A recent analysis piece on Sky Sports looked at the “meat wall” corner kick tactic of placing attackers in the path of the keeper to prevent him from claiming the ball even when it is served into the center of the six yard box. David Reed found that the number of attackers in the six yard box on corners has increased by about 70 percent over just two seasons. In another study, Laura Hunter identified that as the season has gone on, teams have changed their tactics on corner kicks to prioritize inswingers to the goal mouth, with Liverpool having a particularly dramatic shift. This strategy is gaining new converts rather than slowing down.

On throw-ins, there are even proposed rules reforms. Setting up a long throw takes time, especially to move the defenders into position and bring the long throw specialist over to pick up the ball. A recent piece on Opta Analyst by Ali Tweedale found that the time spent with the ball out of play before throw-ins has increased by nearly 20 percent this season, from a little under 9 minutes per match to a little under 11 minutes per match. IFAB, the commission responsible for setting rules across football, has proposed to empower referees to begin a five-second count if a team seems to be taking too long to play a throw-in or a goal kick. It remains to be seen if this time limit will be enforced in a way that makes the long throw tactic less effective or simply speeds it up a little bit.

And on the pitch, the set piece goals keep coming. Set pieces account for just under 30 percent of goals scored this season, up from a little over 20 percent in recent seasons. And as the last newsletter shows, this sudden tactical shift in the Premier League has arisen not from a complex array of factors but from two clearly identifiable new ideas: the long throw and the “meat wall” corner kick.

At the same time as these tactics have increased goal production from set pieces, open-play goal scoring remains depressed. As a result, corners and throw-in set pieces combined have accounted for about 23 percent of goals scored in the Premier League this season, compared to an average of about 15 percent in recent seasons. Nearly a quarter of goals now come from corners or long throws.

How did this happen? Who is responsible for this sudden tactical shift, and how did it take over the Premier League so quickly?

A Revolution, But Who Are the Revolutionaries?

This question has a short answer, and then it also has a long answer that digs back into an earlier era in the Premier League and offers a set of contrasts that can help us understand better what is going on today. That longer answer will take up much of the length of this newsletter. It will also introduce the paradox of the Rory Delap Trap.

Still, there is a short answer.

It was Brentford and Arsenal.

In the 2025–26 season, fully ten Premier League clubs have taken at least half of their deep throw-ins long into the penalty area. In the two preceding seasons, only Brentford at about 64 percent broke that barrier. Likewise Mikel Arteta’s Arsenal, with over 75 percent of their corner kicks played to the six-yard box, stand out as the clear innovators whose corner tactics are now being copied by the rest of the league. (Arsenal’s claims prevented numbers in 2024–25 also lapped the league before the “meat wall” was copied by many of their competitors.)

One striking finding here is that the two tactics have distinct innovators, and no club before this season was executing them both. Brentford has retained a strategy on corners that makes less use of crowding the keeper than most, while Arsenal still do not play the majority of their deep throw-ins long. Instead, new adopters like Everton, Crystal Palace, Spurs and Bournemouth have taken ideas from both Arsenal and Brentford to try to maximize all of their set pieces.

It is important to grasp why Premier League teams have so suddenly altered their approach to set pieces. Clubs and managers have adopted these tactics because they work. And this is true not merely at the level of individual plays, as the previous newsletter demonstrated, but also at the team level. Clubs that adopted these tactics have been rewarded with better set piece results.

The New Corner Kick Model Takes Over

These new corner kick tactics have clearly been effective over the last several seasons for clubs which have chosen to adopt them, first Arsenal and then a spate of copycats.

And just as many teams adapted their set piece strategies over the last summer to adopt Arsenal’s and Brentford’s innovations, this has also happened during the season. Along with Liverpool, as Laura Hunter had noted, West Ham likewise in the late fall began targeting the six-yard box at much higher rates.

The speed of the adoption of these corner kick tactics raises an obvious question. While it is not hard to understand why clubs have picked up strategies that help them score more goals from set pieces, why did it take until the last few seasons for teams to find these innovations? There was a longstanding tradition, in English football particularly, of focusing on set pieces as a way to even out the advantages of more skilled and technical possession sides. Were they unable to identify the right tactics?

The Pre-History of the Meat Wall

When looked at over a longer time horizon, the sudden development of a dominant corner strategy in the last two seasons appears as a spike following a long trough. Back in 2009–10 and 2010–11 in particular, there were more teams playing their corner kicks directly to the center of the penalty area, and even, if to a lesser degree, preventing opposition goalkeeper claims.

Sam Allardyce’s Blackburn Rovers, Tony Pulis’ Stoke City and Mick McCarthy’s Wolves all show up here with strategies that resemble Arteta’s contemporary version.

Why didn’t they catch on? The simple reason is that these clubs were not particularly effective in converting their corners. Blackburn, Stoke and Wolves ran only slightly above a league average of 0.04 goals per corner in this period.

And this relative inefficiency was probably not random, as a closer analysis of their approach shows crucial differences from the modern approach.

Read more

In the midst of working on last week’s set piece study, I pulled some data together on player shot and goal production from set pieces, and on teams’ primary set piece takers and set piece targets. I found that player set piece goals and assists are best predicted not by open play statistics, but by a combination of that player’s set piece production and by their role on the team’s set plays. These results suggest that set piece production should be understood as a separate aspect of player statistical production, and players should be evaluated and projected in distinct ways on set piece and open play skills. That is an interesting result even if it ended up being extraneous to the argument of last week’s newsletter.

I also did a few media appearances last week including a guest spot on the FML FPL podcast,1 which covers Fantasy Premier League. And I realized that my little set piece study included some actionable findings for fantasy players: measuring set piece targets helps predict goal-scoring. So that will be this week’s mini-study. But, with all respect to Alon’s lovely podcast, that was not the most important newsletter media this week.

I Made “Meat Wall” Happen

Last week I published a piece on the set piece revolution in the English Premier League and we recorded an episode of the Double Pivot Podcast discussing the study as well.

This weekend, Jonathan Northcroft wrote a column for The Times which discusses and builds upon this study, focusing particularly on how corner kick strategies have changed. With more corner kicks being struck on inswinging trajectories toward the goal mouth and more players crowding the goalkeeper to prevent claims, Northcroft suggests that clubs will prioritize bigger and stronger keepers. Perhaps Manchester City’s signing of Gianluigi Donnarumma already reflects a recognition of this tactical change.

