The 2026 World Cup will begin on June 11, creating the biggest spectacle in the history of the sport: 48 teams, 104 matches, $11,000 tickets, and three hosts: Mexico, Canada and the United States. It will also be a referendum on the state of the U.S. soccer program: one that, based on our recent results, we might not be prepared to pass in a blaze of red, white and blue glory.

But World Cup time also means “model time” to me. I first designed a soccer rating system, Soccer Power Index, for ESPN in advance of the 2010 World Cup; we later extended it into SPI 2.0 at FiveThirtyEight. While SPI was a good system, I’ve spent a lot more time developing sports models since then.

PELE, our new international soccer model, combines the best features of SPI with some tricks we’ve learned from our newer models. It uses a mix of match results, player market values, and historical economic and geographic data, and is obsessively customized to the peculiar contours of the global sport.

We’ll use PELE, Predictive Elo with Lineup Equilibria, to power our official Silver Bulletin World Cup projections once rosters are finalized in a few weeks. But rather than being a placeholder for SPI, we think PELE has a lot of value on its own — in fact, it might be my favorite model we’ve worked on. Starting today, we’ll publish our PELE ratings for all 211 FIFA teams, covering all years since 1872, and update them whenever new matches are played.

You can get a more formal introduction to PELE on the methodology page. But the best way to demonstrate the breadth of the system is by example. So let’s start with our overall PELE ratings. One tip: we’d encourage you to view PELE on the web rather than in email or the app.

Share

PELE ratings

Each team gets a continuously updated pair of ratings: PELE and Tilt. PELE is the headline number that predicts overall team quality. It’s basically a fancy variation of an Elo rating system, though it’s affected by changes in player market values and player ages as well as match results. Argentina and Spain are essentially tied for the top slot as we head into the 2026 World Cup. But you’ll have to scroll down a few pages to find the United States.

The rest of this page contains:

• A comparison of PELE and FIFA rankings;

• Tilt ratings;

• Offense, defense and round-robin results;

• Future match projections, including for the 2026 World Cup group stage;

• The Atlas of PELE, our 12 footballing regions and home-field advantage;

• Market values and projected team trajectories;

• And, detailed historical PELE data since 1872, including the best teams of all-time.

The last few “advanced” sections are a bonus for paying subscribers. We greatly appreciate your support and we’ll consider extending PELE to club soccer depending on the amount of interest in the model.

Subscribe now

Ready for some ranking controversies? You can already get a few hints for how PELE differs from other systems. Because PELE is designed to be predictive, it can be more sympathetic to recent underperformers like Germany (#7 in PELE, #10 in FIFA) and more skeptical of teams on a possibly unsustainable hot streak (FIFA has Morocco #8; we have them #22). It also tends to like South American teams (CONMEBOL1) better than FIFA. And market value data can help teams with star talent like Norway in PELE.

Tilt ratings

Offense and defense are more fluid in soccer than in sports like American football, but there are some predictable patterns. Tilt ratings indicate a team’s tendency toward attacking or defensive play. They are based on both long-term tactical preferences and the personnel on the field — the presence of Erling Haaland makes Norway more attack-minded than they otherwise would be. Tilt measures the impact on the total number of goals scored combined between both teams.2 For instance, if Germany’s Tilt rating is +.6 (“very attacking”) that means games involving Germany will tend to feature 3+ goals rather than the ~2.5 that is more common in modern international matches.

Offense, defense and round-robin results

Now that we have both PELE and Tilt ratings, we can do some neat things. Most importantly, we can forecast matches! Our round-robin results project every pairwise combination of match results between the 211 current FIFA members (weighted slightly toward more challenging matchups3). Also, this process allows us to derive offensive and defensive ratings for each team, as measured by the projected number of goals scored and allowed against the same competition. Teams with negative Tilt ratings like Colombia and Senegal will produce lower-scoring matches with more draws.

Future match projections

But better yet, let’s forecast some actual games. We’ve uploaded the schedule for all future international matches that we’re aware of. This includes group-stage matches for the 2026 World Cup, though our ratings for World Cup games will be slightly tweaked later once rosters are finalized.4

Read more

🕒 The latest on Trump’s approval rating

Updated May 15, 2026

Donald Trump’s net approval rating hit a new second term low of -18.9 in the Silver Bulletin average today. That’s not much different from his previous low of -18.8, but still, it’s bad news for an already unpopular president.

Things look even worse under the hood. Among just US adults — as opposed to our standard average that also incorporates polls of registered and likely voters — Trump’s net approval is -20.6. And about 48 percent of Americans strongly disapprove of Trump’s job performance. These numbers are partially why Democrats are on track for a strong performance in the midterms, even given their recent setbacks on the redistricting front. -EMD, 5/8/26

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

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.

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

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

Subscribe now

The NBA draft is the league’s most consequential event. We touched on this in our Future of the Franchise rankings, but most teams are in their respective positions because of their decisions in June. This has held true since the NBA’s inception — contrary to popular belief, only a handful of champions were built primarily through free agency or trades. Most title teams, like last year’s Oklahoma City Thunder, are born of smart drafting, supplemented by shrewd moves on the trade market rather than the other way around.

Teams are aware of this, and, of course, are willing to throw away entire seasons because of it. (With the league’s proposed anti-tanking solutions leaving something to be desired, we’re still working on our own NBA draft reform plan. Look for that soon.) In the years since Sam Hinkie’s “process”, tanking and its more polite cousins like “strategic rebuilding” have become a fixture of the league’s competitive landscape. Front offices will openly gut rosters, bench healthy veterans, and punt on entire seasons, all for the chance to move up a few slots.

You would think, then, that NBA teams would be remarkably good at making draft picks. If the draft is important enough to sacrifice entire seasons for, surely the evaluation process behind it must be at least very good? But every year, we see the same pattern of misses. There are a few reasons for this.

Yes, scouting is inherently noisy. Even with a sound approach, you’re trying to project outcomes years into the future from a small sample of games among a set of 18-year-old outliers. This isn’t to excuse bad process — there’s definitely plenty of that across the NBA. But even analytically sound prospects fail to translate for reasons that are hard to foresee.

NBA teams undervalue production and overvalue “potential”

Cognitive biases also distort where players end up getting drafted. Consensus evaluation tends to prize linear ordering — ranking players from “#1 option” down to “role player” — even when that hierarchy doesn’t map onto how basketball value actually works for most teams. The result is that, relative to more “reliable” players, teams systematically overweight creator archetypes and the sorts of players who can turn into what our friend Jeremias Engelmann calls “quagmires”. At each slot, teams should be selecting for the highest expected value, not just the highest ceiling. Team-specific needs complicate things further, but figuring out how to define expected value is most of the problem.

Current production is also chronically underrated. Simply put, the strongest predictor of future NBA success is current success. This shouldn’t surprise anyone, and yet every draft cycle, evaluators find new ways to talk themselves out of good players. Paolo Banchero was drafted ahead of Chet Holmgren in 2022 largely on this basis despite far less efficient college stats, but Holmgren has almost certainly been the better pro from an impact standpoint. So our new model, PRISM1, focuses on identifying good, impactful players for their age without glaring holes in their statistical profiles.

