his second term, Trump’s approval rating in our average is (just barely) below 40 percent at 39.7. And his net approval rating is -17.4, also a new low. The recent decline has been pretty steep: about 5 points of net approval over the past several weeks.

From a political standpoint, the most important events since November are probably the killing of two American citizens by border agents in Minneapolis in January, and then, of course, the Iran War, which began on Feb. 28.2 The effect of Minneapolis on Trump’s topline numbers is more debatable. Trump’s ratings have declined on immigration, and the government’s conduct in Minneapolis was deeply unpopular with swing voters. Immigration and “border security” remain just about the only issues where Trump’s ratings remain close to breakeven, however.

Iran has had a much clearer impact. So far, though, it’s probably less about the war itself than about the impact on gas prices, which are now at about $4 nationally, having risen by more than a dollar over the past month.

It’s still the economy, stupid

The start of the war did not produce a traditional rally-around-the-flag effect, but it also didn’t seem to immediately hurt Trump’s topline numbers. We’ve started publishing an Iran poll tracker, and it finds that net approval for the war is -15.3. So it’s unpopular, notably unlike other recent conflicts that began with more public support, but no more unpopular than Trump himself.

It was, however, about a week into the war when markets started to freak out over the likelihood of a prolonged disruption to oil shipments in the Persian Gulf. Here is the effect on prices at the pump:

Gas prices are likely to a sore point for a president who won an election largely on inflation. In 2024, 40 percent of voters in the exit poll said that “high prices for gas, groceries and other goods” was the single most important factor deciding their vote, and they broke 2:1 for Trump. It would be hard to pick a more visible indicator of affordability.3

Although there’s also the stock market. With its ubiquity on TV tickers, it’s closer to underrated than overrated as a measure of overall “vibes”. If you’re checking your 401K (and more Americans are than usual this time of year because it’s tax season), you’ll find that your wealth has declined by a chunk.

We saw some similar effects during the government shutdown. As Democrats struggled to craft the right message, Trump’s numbers actually improved during the early days of the shutdown. But they began to decline right when the White House threatened people’s SNAP benefits. As much as political discourse on social media can be centered on abstract, “post-material”, “culture war” issues, these debates occur mostly among people who are already strong partisans and who aren’t likely to have their overall opinion about Trump changed much in either direction. It’s when people’s pocketbooks are affected or their daily lives are disrupted that broader public opinion tends to change.

Trump will bounce back, right? Well, that’s not so clear.

So far, however poor, Trump’s approval ratings remain within the range of other recent presidencies. Biden bottomed out at a -20.5 on July 4, 2024 (after his debate with Trump but before he dropped out). Trump’s first-term low was -20.8 on Aug. 6, 2017; he also hit a -19.1 a week or two after the events of January 6, 2021. Even Obama had some rough stretches, hitting his all-time low of -13.2 on Nov. 18, 2013.

There’s no reason to doubt that the extreme political polarization puts a cap on a modern president’s approval ceiling — and probably also raises his floor. And yet, when I look back on that chart of Trump’s numbers, here’s what I’m struck by: so many of these political wounds have been self-inflicted.

In fact, you could argue that he’s actually been lucky not to have more problems. There haven’t been a lot of natural disasters during Trump 2.0, or major wars (like Ukraine or Gaza) breaking out that Trump didn’t start himself. The biggest economic shocks have also been Trump-caused: the tariffs last year, and now the oil shock. Meanwhile, he’s benefited from the boom in AI investment that has helped to keep tech stocks afloat, without which we might be in full-on bear market.

Read more