Northcroft also picked up on a phrase we used on the podcast, describing how teams form a “meat wall” of bodies to prevent goalkeepers from reaching crosses hit to the six-yard box, which normally they would be able to claim. It was exciting to see this newsletter get picked up in major sports media, and I cannot say enough good things about the image that the Times photo editors selected of the “meat wall” in action for Arsenal against Manchester United.

(Thank you to Bluesky user Nick of the Northbank (@nicktheinventor.bsky.social) for alerting me to the article and sharing this image of the physical newspaper.)

We can only hope that “meat wall” gains momentum from here and becomes a football cliche. It has been my dream since I was a little boy that I might introduce a useful and evocative, but vaguely nauseating, term into common sports parlance.

To step back from my own feelings, the fact that this newsletter did reach a wider audience may point back to one of the study’s conclusions. Anyone who has been watching the Premier League this season has seen the set piece explosion and could easily describe the twin tactics of long throws and crowding the keeper which account for most of its revolutionary effects on the game. This study collected the data to confirm more precisely those observations and intuitions, but those observations are widely shared. Public recognition of the set piece revolution is a necessary first step to any experimentation with rule changes to rebalance the game tactically and discourage these simple but seemingly dominant tactics. When and whether such reactions arise will be one of the major stories to follow in football over the next year.

And as long as set pieces remain of elevated importance to the game, it will be even more important to understand their place in the measurement of player production.

The Mini-Study: Predicting Set Piece Statistics

I have written a set of studies for this newsletter on projecting player production using past data, incorporating age curves into projections, and accounting for positional and tactical context. But in those studies, when I looked at shots, goals, assists and shot assists, I did not separate set piece statistics from open play statistics. This appears to have been a mistake.

The year-to-year correlation of these statistics shows that future set piece goal-scoring has a much closer relationship to past set piece shot production than to overall or open-play shot production. Now, correlations do not necessarily tell the whole story, as they are a rough measure which is also affected by underlying variance as well as real relationships. As I go through the data we will find some uses for open play statistics in projecting set piece numbers as well.

Note that direct free kick shots have been removed from set piece goals for this analysis. DFK shots have an easily knowable value and assignment process, and so they should be handled separately from set piece chances on crossed or launched balls.

Read more

As an analyst, I have watched these questions closely for several reasons. I have always had the intuition that the answer to the first two questions was “yes”, and that a skilled team of coaches and analysts could impose their ideas on the sport via set pieces to an extent that is impossible to achieve in dynamic, open-play football. Many analytics consultancies, most notably StatsBomb, have sought to develop an edge by working on set piece design.

The final reason that I have been watching these developments is that I came to soccer analytics from baseball analytics, and baseball was my first love as a sports fan. I saw the so-called “three true outcomes” take over the sport as teams, informed by analytics, realized it was a better strategy for batters to play for walks and home runs and for pitchers to play for strikeouts. The rate of balls hit into play and the amount of action on the field decreased for the simple reason that the best way to play the sport, based on the rules on the books, was not to try to put the ball in play as much. In soccer, a revolution in set pieces seems like the most obvious way that specialized analysis could change the game.

Of course, it is not necessarily bad for the game to change. A shift in strategy might lead to more goals and more excitement. Perhaps set pieces could become situations which reward clever play design that fans come to appreciate. The question, then, is what are the peculiar characteristics of the set piece moment? Which situations are being exploited to new ends, and in what ways? That will provide a more objective basis for evaluating what kind of change the game is undergoing and how the leagues should respond to it, if at all.

Set Piece and Open Play Goals in the Premier League

I am not the first person to notice a major shift in the composition of goal-scoring in the Premier League this season. Ryan O’Hanlon wrote about set pieces for ESPN already in October. An excellent recent article by Hamzah Khalique-Loonat at The Times extended this analysis and emphasized the same key point. Set piece goal-scoring has increased by a significant margin, and yet there have been fewer goals scored in the Premier League than in recent seasons, not more. Open-play goal-scoring has decreased by even more than set piece goal-scoring has increased.1

It is certainly intuitive that as football teams determine they can score goals from set pieces using new, optimized techniques, they will also take fewer risks in attacking possession. But this graph shows another, or perhaps a related explanation. The rate of open-play scoring in the 2020s has been significantly elevated compared to the previous decade, and so part of this effect appears to be the winding down of a short era of greater open-play excitement.

This data suggests that the drop in goals was caused by confluence of two factors: a broader defensive reaction against the high-scoring game of the early 2020s, in combination with a new emphasis on set pieces which enables less risky open-play attacking.

One rough estimate of attacking play is penalty area touches, and this shows a significant drop in penalty area touches for players at attacking positions from a peak in the early 2020s. It also shows that set piece penalty area touches make up a greater and greater percentage of total touches.

As Khalique-Loonat noted in another piece for The Times, touches, shots and goals by strikers are all down this season. This analysis shows that other attackers, although to a lesser degree, have also seen fewer open-play penalty area touches compared to their recent averages.

It seems likely that improvements in defensive tactics and conditioning have also made attacking more difficult, and the shift to set pieces can be understood both as a reaction to tactical trends and a cause of these trends.

But up until this point I have been talking about “set pieces” as an undifferentiated group. An analysis of the specific types of set pieces that are being exploited and the different means of this exploitation can bring to light more precisely what is happening in the Premier League this season.

Crosses and Throw-ins: A Working Theory

The “set piece revolution” may be a misnomer.2 The increase in goal-scoring has not come from every set piece equally, and in particular it has arisen from two unusual types of play.

Broadly, there are three types of dead balls which can be used to create set pieces.3 There are throw-ins, which can be launched into a crowded penalty area from the sideline, and corners likewise crossed in from the corner flag. And there are free kicks, a much more diverse set of opportunities spread all over the pitch which can be used for direct shot attempts, crosses, launches, or a variety of trick plays. The effects of the set piece revolution are seen entirely in the first two categories, while scoring from free kicks has actually decreased.

The long throw-in cross has been a marginal-at-best part of most Premier League teams’ arsenal for the last decade, and it has suddenly risen in importance to be nearly equal to free kicks. Corner kicks have seen more variance over time, but this season certainly marks a major new high.

It may be that these changes are unrelated, but this result suggests at minimum a hypothesis about what part of the rule book the set piece revolution exploits.

What do corners and throw-ins have in common that free kicks do not?

There is no offside rule on corners and throw-ins when they are taken.