Subscribe now

Admittedly, athletic characteristics like vertical leap, agility, speed, and strength are hard to measure outside of combine testing, which only becomes available in late May. And to be fair, our PRISM projections still incorporate consensus scouting rankings — though we’ll show you our “pure stats/computer” version just for fun later on.

But do combine measurements tell us which of those traits are functional — that is, do they actually correlate with positive NBA outcomes? The best studies we’ve seen say not really. If you want an athletic guard because athletic guards get to the rim, a guard who already gets to the rim is a good bet, regardless of what he tests at the combine or what the aesthetics look like.

When we remember the great teams in NBA history, it’d be hard to not rank the Golden State Warriors dynasty from 2015 through 2022 near the top. Most narratives assume they caught lightning in a bottle, drafting Stephen Curry in 2009 and Draymond Green in 2012. But what if I told you those picks were layups? That the Warriors did something as simple as ignoring aesthetic biases — Steph was considered too frail to translate and Draymond didn’t fit a specific archetype — and simply selected the most productive players at their positions. Curry and Dray were two of the best college players two years before they were drafted. There are actually plausible arguments that both should have been drafted top five in 2007 and 2010, respectively. The cornerstones of a dynasty were not that difficult to identify years before anyone considered them future Hall of Famers.

“Analytical” scouting, which isn’t restricted to models, certainly has its place in NBA front offices today, but it’s not nearly as widespread as most fans would think. There seems to be this assumption that front offices are making decisions with extreme efficiency and little bias, but many teams still default to silhouette scouting — evaluating players based on physical traits and aesthetics rather than production, and projecting outcomes based on who a player looks like rather than what they’ve done. This might explain why Kon Knueppel got Joe Harris comparisons last season.

When a player consistently drives winning outcomes against high-level competition, the burden of proof reverses. The question shouldn’t be whether their game “looks” translatable, but whether there’s any structural reason it wouldn’t be. “Structural” might be doing a lot of work in that sentence, but we mean this should be grounded in real, current gaps in production — not just perceived translatability concerns. Luka Doncic shredded the EuroLeague before entering the NBA, winning its MVP in 2017–18, but fell to third behind Deandre Ayton (!!!) and Marvin Bagley III (?!?!?) because evaluators worried his lack of athleticism would neutralize his otherwise extraordinary skill set. It didn’t. Luka never became a great athlete; the traits that made him dominant in Europe simply remained dominant in the NBA. On the other hand, Jahlil Okafor was genuinely productive in 2014–15, but fell out of the league because of his low passing and defensive indicators — real holes in his profile, not just aesthetic concerns.

PRISM is built to operationalize these principles. It anchors on production — what players have actually done against real competition — and layers in the contextual factors that shape projection: shot creation volume and efficiency, defensive and offensive roles, positional size, age-adjusted performance curves, and more. So let’s take a look at our top 5 for this year’s class. Note that PRISM currently only rates players who have spent at least some time with a Division I NCAA program2, but virtually all of the top prospects fall into that category this year.

Intrigued? You can find a more complete methodological description of PRISM here. Spoiler alert: Kansas’s Darryn Peterson, still considered the potential #1 overall pick, is just a fraction of a point outside of our top five. But his ranking has consistently fallen over the course of the college season as we’ve been refining PRISM, and the Jayhawks’ washout from the NCAA tournament dropped him to #6. Some of that reflects the depth of the class: he’d be as high as #2 on PRISM’s board in other recent drafts. But the possibility that he could wind up as a “tweener” or even one of Engelmann’s “quagmires” has been increasing as his production hasn’t quite matched his scouting reports.

In typical Silver Bulletin fashion, we’ve got a lot of detail for you in the rest of this newsletter:

• Prospect profiles for Cameron Boozer, AJ Dybantsa, Peterson, and most of the other top ~20 players in the upcoming 2026 class

• Full PRISM rankings for the 2022, 2023, 2024, 2025 and 2026 classes

• A comparison of PRISM rankings versus consensus rankings

• Projected offensive and defensive roles

• Draft strength scores: Is the 2026 draft as good as reputed?

• Raw PRISM ratings based purely on stats, with no prior based on scouting rankings

• Projected player development trajectories and volatility ratings

• And a draft simulation factoring in young core, fit, NBA lifecycle, and more.

Here are PRISM’s full rankings. (Note that you’ll want to scroll through the pages to see beyond the top 15.)

Read more

🕒 The latest NFL QB ratings

Updated February 10, 2026.

These are our final ratings for the 2025-26 NFL season.

This is the landing page for QBERT1, our NFL quarterback rating system. As with our other landing pages, it’s best viewed on the web rather than over email or in the Substack App. The charts and tables will be updated regularly, and some of the text will change, too.

Subscribe now

Here’s our leaderboard of the top five quarterbacks as measured three ways: by their adjusted QBERT rating in 2025-26, by their 2025-26 WAR (wins above replacement), and then by their projected rating for their next forthcoming start.2 The projected rating also incorporates data from previous years and factors related to a QB’s age and experience.

Congratulations to the Super Bowl Champion Seattle Seahawks, a team that QBERT’s companion model, ELWAY, was fond of this year. QBERT considers Sam Darnold to be good rather than great, however — so he doesn’t appear on any of the leaderboards you see above. We have Darnold as the 14th-best QB in the league on a forward-looking basis and the 8th best in the league this year by WAR, where he gets some extra help since playoff stats count in our calculations. Darnold’s QBERT rarting for Super Bowl LX, 82.3, was close to his overall rating (86.4) for the year.

It was a much more disappointing performance for Drake Maye, QBERT’s choice for MVP. His rating for the game was 75.9, somewhere in between replacement-level and average, and that feels generous since his decent-looking raw stat line (27-43 for 295 yards and 2 TDs) was propped up by a largely noncompetitive second half. Although Maye has faced some tough defenses in the playoffs — although QBERT gives him credit for that — his shoulder injury may have been a little worse than the Patriots were initially letting on.

Thanks for following QBERT this season, and we’ll see you again in the fall! 🏈 —NS 2/10/26

See also: ELWAY NFL team projections.

QBERT in a nutshell

QBERT3 is both a rating system and a projection system. It evaluates virtually everything that a quarterback does on the field:

• Along with more traditional statistics, it incorporates components like rushing, generating first downs and even comeback wins that other systems may ignore.

• By accounting for pass pressure, yards after catch, and the performance of a team’s running backs, QBERT also attempts to allocate credit between the QB and other offensive players.

• QBERT ratings place past and present quarterbacks on a level playing field by adjusting for overall leaguewide trends, as well as the strength of the opposing defense, home-field advantage and even the weather.

But QBERT also projects every quarterback on a forward-looking basis, based on a rolling rating of his performance in recent games and recent seasons, with some assumptions4 baked in based on his experience, the aging curve, injuries and — for young quarterbacks — college performance. Lamar Jackson or Patrick Mahomes can survive a rough week and still rank in our top 5 in projected QBERT, but the rating system is more sensitive for less experienced quarterbacks.

For a thorough description of QBERT as well as our historical ratings, see our feature on the best quarterbacks of all time. The scale is designed to be similar to traditional NFL passer rating: 80 is average, 68 is replacement level, and anything 100 or above is great.