In these two situations, unusually within the game of football, the defending team cannot determine by its shape the area within which the attacking team must play. The attacking team can then pack the penalty area with bodies in whatever formation they choose. It appears that this particular gap in the rules has been the primary opportunity identified and exploited in the Premier League this season.

If Anything is a ‘Revolution’, it is the Long Throw

There have been over 0.12 goals per match this season from throw-in set pieces. This means that in a typical weekend of watching the Premier League, you would expect to see a goal scored from a long throw. This is an increase of more than double from the 2024–25 season, which saw about 0.055 goals from throw-in set pieces, and an increase nearly four times over from the longer-run baseline of about 0.03 goals from long throws.

The basic explanation for this is simple. Teams play far more throw-ins directly into the penalty area than ever before.

The percentage of throws taken from attacking areas directly into the penalty area has nearly tripled compared to the 2021–25 average. Somewhat fewer throws are taken backward, while only half as many are taken to the attacking wing zone.

This radical shift in throw-in strategy can also be seen from a graph of frequency over time.

While small changes were already visible before this season, the change has been truly radical. Within one season, crosses surpassed both types of short throws to become the most common play used on a throw-in from attacking areas.

This mass change in the approach to throw-ins has required a new focus on talent, as teams have identified which of their players has the strongest overhead throw. And it has required specialized work to draw up ways to turn the floated ball in the middle of the penalty area into the best scoring chance possible. But to a great degree, this is simply an opportunity that has been hiding in plain sight for years.

When crosses accounted for less than 20 percent of attacking throw-ins, they led to shots and goals at much higher rates than short throws. On a tactical level, it seems obvious that teams should investigate whether they can increase their usage of long throws to realize these better results at scale. And that is what has happened.

The increased tactical focus on throw-in set pieces appears to have allowed Premier League teams not only to maintain existing levels of efficiency on these plays but even to increase their production. Long throws over the past four seasons have led to shots about 21 percent of the time and goals two percent of the time. This season, even as teams have taken more and more throws long, success rates have increased, with almost 24 percent of long throws leading to shots, and 2.4 percent to goals.

Further, it is possible that the increased use of long throws has had other, positive knock-on effects. The rate of shots and goals from throw-ins taken short has increased by about the same amount as on long throws. When opposition teams have to defend the penalty area against a potential long throw, the attacking team also has the option to run other plays or simply take a short pass in for a cross.

These success rates suggest that, so long as football’s rules remain the same, teams will have little excuse for not making throw-in set pieces a core part of their attacking game. The gains cannot be denied.

Share Expecting Goals

Free Kicks: The Opportunities Left Behind

We know that goals from free kick set pieces have decreased this season in the Premier League. Expected goals tell a similar story, with expected goals per free kick running at about 0.12 per match compared to a past average closer to 0.13.4

There has been a small increase in the use of free kicks as set pieces, especially from midfield, but the amount of change is much smaller than the numbers with throw-ins.

The clearest change here is a meaningful, but not enormous increase in the rate of set pieces played from central areas in the attacking half but not the attacking third. Teams are launching these balls forward more often, but not by more than five to ten percentage points. The large majority of free kicks from these deep areas are still taken short.

There are some peculiar drops in free kick set piece frequency in a few of the areas near the goal. I do not have good explanations for these zones specifically. My primary hypothesis is that teams are playing somewhat more complex set pieces occasionally, and the rules I have set up do not capture these well. One notable point about free kicks in the attacking third, however, is that teams already use these for set pieces the vast majority of the time anyway. If there were to be a large increase in set pieces, it would have to come in what John Muller has called the “Dyche Zone” of launched free kicks from deep areas.

To try to capture more precisely if free kick strategies have changed, I created a model of the likelihood of a free kick being taken as a shot, a cross or a launch. This season there have been somewhat more launches than expected, but only by a small amount.

These results suggest that there is not much room left for teams to play more crosses or shots. But the launched “Dyche Free Kick” still accounts for less than one-third of free kicks from the deepest areas of the attacking half, and only about ten percent in the most advanced areas of the defensive half. John Muller has shown the potential of these free kicks, and even in the midst of the set piece revolution there is little evidence they are being exploited to the fullest. Possibly, when teams feel they have finally maximized crowding the keeper on throw-ins and corners, there will still be more new set piece chances to create.

Corner Kicks: A Dominant Strategy

One might have thought there would not be more room to maximize corner kick goal scoring. Almost all corners are already used as set pieces and kicked long. But it turned out that there are better and worse options for corner kick placement, and significant value to be gleaned from these adaptations.

In a study for Sky Sports, Adam Smith, Nick Wright and David Forster showed that one major component of the increase in set piece goal-scoring has been an aggressive focus on in-swinging corners taken to the six-yard box. The area directly in front of the goal-mouth has been the main target, now accounting for nearly three in ten corner kicks. And that is probably only some of the corners aimed at the goal mouth, as even the best corner kick takers do not hit their spots every time.

The graph of the increase in corners to the six-yard box resembles the long-throw graph for its dramatic shift, although this one does take place over two seasons rather than one.

These crosses obviously target locations where the goalkeeper should be able to get to the ball. But despite this, the rate at which goalkeepers are claiming crosses has gone down in the Premier League. The current trend of surrounding the opposition goalkeeper with bodies, so they cannot patrol the area closest to goal and prevent shots from close range, is clearly working.

I built a model to estimate the likelihood of a cross being claimed, using a similar method to the free kick model, to account for the significant increase in crosses played into the vicinity of the goalkeeper. The model suggests that this season in the Premier League, keepers have been able to claim only about two-thirds of the crosses they normally would under past conditions. Strikingly, in 2024–25, the first season where crosses to the six-yard box surpassed all other crosses as the most common corner kick routine, the rate of claims also spiked and the model more or less predicted the number of times a keeper would rise to catch the ball.

It is true that keeper punches on corner kicks have increased this season, and so keepers are at least getting to somewhat more crosses than these numbers suggest. But punches are always a secondary option for a goalkeeper, which they take when they are unable to safely claim. Again, this suggests the crowding the keeper strategy is working.

This strategy may help explain one of the primary oddities in the corner kick data. Unlike the situation with throw-in crosses, the increase in goal-scoring from crosses to the six-yard box has not been paired with an increase in shots from these chances.

Once again, I included the expected goals numbers in the hopes of explaining why so many more goals have been scored from a very small increase in shot attempts. Here, xG can offer only a partial explanation. Clearly the chances being created in the six-yard box are better than in past seasons, according to the xG model, but only to a degree that explains somewhat less than half the increase in goals scored from these shots.

Two possible explanations present themselves, both of which likely have some merit. The first is that part of the corner kick goal explosion this season has been a fluke of finishing, and the scoring rate from corners should come down to some degree. The second is that the strategy of crowding the keeper is throwing off the xG model, which was not trained on data from this tactical context.

There are two other points to draw from this graph. First is the own goal number. Own goals make up a small, but still significantly larger percentage of goals from corner kicks taken to the six-yard box than corner kicks taken to other locations. Playing a ball more or less at the goal increases the chances it will take a deflection off a defender on the way in.

Second, we see a large increase here in goals scored from short corners. The xG data suggests this is probably more of a fluke than anything else. But it seems possible again that changing the direction of attack unexpectedly, with the keeper surrounded, might also lead to benefits that the xG model is not able to pick up.

One reason I am emphasizing the importance of the tactical change of goalkeeper interference is that past data suggests aiming crosses to the six-yard box was not an obvious winner. This is different from the evidence on throw-ins where playing them long has always been more successful than taking them short. With corners, the inswinger hit to the goal-mouth only becomes a dominant strategy in combination with other changes which magnify the advantages of such crosses.

Conclusion: Looking to the Future, and Germany

Anyone who has been watching the Premier League this season could have told you that set pieces have taken over ever more of the game. This newsletter, to a great degree, serves only to confirm what your lyin’ eyes have already told you. But what is going to come next?

In his piece for The Times, Hamzah Khalique-Loonat reported that he had spoken to an IFAB official about set pieces in September and they had no concerns at the time. While some changes in the rules may be implemented, in particular a shot clock for bringing the ball back into play, I am skeptical these will make any major dent in the effectiveness of these strategies. A long throw specialist sprinting rather than jogging over to take a throw-in will save some time, but it is unlikely to make long throws meaningfully less of a dominant strategy.

What we will have to see in the Premier League is whether any tactical counterstrokes exist. So far none has arisen. It is possible that the set piece trend is like innumerable previous tactical trends, an innovation just waiting for a counter-innovation to come along and re-balance the game. But it is also possible, and it tends to be my hypothesis, that set pieces are a real gap in the football rulebook, situations where a dominant strategy exists and clubs will have few choices but to follow the trend, at least until more significant reforms to the rulebook are implemented.

If that is true, the question may not be what responses to set pieces there are, but rather whether set pieces have been fully exploited yet. In particular I wonder if launched free kicks will become a larger part of the game as teams look for new edges.

But even more than the Premier League, I will be watching the top leagues around continental Europe. If these strategies work, they will surely be picked up elsewhere.

In particular, the Bundesliga shows signs of its own tactical shift to set pieces.

While the increase in goals from throw-ins this season is small, that is because the long throw had already peaked last season. The rates of goal-scoring from long throws in the Bundesliga in 2024–25 and 2025–26 are the two highest of any league since 2010–11 other than the Premier League this season.

On corner kicks, the data points in the same direction.

Goals from corner kicks are up in the Bundesliga, and crosses hit to the goal-mouth have increased significantly. Nearly 40 percent of corners are being placed in the six-yard box.

This appears to also correlate with a decline in keeper claims, although it is worth noting that Bundesliga goalkeepers are far more reticent to claim crosses at baseline.

Watching matches in the Bundesliga, I have noticed several teams making use of recognizable Premier League strategies around the goal during corner kicks. I would love to hear from more dedicated German football fans if this trend is something they have noticed as well.

One of the theories of the Premier League’s set piece revolution has been that Premier League refereeing standards are a core factor. The idea that goalkeeper “should be stronger” when he is blocked by a wall of bodies from reaching the ball is certainly common in English commentary. It may be that refereeing norms which are kinder to keepers elsewhere will limit the effect these tactical trends. It is also possible that refereeing standards simply have not been stress tested yet, and there is more room to make use of physicality, or just masses of human bodies, in the continental leagues as well.

What the Bundesliga data suggests is that it is already happening, and it is already succeeding, if to a lesser degree.

I am by no means certain that the set piece revolution has only just begun. Any number of dynamic responses, from defensive adjustments to national norms of play and refereeing, could prevent the continuation of these trends in England or their extension to other football leagues. At the same time, corner kick and long-throw tactics have been adopted and demonstrated their effectiveness in the Premier League with remarkable speed. At this rate of spread, top football leagues across the world could see rapid shifts in goal-scoring and tactical trends within the year. I am particularly interested in whether this summer’s World Cup will crystallize these trends for an even larger audience.

There is every chance that the set piece revolution has only just begun.

Appendix on Method: Defining “Set Pieces”

Identifying which situations can be identified as set pieces begins as the most basic football question and eventually develops into an irreducibly uncertain quandary that can only be resolved with some best-guess heuristics.

Obviously when a player plays a dead ball cross into the center of the penalty area and their teammate heads it home, that’s a set piece goal. If a player launches a long throw into the penalty area and it isn’t cleared and in the chaos a defender kicks it into the net, that’s a set piece goal. But there are two important edge cases. The first is, when does a dead ball play constitute a set piece? How far does the ball need to be kicked or thrown to start a set piece action? And second, when does a set piece end? If a ball is cleared out of the penalty area but recovered by the attacking team and they cross it in again during the second phase, is that still the same set piece?

For this analysis, I made the following choices.

1. All corner kicks constitute set piece opportunities even if they are taken short.

2. Throw-ins are counted as set pieces when they are thrown into the penalty area, or when the data is tagged as a “throw-in set piece” by the provider, indicating that the action involved a designed play aimed at creating a good shooting opportunity within the next couple touches.

3. Free kicks are counted as set pieces when they are taken directly as a shot, played as a cross, or launched at least 20 yards and into the central area of the pitch, either into the penalty area or a region about eight yards extended beyond the penalty area. Likewise, when an action is labeled as a “set piece” for some other characteristic not immediately visible in the on-ball data, that will also count the action as a set piece.

Further, I chose to count second phase opportunities within the same set piece action. A “second phase” ends when the possession itself ends, either with a turnover or a dead ball, or when the ball is recycled into the defensive half of the pitch, or 20 seconds elapse from the initial playing of the dead ball. These definitions mean that my numbers may differ slightly at the margins from other studies, but they clearly accord with the findings of others in broader terms.

All data from Opta unless otherwise noted.