In addition to the leaderboard you saw at the top of this page, we have a lot of numbers to show you:

• QBERT ratings for each QB so far in 2025

• Projected QBERT ratings for each QB’s next start

• A breakdown of each QB’s weekly ratings for 2025, allocating credit between his passing performance, his rushing and other factors

• A series of charts with historical career tracking for each current QB15

• An updated lifetime leaderboard of the top QBs of all time

• Lifetime QBERT playoff leaders

• A spreadsheet of week-by-week ratings for all QBs since 1950

Other than the first table, the rest of this data is an exclusive for paying subscribers.

Who are the best QBs in the NFL?

Let’s see how that distinction between empirical QBERT and projected QBERT plays out in practice. Here’s our 2025 leaderboard based on performance for the season to date. It’s sorted by WAR, which accounts for adjusted QBERT ratings relative to the replacement-level line of 68 as well as the number of plays6 each QB makes. You can also see each QB’s unadjusted QBERT, which is nearly always higher than his adjusted QBERT because we’re living in a QB-friendly era. Except where otherwise specified, we’re almost always referring to adjusted QBERT when we describe a player’s rating.

In comparison, here are our projected QBERT ratings for the upcoming week. They incorporate performance prior to this season, as well as long-term factors like aging, experience, injuries, and college stats. Although we’ve listed projected QBs for every quarterback currently on an NFL roster, it’s important to note that the projections assume the QB starts the next game; projected QBERTs gradually deteriorate if a guy is sitting on the bench or the IR.

Read more

🕒 Our latest NFL team ratings

Monday, February 2

Happy Super Bowl week. We’ve updated our projections with the latest injury information. Although the Seahawks were already favored, this helps them further at the margin as the Patriots have a longer injury list. That includes QB Drake Maye, who is officially listed as questionable with a shoulder injury. To be honest, I’m not taking that entirely literally: it’s the Super Bowl, and he’s probably going to play unless he’s missing a limb or falls into the Springfield Mystery Spot. Instead, I have Maye listed as “probable”, which implies a 90 percent chance of suiting up (“questionable” would be 70 percent). But not being able to throw the deep ball reliably could considerably limit the Pats’ big-play potential.

ELWAY has had a pretty good postseason after a regular season in which, like a lot of prognosticators, it was disappointed by the Chiefs’ and Lions’ demise. (Annoyingly, also my favorite teams.) When the playoffs began, it had the Seahawks as the most likely Super Bowl champion by some margin, not necessarily because they were all that much better than the Rams — they weren’t last night — but because they’ve consistently been very good and getting a first-round bye plus getting to play at home in a stadium with a larger-than-average home-field-advantage is a huge positive.

In the AFC, we had the Broncos as 1a and the Patriots as 1b, though there was a gap between them, again mostly coming down to a bye and home-field advantage. Empirically, home field is the largest under worse weather conditions, which you’d think would help the Broncos in a literal blizzard. (Sean Payton should probably have checked the weather forecast.) But Drake Maye actually has more lifetime experience in cold weather than Jarrett Stidham, Bo Nix’s replacement, who had only four lifetime starts of any kind prior to Sunday. QBERT, our QB rating model that plugs into ELWAY, actually liked Maye’s performance, giving him a 96.9 rating for the game despite just 86 passing yards. Maye’s 6 rushing first downs were basically the thing that either team had going for them the entire game.

ELWAY updates its ratings after each game based not just on the final score but also lots of underlying data. Both the Seahawks’ and the Patriots’ ratings were essentially unchanged after last week. The Seahawks certainly had the conventionally better performance — the Patriots averaged just 3.2 yards per play — but it’s going to be very forgiving to a road team playing in a blizzard (yes, there’s a weather adjustment). The system actually thought the Rams had the best performance of the week, but they’re stuck watching the Super Bowl from the sidelines.

It’s no huge secret what either team’s biggest question is. For the Seahawks, Sam Darnold is a perfectly fine quarterback; his 86.6 QBERT rating is (slightly) above average. Still, this ranks just 108th of 120th Super Bowl starters all time. Meanwhile, the Patriots had the league’s softest regular-season schedule. And the QBs they faced in the playoffs — injured Justin Hebert, slumping C.J. Stroud, and Stidham — are not exactly a murderers’ row, although they faced three very tough defenses and the Pats’ own D has been outstanding.

Since our overall ELWAY ratings are public, you can see that we rate the Seahawks higher. While there are other factors the system considers that can sometimes complicate the forecast, the most important of these are 1) rest and 2) various elements of home-field advantage, not a factor in a neutral-site game after an extra week off. (Although the Seahawks will be playing in their natural time zone.)

So, yes, we have Seattle favored, although paid subscribers can scroll down to the “ELWAY future game projections” section for a more detailed forecast (including, yeah, a point spread, win probabilities, and an over/under line). Soon, we’ll also have a separate Super Bowl analysis coming for paid subscribers, with more analysis and even a guide to the most valuable Super Bowl squares.

Subscribe now

The rest of this page contains:

• Different versions of our ELWAY ratings;

• Odds for each remaining stage of the playoffs and winning the Super Bowl

• Detailed projections for this week’s games, and conditional projections for next week, including point spreads and totals (yes, there is gambling in this casino);

• And a series of 32 mini-graphs that track each team’s ELWAY rating from week to week.

The ratings themselves are free for everyone, while the other features are exclusive to paying subscribers. -NS, 2/2/26

See also: QBERT NFL quarterback ratings.

The best (and worst) teams in the NFL

The default version of ELWAY ratings includes adjustments based on a team’s projected starting QB for the upcoming game and current injuries. You can click the tabs in the table to view the other versions. “Healthy QB1s” backs out injuries completely for both QBs and non-QBs; the Bengals magically get Joe Burrow back. This might be a good indication of a team’s intrinsic strength later in the season. Meanwhile, the “team performance” version is most comparable to traditional NFL power ratings. It accounts for offseason roster movement and carries over a portion of the ratings from one season to the next, but otherwise takes a more “set it and forget it” approach.

ELWAY’s offensive and defensive ratings are essentially projections of how many points a team is expected to score and allow in forthcoming games. (Lower defensive ratings are better since it means a team is projected to allow fewer points.) By design, these ratings tend to be correlated, as you can see from this scatterplot:

Why such a strong relationship? Well, it’s because in a very literal sense, “a good offense is the best defense” in the NFL. Holding onto possession of the ball prevents the other team from scoring — defensive TDs are rare — eats up the clock, and leaves them with worse field position.

ELWAY projected records and playoff odds

We simulate the remainder of the season 10,0001 times based on these ratings, projected injury timelines, and a whole series of other adjustments. A team like the Packers will tend to have a bigger home-field advantage later in the season, for instance, because winter weather tends to help home teams that are accustomed to it. (There’s a reason that luxury housing is relatively affordable in Green Bay.)

Here are our projected playoff odds.

Read more

There are some minor differences between our approval rating calculations and our generic ballot numbers, which we’ll explain throughout the text. There are also some differences between these “simple” polling averages and the methods we use to calculate our election forecasts, which include additional steps and leverage the larger volume of polling data available (such as the ability to infer information from both state and national polls). Our forecast for the 2026 midterms will launch mid-year. This page describes solely our continually-updating averages.