You can read the initial newsletter introducing the model for a full description of the method, but broadly it is based on a combination of expected goals created and conceded, actual goals created and conceded, and estimated team value from TransferMarkt, with adjustments for schedule difficulty, home field advantage, and time played with a man advantage or disadvantage.

This post will include the updated Expecting Goals Power Rankings for each of the big five European leagues: the Premier League, the Bundesliga, Serie A, La Liga and Ligue 1. Because the model is based on league football only, the ratings are only scaled relative to the league, they do not project how well two teams from different leagues would fare against one another.

Here’s what the columns in the table mean.

Glossary

Team Performance: This is a measure of how well the team has played in its matches this season, based on statistics that best project future quality, adjusted for red cards and opposition quality. It is expressed in goal difference per match. You can think of this as, how many goals better or worse than the average team has this club played over the season?

Performance +/-: This is the change in the team’s performance rating from the previous published version of the power rankings (in this case December 5).

Attack Performance: This is the attacking component of Team Performance. How many goals better has this club’s attack been, compared to the average team, over the season so far?

Defense Performance: This is the defensive component of Team Performance. How many fewer goals and scoring chances has this club’s defense conceded, compared to the average team, over the season so far? Note that better defensive performances here are negative, in the sense that the team has conceded fewer chances and goals than average.

Schedule Difficulty: This is the opposition quality adjustment, also expressed in goal difference per match. A team with “+0.1” schedule difficulty has played a schedule which is harder than average by a margin of 0.1 goals per match. This measurement takes into account home field advantage and the team ratings estimated quality of opposition. A positive Schedule Difficulty reflects a harder schedule, in the sense that the team’s typical opponent is better than average.

Team Rating: This is the current overall rating for this club, representing how my system will project their future performance. It is based on a weighted average of team performance combining goals and xG, regressed to estimated team value, and for promoted teams past performance is adjusted for league difficulty. Team Rating is scaled to 1.0, with 2.0 being a team that is roughly twice as good as the average team and 0.5 being a team that is roughly half as good as the average team.

Rating +/-: This is the change in the team’s projected rating since the previous published version of the power rankings (in this case December 5).

Attack Rating: In the model itself, it is Attack and Defense Rating which are used to project matches. In the newsletter below I will get into the guts of what makes these up, but the basic components are what I listed under Team Rating. Like Team Rating, Attack Rating is scaled to 1.0, with 2.0 being an attack roughly twice as good as league average, and 0.5 half as good as league average.

Defense Rating: This is the defensive component of team rating. Note that while it is scaled to 1.0, now lower numbers are better. An 0.5 Defense Rating reflects a defense roughly twice as good as league average (that is, conceding half as many goals), and a 2.0 Defense Rating reflects a defense roughly half as good as league average.

Premier League Power Rankings

See the actual Premier League table from fbref.com here.

Commentary

Read more

You can read the linked post for a full description of the model, but broadly it is based on a combination of expected goals created and conceded, actual goals created and conceded, and estimated team value from TransferMarkt, with adjustments for schedule difficulty, home field advantage, and time played with a man advantage or disadvantage.

This post will include the updated Expecting Goals Power Rankings for each of the big five European leagues: the Premier League, the Bundesliga, Serie A, La Liga and Ligue 1. Because the model is based on league football only, the ratings are only scaled relative to the league, they do not project how well two teams from different leagues would fare against one another.

Here’s what the columns in the table mean.

Glossary

Team Performance: This is a measure of how well the team has played in its matches this season, based on statistics that best project future quality, adjusted for red cards and opposition quality. It is expressed in goal difference per match. You can think of this as, how many goals better or worse than the average team has this club played over the season?

Attack Performance: This is the attacking component of Team Performance. How many goals better has this club’s attack been, compared to the average team, over the season so far?

Defense Performance: This is the defensive component of Team Performance. How many fewer goals and scoring chances has this club’s defense conceded, compared to the average team, over the season so far? Note that better defensive performances here are negative, in the sense that the team has conceded fewer chances and goals than average.

Schedule Difficulty: This is the opposition quality adjustment, also expressed in goal difference per match. A team with “+0.1” schedule difficulty has played a schedule which is harder than average by a margin of 0.1 goals per match. This measurement takes into account home field advantage and the team ratings estimated quality of opposition. A positive Schedule Difficulty reflects a harder schedule, in the sense that the team’s typical opponent is better than average.

Team Rating: This is the current overall rating for this club, representing how my system will project their future performance. It is based on a weighted average of team performance combining goals and xG, regressed to estimated team value, and for promoted teams past performance is adjusted for league difficulty. Team Rating is scaled to 1.0, with 2.0 being a team that is roughly twice as good as the average team and 0.5 being a team that is roughly half as good as the average team.

Attack Rating: In the model itself, it is Attack and Defense Rating which are used to project matches. In the newsletter below I will get into the guts of what makes these up, but the basic components are what I listed under Team Rating. Like Team Rating, Attack Rating is scaled to 1.0, with 2.0 being an attack roughly twice as good as league average, and 0.5 half as good as league average.

Defense Rating: This is the defensive component of team rating. Note that while it is scaled to 1.0, now lower numbers are better. An 0.5 Defense Rating reflects a defense roughly twice as good as league average (that is, conceding half as many goals), and a 2.0 Defense Rating reflects a defense roughly half as good as league average.

Premier League Power Rankings

Commentary

Read more

So I did what anyone in this situation would do. I built a statistical model.

In some ways this is just a newsletter about how cool machine learning is. I put together a set of data with election results and census demographics by New York City election districts and I fed them into a “Gaussian Mixture Model” to create clusters of similar election districts, and of the 20 clusters it created, 15 had clear and coherent stories, and shockingly coherent geographies, despite the lack of geographic markers in the data. I merged two of the clusters into one group because they made more sense together than apart, and so that left me with 14.

(For a little more on the method and process, see the Appendix on Method at the conclusion of this newsletter.)

I should say here that writing about political geography entails making generalizations about populations. In many of these areas, there are large groups making up 20 percent or more of the population whom I do not discuss. That does not mean they are unimportant. Further, obviously a majority 70 percent of a particular population, whether black or East Asian, lower income or college educated, is made up of many thousands of people irreducibly different from one another. What this analysis seeks to do is not to deny those obvious facts about people. Rather, I think it is striking that despite the ways people are ever unpredictable and changing, we can still identify population level trends by geography and electoral and demographic factors.

Let’s get into the different groups.