Which polls are included

Our general aim is for inclusivity. We seek to include all professionally-conducted surveys. If you don’t see a poll listed, it may be because it’s included under a different name — we list the name of the polling firm rather than the media sponsor (for example, Beacon Research/Shaw & Co. rather than Fox News) — or because we haven’t gotten around to adding it. (Polling averages are typically updated ~6 times per week.) However, here are certain exceptions:

• We don’t use polls banned by Silver Bulletin because we know or suspect that the pollster faked data.

• We don’t use DIY polls commissioned by nonprofessional hobbyists on online platforms such as Google Surveys. These are becoming increasingly common. (Professional or campaign polls using these platforms are fine.)

• We don’t use “polls” that blend or smooth their data using methods such as MRP.

• We exclude polls that ask the voter who they support only after revealing leading information about the candidates. If, for instance, a poll says “Republicans hate puppies. Who do you plan to support: Republicans or Democrats?” we won’t include it.

Internal or campaign polls are included, provided they meet these other standards. For our approval rating averages, there is no distinction between partisan and nonpartisan polls. For the generic ballot, partisan polls are subject to different assumptions about “house effects” (see below). If ostensibly non-partisan pollsters have a history of producing polls for political campaigns or partisan organizations without disclosing these relationships, they may be automatically classified as partisan.2

For the generic ballot, if a pollster lists multiple versions of the same survey, we prioritize polls of likely voters > registered voters > all adults. For approval ratings, the opposite is true (although we calculate a separate average based on likely voter polls exclusively). For all of our polling averages, we prefer the version of the poll with “leaners” included. If there are other cases where the pollster uses different methods in the same survey, such as multiple turnout models, we simply average all applicable versions.

Also for the generic ballot, two slightly distinct questions are included in the average:

• Polls that ask voters whether they plan to vote for the Democratic or Republican candidate in the forthcoming election to Congress.

• Polls that ask voters which party they’d prefer to see control Congress.

How we weight polls

While all polls are included to the greatest extent possible, more reliable polls have more influence on the averages. We calculate an influence score for each poll based on three factors:

• The Silver Bulletin pollster rating, which in turn is based on its historical accuracy and whether it belongs to professional polling organizations that promote greater transparency and disclosure;

• The poll’s sample size — although there are diminishing returns to this. The importance of sample size is determined empirically. In practice, perhaps because of the complicated methods that pollsters use to weight their polls, sample size makes less difference than it does in theory, and there are considerable diminishing returns from adding additional voters to the sample.

• How recently the poll was conducted. This is also determined empirically, based on which settings best predict new polls over the subsequent two weeks. The settings for presidential approval ratings are more aggressive than for generic ballot numbers or for Elon Musk approval, because approval ratings tend to stabilize more quickly and shifts in the numbers are more likely to be signal rather than noise. For the generic ballot, weights become more aggressive as Election Day approaches.

We also take precautions to prevent any one polling firm from “flooding the zone” with a disproportionate number of surveys. Namely, the weights are based on the aggregate number of voters contacted by the pollster in a given time frame. In other words, if a pollster conducts a new survey every day or every week, it essentially “maxes out” the total weight assigned to the firm, and this weight is divided among all instances of the survey in a given window. Collectively, this firm’s polls may still have quite a lot of influence on our average, but any one from among the several surveys it conducts in a given window will be down-weighted. There is an additional adjustment for tracking polls in which the sample dates overlap.

More precisely, the polling average on a given date is calculated using local polynomial regression. Technically speaking, we use a blend of more aggressive and more conservative settings. However, these settings are determined empirically based on what produces more stable averages that are free from autoregression. In other words, the current polling average should be the best predictor of forthcoming surveys. If Trump approval is 46 percent on Tuesday and 43 percent on Wednesday, whatever average you calculate (i.e., 44 percent or 44.5 percent) should minimize error on what a new survey would say on Thursday.

While local polynomial regression is a fairly intuitive method, the right settings aren’t necessarily obvious. Too little smoothing can make the curve jut up and down unnecessarily and will result in overfitting of the data. If you smooth too much, however, the curve may be aesthetically pleasing but won’t do all that good a job of describing the data and may be slow to catch up to new trends. As an empirical matter, we’ve found that generic ballot can suffer from under-smoothing, making the average too “bouncy,” while this is less true for approval ratings, and our settings are more aggressive.

House effects

House effects are persistent differences between a given firm’s polls and those of other firms that survey the same race. For example, if a given firm’s polls are on average 3 points more favorable to Democratic candidates than the average of other polls conducted at the same time, we back a proportion of that difference out (in this example, this would mean shifting the firm’s surveys toward Republicans). The proportion subject to the house effects adjustment is based on the aggregate sample size of polling conducted by the firm. For pollsters that survey a given question extremely often, meaning that we can be highly confident about its house effect, close to 100 percent of the difference between a given firm’s polls and the polling average may be adjusted for. For a firm that has surveyed a race just one or twice, only a small fraction is, conversely. Our generic ballot average uses a slightly different formula than our approval ratings, which starts out with a lower adjustment but ramps it up more quickly as the aggregate sample size increases. We believe this formula is slightly more accurate and will probably eventually transition our approval rating averages to using it, although in practice, it only makes a difference in the decimal places.

More precisely, a separate house effects adjustment is applied to approval and disapproval ratings (or Democrats and Republicans in the case of the generic ballot). Certain firms persistently include more or fewer undecided voters than the polling average, and so we adjust for this, too.

Firms with higher pollster ratings have more influence on what our model considers to be “true north”, i.e. the unbiased average that other polls are adjusted toward. For the generic ballot, polls with partisan sponsors are excluded from our calculation of the “true” average.

The house effects adjustment is determined through an iterative process. Essentially, we first calculate a polling average without house effects, then adjust for house effects based on how a given firm’s polls line up relative to the polynomial curve, then recalculate the polling average with house-effect-adjusted polls, then recalculate house effects based on the new version of the average, and so forth. The model loops through the process several times until there are no further gains to be had.

For the generic ballot, explicitly partisan polls start out with a prior: in other words, we assume they’re slightly biased toward the sponsoring candidate or party. (More specifically, by a net of about 1.7 points. That is, if the generic ballot polling average is D +2, we’d expect a partisan poll to show something like D +3 or D +4 instead.) For nonpartisan polls, we use a zero prior instead (we default toward assuming unbiasedness). In both cases, the prior is eventually overridden by the hard evidence on how a firm’s surveys line up relative to others that ask the same question. This prior is applied to the generic ballot and the “horse race” polling averages used in our forecasts, but not to presidential approval ratings.

Error bars

Our generic ballot and approval rating averages incorporate an estimate of uncertainty, shown with shaded regions on our charts. The shaded areas indicate where we project 90 percent of new polls to fall based on our historical analysis of these averages over the past several decades. The error bars are wider when there are fewer polls of race and when the polls are less consistent with one another.

State benchmarks (generic ballot only)

For the generic ballot average, we calculate a current benchmark in each state based on the following formula:

Benchmark = generic ballot + partisan lean score * elasticity

The partisan lean score is calculated based on a combination of how states have voted relative to the rest of the country in the past two presidential elections and in recent state legislative elections. More precisely, the formula gives 50 percent of the weight to the 2024 presidential election, 25 percent to the 2020 presidential election, and 25 percent to a rolling average of the past five state legislative elections.3 Meanwhile, elasticity is a longstanding Silver Bulletin/FiveThirtyEight concept. Some states, like New Hampshire, have a larger proportion of independents and swing voters, typically upper-middle-class suburbanites. Whereas in other states, like Mississippi, everyone is pretty much either extremely Democratic (mainly Black voters) or extremely Republican (mainly white evangelicals).

Our current elasticity scores, calculated from microdata in the 2022 and 2024 Cooperative Election Studies, which each surveyed 60,000 voters, range from 0.90 in Mississippi to 1.23 in Alaska and 1.25 in Hawaii. More precisely, the elasticity scores are calculated based on assigning a probability of each voter in the CES voting for the Democratic or Republican candidate. If a voter is toward the tail end of the bell curve, i.e. with nearly a 100 percent chance of voting for the Republican or the Democrat, shifts in the political environment likely make little difference in their vote. Conversely, they’re more likely to switch preferences if they start out somewhere near 50/50. Technically speaking, these probabilities are based on an average of two formulas, one that accounts for political questions in the survey (specifically, how strongly a voter leans toward the Democratic or Republican parties and where they rate themselves on a spectrum from liberal to conservative) and one that is exclusively based on demographic data. A “cross-pressured” voter (i.e., one who has some characteristics that typically predict Democratic voting and others that predict Republican voting) will intrinsically start out closer to 50/50, and so states with more of these voters have higher elasticity.

Current elasticity scores are listed below.

We’ll update this document in the future with any significant changes to the methodology.

Do you want to know which American pollsters rank highest (and lowest) based on historical accuracy and transparency? Curious about how accurate a particular pollster might be in future elections? Well, you’ve come to the right place.

Welcome to the landing page for the Silver Bulletin pollster ratings. Below, you’ll find our most up-to-date data on how accurate and transparent each pollster in our database has been in past elections — and how accurate it might be in the future. Plus a ton of extra data on pollster quality.

Subscribe now

These ratings apply the same methodology as the previous, Nate-era FiveThirtyEight pollster ratings. You can find the data from previous iterations of our ratings below:

• February, 2025 update.

• June, 2024 update.

We’ve since updated the ratings with polls and election results from 2025, namely:

• The 2025 Virginia gubernatorial election

• The 2025 New Jersey gubernatorial election

• Special elections for the U.S. House in 2025

Because 2025 was an off-year, we’ve only added about 40 polls since the last update update. So don’t expect massive shifts in the ratings. The top-rated pollsters — The Washington Post1, Marquette University, NYT/Siena — will still be familiar to longtime readers.

Note that we don’t include polls of the New York City mayoral election in our database. (Historically, we haven’t used mayoral polling when calculating pollster ratings, and there’s no principled reason to start using polls from NYC but not from other mayoral races.) That means pollsters won’t be penalized for their misses in the Big Apple.

Since there haven’t been any methodological changes since our last update, let’s just go ahead and get to the numbers. The columns in the main pollster ratings table are as follows:

• An overall grade based on a pollster’s Predictive Plus-minus rating. Grades for pollsters with sparse data are rounded (e.g., to “A/B” rather than A-) and banned pollsters automatically receive a grade of F.2

• The Predictive Plus-Minus rating itself, which is how accurate the model expects the pollster to be going forward based on a combination of its historical accuracy and its transparency/disclosure standards — pollsters get a bonus if they’re either a member of the AAPOR Transparency Initiative or share their data with the Roper Center archive. Negative plus-minus scores are good and imply that we expect the pollster to be more accurate than average going forward.

• Mean-reverted bias — that is, a pollster's historical average statistical bias toward Democratic or Republican candidates, reverted to a mean of zero based on the number of polls in the database. This is calculated only for races in which exactly one Republican and one Democrat are the two leading candidates (so it doesn’t apply for presidential primaries, for instance). A score of "D +1.5", for example, means that the pollster has historically overrated the performance of Democratic candidates.

• Finally, the number of polls included in the calculation. The database goes back to 1998, though polls from more recent years are weighted more heavily. In general, these ratings cover all polls in the 21 days prior to presidential, Congressional and gubernatorial general elections, and presidential primaries.

But wait, there’s more. Predictive-Plus-Minus might be what we use to assign letter grades, but it isn’t the only measure of pollster accuracy we generate:

• Simple Average Error is the firm's average error, calculated as the difference between the polled result and the actual result for the margin separating the top two finishers in the race.

• Simple Plus-Minus is a firm's Simple Average Error minus the expected error for the race the firm surveyed, which accounts for the type of election polled, the number of days until the election, and the poll's sample size.

• Advanced Plus-Minus is (as advertised) a more advanced plus-minus score that also accounts for the performance of other polling firms surveying the same races, and which weights recent results more heavily.

• Mean-Reverted Advanced Plus-Minus is the Advanced Plus-Minus score, reverted to a mean of zero. The amount of mean reversion is based on how many polls the firm has conducted, but with polls in previous years discounted.

Note that for this chart and the following one, we’re only listing survey firms with at least 20 polls in our database, since most of these measures aren’t mean-reverted and are subject to large sampling errors. You can find ratings for all firms in the Excel file blow, but interpret them with extreme caution for pollsters that only survey elections irregularly.

We also calculate other data about how each pollster has performed in past elections. Including how often a pollster called the winner of the race correctly3 and missed outside the margin of error, its house effect, and how much we penalize it for potential herding.

House effects show how a firm’s results compare against other polls. For example, if a pollster shows the Republican candidate leading by 3 points in a race that every other pollster had tied, that firm would have a Republican house effect. Note that house effects are distinct from statistical bias. If the Republican candidate actually won this hypothetical race by 5 points, our imaginary pollster would still have a 2-point Democratic bias even with their Republican house effect.

Our Average Distance from Polling Average (ADPA) score measures how far a firm's average poll differed from the average of other polls in the field at the time it was released. A low ADPA is potential evidence of herding. A herding penalty is triggered when a firm's ADPA is lower than the theoretical minimum based on the sampling error in its polls. This penalty is added to a firm's Advanced Plus-Minus score before calculating Predictive Plus-Minus.

Finally, here are two files containing even more data. The first file includes all of the data shown in the tables above, alternative versions of the plus-minus calculations, and details about the methodological transparency of each pollster.

Pollster Stats Full 2026

129KB ∙ XLSX file

Download

Download

And here is the raw data used to calculate the averages, including topline numbers from more than 12,300 polls. You are welcome to use this data free of charge for any purpose, but please attribute it to Silver Bulletin.

Rawpolls 011226

1.95MB ∙ XLSX file

Download

Download

Do you want to know which American pollsters rank highest (and lowest) based on historical accuracy and transparency? Curious about how accurate a particular pollster might be in future elections? Well, you’ve come to the right place.

Welcome to the landing page for the Silver Bulletin pollster ratings. Below, you’ll find our most up-to-date data on how accurate and transparent each pollster in our database has been in past elections — and how accurate it might be in the future. Plus a ton of extra data on pollster quality.

Subscribe now

These ratings apply the same methodology as the previous, Nate-era FiveThirtyEight pollster ratings.

You can find the data from our previous ratings in June 2024, which were in effect for our 2024 presidential election forecast, here.