The Mamdani Base

The core of Zohran’s base in the primary and the general lay in the Brooklyn and Queens waterfront neighborhoods of Greenpoint, Williamsburg, Bushwick and Astoria.

Mamdani put up margins of 60 points and more in these districts, amid surging voter turnout. This cluster also has the highest percentage of younger adults of any that the model identified. It is professional-class and highly educated but not exceptionally high-income. These are areas with multiple Democratic Socialist local representatives already, and it is hardly surprising they supported a DSA Mayoral candidate.

Several other areas also gave large support to Mamdani, and while these areas are significantly different, perhaps what ties them together most is a high percentage of young people and professional-class voters.

This captures large sections of the neighborhoods of Clinton Hill, Crown Heights, Bedford-Stuyvesant, Prospect-Lefferts Gardens and Flatbush in Brooklyn, and Central Harlem in Manhattan. These are majority black areas with significant young, college-educated, left-progressive populations. Mamdani did not quite rack up the same majorities here as he did in North Brooklyn and Astoria but it was close.

Although Mamdani got far fewer 1st place votes in the ranked-choice primary here than he did in the previous cluster, nonetheless when the ranked choice voting was completed he had two-thirds of the vote. In these affluent, professional-class, white progressive areas, voters used their full ballots and Brad Lander got his best numbers in the mayoral race. Thanks to the ranked-choice system and Lander’s full-throated cross-endorsement with Mamdani, those votes were eventually assigned to Mamdani.

This map shows the neighborhoods of Park Slope, Carroll Gardens and Boerum Hill in Brooklyn, Long Island City in Queens, and a bit of the East Village, Chelsea and Hell’s Kitchen in Manhattan. The best thing about this map is how it identified Morningside Heights, near Columbia University, as sharing these qualities with Park Slope.

I am not sure if the identification of these particular Manhattan neighborhoods—Morningside Heights aside—as being significantly different from the rest of affluent, primarily-white Manhattan is correct. As we go through more of these and my grasp of the political geography gets progressively weaker outside Brooklyn, I will have more uncertainty.

This cluster, even more so than the previous one, is not quite as geographically coherent, with a few little dots strewn around on perhaps dubious grouping logic. But the core of it remains compelling to me. We have chunks of Williamsburg and Bushwick in North Brooklyn, Red Hook and Sunset Park on the Brooklyn waterfront, a bit of the Lower East Side, Spanish Harlem, West Harlem and Washington Heights in Manhattan. These are lower-income areas, majority Hispanic with significant white and black populations, but again very young and with a significant college-educated minority. Mamdani’s margins here were not as enormous as in the above clusters, but he won in both the primary and the general by over 20 points.

Here we also see the first story of the 2020s shift to the right. Trump gained seven points in these districts between 2020 and 2024, but in the down-ballot Comptroller race in 2025, Mark Levine won those voters back and more.

For a fuller analysis of the recent Republican trend in New York City’s voting and whether 2025 saw it reversed, see the 2025 election analysis newsletter, to which this post is a companion.

This map cuts a swath through a series of central and eastern Queens neighborhoods, including Richmond Hill, Ozone Park, and South Ozone Park. These are mixed-race and heavily South Asian areas where Mamdani ran shockingly strong in both the primary and the general election compared to the recent level of Democratic support. He won 61 percent of the vote in the November election, just about exactly equal to Kamala Harris’ share in 2024. If you wanted to look for Trump-Mamdani voters, this is where you should go. It seems very plausible that a South Asian candidate running for Mayor was able to connect with and win over an unusual number of South Asian voters, even among those who had trended away from Democrats in recent elections.

Here, support for Mamdani coincided with a dramatic shift back toward Democrats. Harris had lost 16 points of support from Biden in 2020, and Levine won it all back even beyond Mamdani’s impressive win.

Swing Areas in the Primary

During the primary, Mamdani ran surprisingly strong in a number of areas where Democrats had mostly struggled recently. Generally he did not end up with majorities in these neighborhoods in the general election because large numbers of conservatives did eventually turn out to vote. But his margins in the primary made up a meaningful part of his victory, and we see a number of hopeful signs for Democrats in the general election results.

This is another somewhat spotty cluster, but I think it still captures something. We have heavily Hispanic, working class neighborhoods of Woodhaven, Elmhurst, and Corona in Queens, as well as some chunks of the Bronx. We can see here that lower-income, more moderate Hispanic voters supported Mamdani in significant numbers in the primary and the general election. And just as with the heavily South Asian areas above, this coincided with a massive reversion to Democratic support in down-ballot elections. Where Harris had fallen to 63 percent support in 2024, Mark Levine won 79 percent just one year later.

Mamdani actually won a small majority in the primary in Flushing, Sunset Park, and Manhattan Chinatown, but a more conservative general electorate made these swing regions in November. Here, strikingly, we see the same pattern of recent Democratic losses being reversed, but now in an area where Andrew Cuomo also received a large share of the vote. Voters in New York City’s Chinatowns were splitting their tickets, not supporting the Democratic nominee for Mayor but backing down-ballot Democrats like Mark Levine by high margins that hadn’t been seen in some years.

Large areas of Southern Brooklyn and Northeastern Queens with a mixed White and Asian population have leaned to the right in recent years. In Bayside and Bensonhurst, voters backed Cuomo in both the primary and general election, by notably larger margins in the general. But at the same time, down-ballot Democrats like Mark Levine and Kayla Santosuosso in Southern Brooklyn managed to win majorities. These areas are more conservative on the whole than Flushing and Sunset Park, with larger populations of conservative white voters, but they swung back to Democrats in 2025 nonetheless.

Swing Areas in the General

The story of the 2025 Mayoral election was that Andrew Cuomo couldn’t hold on to his base. The middle- and working-class moderate Democrats who backed him in the primary ended up giving small majorities to Mamdani in the general.

In the black middle class neighborhoods of Southeastern Queens—Jamaica, Springfield Gardens, Queens Village, Rochdale, Cambria Heights, and Laurelton among others—Cuomo managed 40 point majorities in the primary, albeit with depressed turnout. Flatlands and Canarsie in Brooklyn showed similar numbers. But in the general election, majorities of voters in these 90+ percent Democratic neighborhoods ended up preferring Zohran Mamdani. The election was lost when Cuomo could not hold on to these votes, and it was never clear that he or his campaign understood this was where the election would be decided.

We see an extremely similar story in the lower-income neighborhoods of East Flatbush, Brownsville and East New York in Brooklyn, and in the Northeast Bronx. A nearly 30-point shift toward Mamdani from the primary to the general.

Both this and the preceding cluster demonstrate that there was much less of a shift to the right among black voters in New York City than among Hispanic and Asian voters. What small shifts did occur also appear to have disappeared in 2025.

The black and Hispanic South Bronx, which contains many of the poorest neighborhoods in the city, tells a similar story about the Mayoral election as Mamdani gained about 20 points between the primary and the general.

In 2024, the Bronx was a surprising epicenter of Democratic losses, as even in 2022 Kathy Hochul’s vote share in the Bronx had held up reasonably well. Just a year later, Mark Levine won back the double-digit losses Democrats had suffered in these districts.

Cuomo Conservatives and the Manhattan Mystery

Cuomo’s base in this election was Republicans and Manhattan.

I’m sure there is much diversity in the many neighborhoods of Southern Staten Island, in Brighton Beach, Breezy Point, and other highly Republican areas. But the model saw Harris at 25 percent and lumped them all together. And for the purposes of looking at the 2025 election, this is correct. Republicans were motivated to vote against Zohran Mamdani and they did so by enormous margins.

Those margins and turnout, however, were nothing like what Cuomo racked up among Haredi Jewish voters. Mamdani won under 10 percent of the vote in these areas and turnout even increased from the 2024 general election.

One of the reasons I decided I would publish these clusters is that the model found most of New York City’s Haredi communities despite the fact that the census has no data on religious affiliation. The peculiar voting patterns and demographics of these neighborhoods identified themselves to the clustering model.

So finally we are left with Manhattan. These are wealthy, mostly white areas that have historically voted for Democrats by enormous margins. And ultimately, they did support Mark Levine 78–22. But in an election where Democrats were running well ahead of their baseline everywhere, even in Staten Island, it is notable that Democrats lost ground in Manhattan, as well as in Riverdale in the Bronx.

As I said in my election analysis, I would guess that the dynamics here are peculiar to Manhattan and Riverdale. It will remain to be seen whether Manhattan moves to the right while the rest of the city moves left.

On New York City: The Big Apple: The Greatest City in the World

This model helps to demonstrate the immense complexity of New York City’s politics. There is no single ‘Democratic Base’ or ‘Working Class’ vote in New York City. There are many distinct political ecosystems, and the political climate of the 2020s has impacted them in similar but never identical ways.

While the topline result showed a city swinging back to the left, that shift was built on a complex coalition. Young progressives in Astoria, older Black homeowners in Jamaica, and South Asian families in Richmond Hill ended up voting in similar ways, but they all took different paths to get there, through the 2020s and then during the primary. And that’s just in Queens alone. The underlying dynamics which made New York City a bellwether for the nation in recent years likely stem from this very complexity. Whether we want to understand New York City for its own sake, or for what it can tell us about national political dynamics, we must start with this diversity before identifying the trends that can tie different people and different neighborhoods together.

Appendix on Method

This model was built from publicly available data from the US Census Bureau and the New York City Board of Elections. There is an enormous and probably incomprehensible public spreadsheet with the data I used for this analysis which you are welcome to play around with. The census provided information on race/ethnicity, class and education. The election data focused on dynamics since 2020, looking at general election outcomes in 2020, 2021, 2022 and 2024, as well as primaries in 2021 and 2025. This took into account total vote share, changes in vote share between elections, and shifts in turnout.

I merged the census data with the election data by extrapolating from geographical overlaps between census tracts and election districts. Because New York City updates its voting districts every year, sometimes more than once a year, older election results are also extrapolated according to geography. (This is why, if you look at the spreadsheet, it is assigning fractional votes to Joe Biden and Eric Adams. That’s from the geographic extrapolation.)

The “Gaussian Mixture Model” took this data and reduced it to 20 clusters. Five of the clusters had to be discarded as ultimately incoherent. Two of the clusters made much more sense grouped together than separated.

Broadly, machine learning is only as good as the data you feed into it and your ability to make sense of its outputs. I knew which data I had used in the past to explain political trends in New York City, and I provided those to the model rather than every single line of data I could find. The results of the model were, to my eye, shockingly compelling and coherent. But still five of them did not work at all. So I selected the ones that did appear meaningful, and that’s how this newsletter came to be.

New York City in the 2020s is not an outlier; it is a bellwether. The drift of non-white working-class voters away from Democrats began here in 2021, foreshadowing the national catastrophe of 2024. Now, the 2025 results are telling a different story.

Beneath the unusual structure of the Mayoral race, the data reveals a massive swing back toward Democrats, driven not by off-cycle turnout quirks but by former Trump supporters changing their minds. Despite all the drama of 2025 in New York City, the most important story might be how normal the election ultimately was. It is precisely that normality which makes New York City worth deep consideration for electoral analysis.

How Zohran Won the Primary

Now, to explain how the 2025 general election can offer insights into the national political situation, I need to explain the New York City Mayoral primary, which was anything but a normal American election.

Here’s a full map of city results from the primary.

The first way this election was unusual is that New York City has ranked-choice voting in its primaries, which allowed multiple candidates to cross-endorse and encourage their supporters to rank allied candidates. Thanks to cross-endorsements, Mamdani consolidated the support of eliminated candidates in the ranked-choice rounds to expand his lead. He won the first round vote by a little under 8 points, 43.8–36.1, but then in the final round he defeated Cuomo by about 13 points, 56.4–43.6. (In this newsletter I will generally be using the final-round head-to-head vote when I talk about the primary, and that is what the above map displays.)

Second, the primary election was driven significantly by differential turnout. Consistently, rates of voting ran higher in the areas where Mamdani’s support was concentrated.

The area—in blue on the left and green on the right—running from the Queens waterfront down through Brooklyn and around Prospect Park gave Mamdani his largest margins anywhere in the city. These areas also saw some the highest rates of voter turnout, compared to the 2021 Mayoral primary, in the city.

By contrast, most of Cuomo’s strongest regions, from Southern Queens to Central Brooklyn and the Bronx, had some of the lowest rates of turnout in the city. These are light red on the left and purple on the right.

There were 14 Assembly Districts where turnout in the 2025 primary did not surpass 2021 turnout, and Cuomo won 13 of them.