We’ve since updated the ratings with polls and election results since June 2024, namely:

• The 2024 presidential election

• The 2024 congressional elections, and

• The 2024 gubernatorial elections

As after any presidential election year, we've added a lot of new data — about 460 polls since the last update.1 So we’ve seen some shifts in the ratings. But the top-rated pollsters — AtlasIntel, Marquette University, The Washington Post2 — will still be familiar to longtime readers.

Since there haven’t been any methodological changes since our last update, let’s just go ahead and get to the numbers. The columns in the main pollster ratings table are as follows:

• An overall grade based on a pollster’s Predictive Plus-minus rating. Grades for pollsters with sparse data are rounded (e.g., to “A/B” rather than A-) and banned pollsters automatically receive a grade of F.3

• The Predictive Plus-Minus rating itself, which is how accurate the model expects the pollster to be going forward based on a combination of its historical accuracy and its transparency/disclosure standards — pollsters get a bonus if they’re either a member of the AAPOR Transparency Initiative or share their data with the Roper Center archive. Negative plus-minus scores are good and imply that we expect the pollster to be more accurate than average going forward.

• Mean-reverted bias — that is, a pollster's historical average statistical bias toward Democratic or Republican candidates, reverted to a mean of zero based on the number of polls in the database. This is calculated only for races in which exactly one Republican and one Democrat are the two leading candidates (so it doesn’t apply for presidential primaries, for instance). A score of "D +1.5", for example, means that the pollster has historically overrated the performance of Democratic candidates.

• Finally, the number of polls included in the calculation. The database goes back to 1998, though polls from more recent years are weighted more heavily. In general, these ratings cover all polls in the 21 days prior to presidential, Congressional and gubernatorial general elections, and presidential primaries.

But wait, there’s more. Predictive-Plus-Minus might be what we use to assign letter grades, but it isn’t the only measure of pollster accuracy we generate:

• Simple Average Error is the firm's average error, calculated as the difference between the polled result and the actual result for the margin separating the top two finishers in the race.

• Simple Plus-Minus is a firm's Simple Average Error minus the expected error for the race the firm surveyed, which accounts for the type of election polled, the number of days until the election, and the poll's sample size.

• Advanced Plus-Minus is (as advertised) a more advanced plus-minus score that also accounts for the performance of other polling firms surveying the same races, and which weights recent results more heavily.

• Mean-Reverted Advanced Plus-Minus is the Advanced Plus-Minus score, reverted to a mean of zero. The amount of mean reversion is based on how many polls the firm has conducted, but with polls in previous years discounted.

Note that for this chart and the following one, we’re only listing survey firms with at least 20 polls in our database, since most of these measures aren’t mean-reverted and are subject to large sampling errors. You can find ratings for all firms in the Excel file blow, but interpret them with extreme caution for pollsters that only survey elections irregularly.

We also calculate other data about how each pollster has performed in past elections. Including how often a pollster called the winner of the race correctly4 and missed outside the margin of error, its house effect, and how much we penalize it for potential herding.

House effects show how a firm’s results compare against other polls. For example, if a pollster shows the Republican candidate leading by 3 points in a race that every other pollster had tied, that firm would have a Republican house effect. Note that house effects are distinct from statistical bias. If the Republican candidate actually won this hypothetical race by 5 points, our imaginary pollster would still have a 2-point Democratic bias even with their Republican house effect.

Our Average Distance from Polling Average (ADPA) score measures how far a firm's average poll differed from the average of other polls in the field at the time it was released. A low ADPA is potential evidence of herding. A herding penalty is triggered when a firm's ADPA is lower than the theoretical minimum based on the sampling error in its polls. This penalty is added to a firm's Advanced Plus-Minus score before calculating Predictive Plus-Minus.

Finally, here are two files containing even more data. The first file includes all of the data shown in the tables above, alternative versions of the plus-minus calculations, and details about the methodological transparency of each pollster.

Pollster Stats Full 2025

126KB ∙ XLSX file

Download

Download

And here is the raw data used to calculate the averages, including topline numbers from more than 12,300 polls. You are welcome to use this data free of charge for any purpose, but please attribute it to Silver Bulletin.

Rawpolls 021325

1.75MB ∙ XLSX file

Download

Download

🕒 Last updated: 10:20 a.m., Wednesday, April 9

Congrats to the Florida Gators for winning the national title after a clutch comeback against Houston.

We’ll work to restore this page so it reflects our original pre-tournament projections while also providing spreadsheets with the interim projections we made throughout the tournament — though give us a moment with all the other news in the world this week.

Our season-ending men’s SBCB ratings can be found here.

We’ve been hard at work on our college basketball forecasts this year. We’ve recalibrated them based on analyzing more than 250,000 historical college basketball games. We now have our own fully-fledged rating system — SBCB — though our ratings are combined with other systems for purposes of our tournament forecasts. And there are some new wrinkles this year, too, like a better formula to account for injuries¹ — though if you’re reading this from Durham, North Carolina, you’ll be glad to know that it looks like Cooper Flagg will be OK.

Still, March Madness projections are an annual tradition in these parts. We’ve run tourney forecasts through literally every iteration of Silver Bulletin/FiveThirtyEight dating back to 2011. Some of the code1 in the model goes back even earlier than that from when I was trying to win my friend’s unnecessarily complicated office pool as a junior associate at KPMG back in 2002 or 2003. The tournament was also key to our launch strategy at the Disney-fied version of FiveThirtyEight in 2014. And if I’m being honest, it was an inflection point for this newsletter. I was on the fence about running the 2024 version, but I figured it was worth a few hours of work on an idle Sunday. Instead, it took until 3 in the morning. But our paid subscription count basically tripled overnight to the point where I thought it was some sort of glitch. That was all the proof of concept we needed to reboot the election forecast and other things.

How this works

• We’ll update these numbers once per day (often in the morning) after tournament games are played. If you’re reading this by email, you’ll want to use the web version because we can’t magically change the text in your inbox.

• These ratings give half the weight to our Bayesian SBCB ratings2 and half to a composite of other systems, namely:

  • KenPom3

  • The Sonny Moore power ratings

  • ESPN BPI

  • The Massey ratings

  • And the NCAA tournament committee’s S-Curve rankings, i.e. how it seeds all tournament teams from 1 through 68.

• 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. (Getting “hot” in the tourney is a thing.) For more detail about how all of this works behind the scenes, see here.

Let’s cut to the chase: who’s the most likely winner?

Share

Usually, there’s strong alignment between who we think the best teams are and who has the best odds. But this isn’t always the case. There can be discrepancies introduced based on how easy or difficult each team’s draw is, injuries, and the geographic location of each game — plus who’s managed to survive in the tournament so far.

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

• Projected odds for all rounds in all four regions; the South is free for all readers, while the other regions are a benefit for paying subscribers.

• Forecasted win probabilities and margins of victory — which conveniently double as point spreads — for forthcoming games.

• A spreadsheet version of these forecasts with additional precision.

• Additional tables and spreadsheets comparing all rating systems we use on a common scale, plus current injury adjustments.

Subscribe now

I’ll probably send an update or two out over the email list, but we won’t spam you; approximately 27 percent of Silver Bulletin subscribers are degenerate gamblers extremely interested in the tournament while others are not interested at all. For notifications of updates, please follow me on Twitter. Or perhaps better, Substack Notes. It won’t be an overnight transition — but I’m planning to gradually do more of my “microblogging” on Notes given the problems with the alternatives.

Also, we know from past experience that many people subscribe just for the tournament forecasts and then don’t stick around so long. We won’t take offense, and we appreciate your business either way. But we do think the annual subscription price is the better deal, and we plan to gradually reintroduce various sports and politics models over the course of the next year.

Here are our updated odds for the four regions:

Read more

🕒 Last updated: 8:30 a.m., Monday, April 7

Congrats to the UConn Huskies for winning their first title since 2016 and for doing so in an absolutely dominant fashion!

We’ll work to restore this page so it reflects our original pre-tournament projections while also providing spreadsheets with the interim projections we made throughout the tournament — though give us a moment with all the other news in the world this week.

Our season-ending women’s SBCB ratings can be found here.

I don’t want to turn this into too much of a “discourse.” Still, back in the FiveThirtyEight days, we were into women’s basketball before it was cool, running women’s tournament projections along with the men’s bracket and finding that there was a bigger audience for this than you might think.

Now, women’s hoops are much more mainstream. In fact, TV ratings for the women’s final crushed the guys’ numbers last year. While that might not be true this time around — no Caitlin Clark — March and April are much richer months for basketball fans when there are 134 games to watch instead of “just” 67 on the men’s side.

So, I want to uphold that tradition. In fact, we’ve been hard at work on this, creating a women’s version of our Elo-flavored SBCB ratings and tweaking parameters to be more specific to the women’s game instead of just importing code from the men’s version. There are some notable differences, particularly that American women can’t get drafted into the WNBA until age 22, a practice I consider unjust. But it means there’s more carryover in our ratings from season to season. Also, in the women’s tourney, top seeds basically host home games in the first two rounds. Our SBCB ratings calculate team-specific home-court advantage ratings — the more popular programs are generally even more dominant at home — and these are now factored into our projections.

The only real downside is that women’s basketball tends to be fairly chalky in the early rounds — favorites usually win, typically by big margins — notwithstanding the legendary upset by 16-seeded Harvard over Stanford in 1998. In general, the women’s version of the sport is more lopsided. But from the Sweet Sixteen onward, it’s usually a great tournament.

Apart from that, the process here is pretty similar to the men’s version.

How this works

• We’ll update these numbers once per day after tournament games are played.

• These ratings give half the weight to our Bayesian SBCB ratings and half to a composite of other systems. There are fewer of these on the women’s side, but namely:

  • The Sonny Moore power ratings

  • The Massey ratings

  • HerHoopStats ratings — newly added this year because I’m an idiot and hadn’t known about it before — great site by the way.

  • And the NCAA tournament committee’s S-Curve rankings, i.e. how it seeds all teams from 1 through 68. Typically, the committee has not released a complete 1-to-68 list for women, only designating the order of the four #1 seeds, so ties are broken by the NCAA’s quasi-official NET rating system.

• They also account travel distance, and — for the early rounds when games are hosted at campus sites — home court advantage. And newly this year, we do make some attempt to account for injuries for women too, though data on them is much sparser than on the men’s side.1 Also, once the tournament is underway, we look at how teams have performed so far in the tourney relative to the model’s expectations. For more detail about how all of this works behind the scenes, see here.

Drumroll, please …. here are the favorites:

Share

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

• Projected odds for all rounds in all four regions, along with brief commentary; the first region is free for all readers, while the other regions are a benefit for paying subscribers.

• Forecasted win probabilities and margins of victory — which conveniently double as point spreads — for forthcoming games.

• A spreadsheet version of these forecasts with additional precision.

• Additional tables and spreadsheets comparing all rating systems we use on a common scale, plus current injury adjustments.

Subscribe now

Read more

🕒 The latest

April 8, 2025. These are our season-ending women’s SBCB ratings for 2024-25! They’ll be back next season! Congratulations to Florida for winning the national title; they’re also the best team in both versions of our ratings. For the women’s version of SBCB ratings, see here.

Welcome to the homepage for the men’s edition of our Silver Bulletin College Basketball (SBCB) ratings. It will be updated roughly once per day, either in the evening or the morning after the full slate of games is played.1 If you’re viewing this in the Substack app or your email client, you may want to click over to the web version to see the latest data.

There are two versions of SBCB: a “pure” version that essentially only accounts for performance on the court and a “Bayesian” version that also factors in human ratings — specifically preseason ratings in the AP and Coaches polls. For instance, Kansas has a considerably higher rating in the Bayesian version because they were the preseason #1 — although their performance hasn’t lived up to their advance billing. Despite this, the Bayesian version generally yields more accurate predictions of upcoming games. Here is the current top 10 in each system:

The pure version of SBCB is free for all readers, while the Bayesian version requires a paid subscription. Paid subscribers will also get a spreadsheet showing how to translate Elo ratings to win probabilities and projected margins of victory (i.e. point spreads). And they’ll get full access to our men’s and women’s NCAA tournament forecasts once those launch, which blend SBCB with other ratings.

Subscribe now

SBCB is an Elo-based rating system, though with many enhancements and modifications. For a more complete methodological description, see here. I’ll just go over the basics in a few bullet points.

• The most important factor in Elo is simply winning, especially against strong opponents. A team can never lose Elo status in any game it wins — but it gains far more credit for winning games as an underdog.

• However, our formula also accounts for margin of victory and home court.

• Home-court advantage consists of two factors. First, a rolling score specific to each school indicating its tendency to overperform or underperform its Elo rating in home games — teams like Purdue and Duke that are reputed to have substantial home-court advantages often do. Second, a travel distance factor.2

• SBCB ratings never reset to zero. Games from previous seasons have some influence. This adds predictive value since the college basketball season is relatively short, though it may make SBCB less appropriate for comparing tournament resumes based on the current season’s play.

• There are some special procedures at the start of each season. In the pure version of SBCB, each team’s rating is reverted toward the mean of its conference. In the Bayesian version, the mean reversion process considers human polls in addition to conference strength. Despite this, SBCB is generally more bullish on mid-major and “minor” conference teams than its competitors.

• Although the ratings account for more than 250,000 games (!) going all the way back to the 1949-1950 season3, we’ve made some modifications to reflect the changing nature of the sport. There’s more mean reversion in recent seasons, reflecting higher player turnover. And the travel distance component is lesser because of increasingly luxurious accommodations.

Let’s keep the rest of this simple. Here are the pure SBCB ratings for those of you who don’t want any human-flavored inputs:

And here is the Bayesian version, which should be slightly better for predictions:

Read more

This is simply an archive of our men’s NCAA tournament predictions as of the late afternoon on March 18, which we’re posting now that the tournament — well, First Four games, which probably won’t affect your office pool — are underway. It’s posted solely for accountability purposes so that you can check our work later, as the main tournament page will update to reflect results so far. All data and text are preserved “as is” as of ~3:30 pm on 3/18/25. You’ll probably want to click over to the main tournament page unless you specifically want to assess the original forecasts. Comments are off on this post, but we’re always happy to welcome new subscribers.

Subscribe now

We’ve been hard at work on our college basketball forecasts this year. We’ve recalibrated them based on analyzing more than 250,000 historical college basketball games. We now have our own fully-fledged rating system — SBCB — though our ratings are combined with other systems for purposes of our tournament forecasts. And there are some new wrinkles this year, too, like a better formula to account for injuries1 — though if you’re reading this from Durham, North Carolina, you’ll be glad to know that it looks like Cooper Flagg will be OK.

Still, March Madness projections are an annual tradition in these parts. We’ve run tourney forecasts through literally every iteration of Silver Bulletin/FiveThirtyEight dating back to 2011. Some of the code2 in the model goes back even earlier than that from when I was trying to win my friend’s unnecessarily complicated office pool as a junior associate at KPMG back in 2002 or 2003. The tournament was also key to our launch strategy at the Disney-fied version of FiveThirtyEight in 2014. And if I’m being honest, it was an inflection point for this newsletter. I was on the fence about running the 2024 version, but I figured it was worth a few hours of work on an idle Sunday. Instead, it took until 3 in the morning. But our paid subscription count basically tripled overnight to the point where I thought it was some sort of glitch. That was all the proof of concept we needed to reboot the election forecast and other things.

How this works

• We’ll update these numbers once per day (usually in the morning) after tournament games are played. If you’re reading this by email, you’ll want to use the web version because we can’t magically change the text in your inbox.

• These ratings give half the weight to our Bayesian SBCB ratings3 and half to a composite of other systems, namely:

  • KenPom4

  • The Sonny Moore power ratings

  • ESPN BPI

  • The Massey ratings

  • And the NCAA tournament committee’s S-Curve rankings, i.e. how it seeds all tournament teams from 1 through 68.

• 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. (Getting “hot” in the tourney is a thing.) For more detail about how all of this works behind the scenes, see here.

It’s a wide-open tournament this year, but let’s cut to the chase: who’s the most likely winner? I’m not excited about this, because like 98 percent of well-adjusted college basketball fans, I’m not Crazy for Cameron. But the answer is … Duke, by a hair.

Share

You can see that, for now, there’s strong alignment between who we think the best teams are and who has the best odds. (Though we disagree with the selection committee, who has Auburn as the #1 overall seed.) But this isn’t always the case. There can be discrepancies introduced based on how easy or difficult each team’s draw is, injuries, and the geographic location of each game — plus who’s managed to survive in the tournament so far.

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

• Projected odds for all rounds in all four regions, along with brief commentary; the South is free for all readers, while the other regions are a benefit for paying subscribers.

• Forecasted win probabilities and margins of victory — which conveniently double as point spreads — for forthcoming games.

• A spreadsheet version of these forecasts with additional precision.

• Additional tables and spreadsheets comparing all rating systems we use on a common scale, plus current injury adjustments.

As I’ve said, these forecasts will be updated once daily — but as a heads-up, we’ll likely pare down on the commentary once the tournament is underway just because we figure speed and the numbers are at a premium relative to me pontificating. Once the action gets started, I’ll create an archived version of this page so you can see just how right or wrong the original forecasts were, even if our bracket gets busted.

I’ll probably send an update or two out over the email list, but we won’t spam you; approximately 27 percent of Silver Bulletin subscribers are degenerate gamblers extremely interested in the tournament while others are not interested at all. For notifications of updates, please follow me on Twitter. Or perhaps better, Substack Notes. It won’t be an overnight transition — but I’m planning to gradually do more of my “microblogging” on Notes given the problems with the alternatives.

Also, we know from past experience that many people subscribe just for the tournament forecasts and then don’t stick around so long. We won’t take offense, and we appreciate your business either way. But we do think the annual subscription price is the better deal, and we plan to gradually reintroduce various sports and politics models over the course of the next year.

Subscribe now

The South

Can we be honest? It’s not the strongest region. Our SBCB ratings and most other computer systems actually regard Auburn as the weakest #1 seed, not the best. The #3 seed in the South is Iowa State, another team that SBCB regards as weaker than the consensus. Plus, there are two play-in games, one involving the worst team in the bracket, St. Francis. The Red Flash have a 1 in 375,396,779 chance of winning the Big Dance, so never say never, I guess.

That could open the door to my hometown Michigan State Spartans, although Sparty just lost in the Big Ten semifinals. For a deeper sleeper, consider Louisville, whom we have as the 15th-best team — in other words, they should have been a #4 seed, not an 8. UC San Diego — not to be confused with plain old San Diego, or San Diego State — also has a relatively high chance of pulling off a classic 12-vs-5 upset versus Michigan.

The West

Now, this one’s a bit more action packed.

Read more

These are our season-ending women’s SBCB ratings for 2024-25! They’ll be back next season! For the men’s version of SBCB ratings, see here.

Welcome to the homepage for the women’s edition of our Silver Bulletin College Basketball (SBCB) ratings. It will be updated roughly once per day, either in the evening or the morning after the full slate of games is played.1 If you’re viewing this in the Substack app or your email client, you may want to click over to the web version to see the latest data.

There are two versions of SBCB: a “pure” version that essentially only accounts for performance on the court and a “Bayesian” version that also factors in human ratings — specifically preseason ratings in the AP and Coaches polls. For instance, on the men’s side, Kansas has a considerably higher rating in the Bayesian version because they were the preseason #1 — although their performance hasn’t lived up to their advance billing. Despite this, the Bayesian version generally yields more accurate predictions of upcoming games. Here is the current top 10 in each system:

Note that the top women’s teams tends to have higher ratings than the men, as there is more dominance in the women’s game and more lopsided outcomes.

The pure version of SBCB is free for all readers, while the Bayesian version requires a paid subscription. Paid subscribers will also get a spreadsheet showing how to translate Elo ratings to win probabilities and projected margins of victory (i.e. point spreads). And they’ll get full access to our men’s and women’s NCAA tournament forecasts once those launch, which blend SBCB with other ratings.

Subscribe now

SBCB is an Elo-based rating system, though with many enhancements and modifications. For a more complete methodological description, including some subtle differences for the women’s version of the ratings, see here. I’ll just go over the basics in a few bullet points.

• The most important factor in Elo is simply winning, especially against strong opponents. A team can never lose Elo status in any game it wins — but it gains far more credit for winning games as an underdog.

• However, our formula also accounts for margin of victory and home court.

• Home-court advantage consists of two factors. First, a rolling score specific to each school indicating its tendency to overperform or underperform its Elo rating in home games — men’s teams like Purdue and Duke that are reputed to have substantial home-court advantages often do. Home-court advantage is generally lower in the women’s game, though the most popular programs tend to have more of one. Second, a travel distance factor.2

• SBCB ratings never reset to zero. Games from previous seasons have some influence. This adds predictive value since the college basketball season is relatively short, though it may make SBCB less appropriate for comparing tournament resumes based on the current season’s play.

• There are some special procedures at the start of each season. In the pure version of SBCB, each team’s rating is reverted toward the mean of its conference. In the Bayesian version, the mean reversion process considers human polls in addition to conference strength.

• The women’s ratings account for more than 100,000 games3 going back to the 2002-03 season4. Some of the modifications we’ve made to reflect the changing nature of men’s basketball don’t apply to the women. For instance, there’s higher ratings carryover from season to season because players can’t enter the WNBA until they’re 22, meaning they retain more of their top talent.

Let’s keep the rest of this simple. Here are the pure SBCB ratings for those of you who don’t want any human-flavored inputs:

And here is the Bayesian version, which should be slightly better for predictions:

Read more