But the few districts where Cuomo did perform well in the primary and where voter turnout did increase from the 2021 baseline are perhaps just as indicative of the dynamics of this primary. AD45 and AD48 in Southern Brooklyn are home to large Haredi Jewish communities who have recently strongly supported Trump, as well as significant numbers of Republicans. AD62 in Southern Staten Island is the most consistently Republican area of the city, while AD76 in the Upper East Side is a moderate, affluent Democratic area. All four regions are heavily white. In areas where Cuomo’s support was highest among white voters, his campaign turned out supporters at good rates.

But most of Cuomo’s support did not come from affluent moderate or conservative whites. It came from black and Hispanic middle class and working class voters in Brooklyn, Queens, and the Bronx. And those are precisely the areas where voter turnout fell.

(If you are interested in the political geography of New York City, see the accompanying newsletter, “The Political Geography of New York City”.)

The core of Cuomo’s failure in the primary campaign was that his campaign, with its focus on Israel-Palestine issues and on the dangers of socialism, did successfully turn out voters in areas where these topics were pertinent. But for the working and middle class black and Hispanic voters that made up Cuomo’s real base, the campaign had little in the way of message or outreach. Further, with his median-voter focused message on affordability, Mamdani was able to reach moderate voters who may not share his larger ideology. Mamdani cut down the margins of his losses in most of his opponent’s base areas, winning on average about 40 percent of the vote there while racking up 70 percent or better numbers in his best districts. And of course Cuomo failed to drive turnout among most of his base.

Mamdani, by contrast, motivated a historic shift in primary voter turnout.

The 2025 New York City General Election: A Normal Election

While the primary was marked by an unusual voting structure and massive shifts in the voting population, the general election ended up a surprisingly normal affair. Even though Andrew Cuomo ran an independent, third-party campaign and Republican nominee Curtis Sliwa refused to step aside for him, Mamdani’s eventual winning coalition was not all that different from a typical Democrat’s coalition.

Obviously Mamdani, running in a three-way race against a former Democratic governor, ended up with a much lower overall vote share than Kamala Harris (51 percent as opposed to 68 percent). However, the shape of his support is easily recognizable. Mamdani lost Republican areas and won Democratic areas, if with lower margins in some of the Democratic areas. The only real exception is the Upper East Side of Manhattan, a Democratic area where Cuomo nonetheless notched a clear victory.

But for the most part, Mamdani won because Democrats voted for the Democratic nominee in a general election.

A map of the change in vote share from the primary shows that Mamdani gained votes in Democratic areas and lost votes in Republican areas.

It was precisely in heavily Democratic areas of the Bronx, Southern Queens and Central to Eastern Brooklyn where Mamdani made major gains compared to the primary. Republicans turned out in force in Staten Island, Southern Brooklyn and Northern Queens to vote against Mamdani. But at the same time a critical group of Cuomo supporters from the primary, moderate and mostly black and Hispanic Democrats, flipped to back Mamdani and gave him the margin he needed.

Once again, Andrew Cuomo’s campaign, buoyed by support from Donald Trump, successfully won over many of the Republicans they were targeting. But they lost Democratic support for every Republican they won and that’s not a good trade in New York City.

It’s worth returning to the point that Republicans turned out to vote. While Mamdani had a massive turnout advantage in the primary, the November electorate showed no obvious trend either way compared to 2024.

One of my favorite little indicators of how far to the right a New York City electorate tilts is the percentage of the vote that came from Republican Staten Island, from everywhere in the borough excluding the North Shore.

Unlike in 2018 when Democrats massively outvoted Republicans in a blue wave year, here Republicans made up a healthy slice of the electorate.

There are two big takeaways from this point. The first is that the Mamdani campaign did not discover a way to change the electorate with a new message. Instead, it seems that Republicans were unusually motivated to vote against a socialist candidate, leading to an electorate that was more balanced, or perhaps even slightly tilted to the right, compared to expectations in an off-cycle election.

Republican voters have agency, and they expressed their preferences in clear terms. Right-wing media, influential figures, and voters themselves recognized and opposed Mamdani’s ideology in particular.1 Conservatives are nowhere near a majority in New York City, and so they were outvoted. The story here is that a Democratic Socialist running on the Democratic line can absolutely win a high-stakes election in a large Democratic city, but not without facing a substantial right-wing backlash.

At the same time, even if this election result doesn’t validate a new way of winning elections for the left, it does offer a surprisingly representative view of the electorate right now. A “normal” election in an off cycle year doesn’t come along very often, and this is an opportunity for deeper analysis.

A Normal Election is a Representative Election

The problem with drawing any conclusions from the November election is that the Mayoral contest itself can’t offer any particular insights because it was a three-way race. But there were two citywide elections, for Comptroller and Public Advocate, contested between Democrats and Republicans, appearing on the same ballot.

Mark Levine won the two-party vote for Comptroller by a 77–23 margin, and Jumaane Williams won the two-party vote for Public Advocate 75–25. These were the best citywide candidate numbers since 2020. (Chuck Schumer managed 74–26 in his Senate re-election in 2022, that’s the closest one.)

These were down-ballot candidates, where odd things can happen. However, it is worth noting that in the 2021 citywide election, the Public Advocate and Comptroller candidates (Williams and Brad Lander) ran only one or two points ahead of Adams.

Further, there were some tightly-contested City Council races in New York City this cycle which offer a more granular view of voter shifts. In particular, Council District 30 in Central Queens and Council District 47 in Southern Brooklyn had open seat elections, which means they were unaffected by incumbency effects. Both were tightly contested, as Republicans had made major gains in these areas since 2020. The Democratic candidates, Kayla Santosuosso in CD47 and Phil Wong in CD30, won by double digits. Levine and Williams put up similar numbers.

It is important to note that it was Democrats in two-ways contests who ran strongly here, not Zohran Mamdani. Cuomo won both CD30 and CD47, by three and five points respectively. What the down-ballot races show us is that while Mamdani gained most of his votes from Democrats and liberals, there was a crucial slice of voters who chose Cuomo while voting Democratic down-ballot. These swing areas of New York City have swung back, but the candidates who can win here need to win over a significant minority of Cuomo voters.

The down-ballot races all seem to confirm that, beneath the three-way Mayoral election, Democrats had an extremely strong election in New York City. With an electorate probably more conservative than the 2024 electorate, Democrats outperformed Harris’ baseline by something in the range of six to eight points, and by more in these swing council districts.

As I have previously written, New York City was a bellwether for the nation in the Biden years. The collapse of support for Democratic candidates in 2021 and 2022 in New York City ended up being a preview for the nationwide Democratic losses in 2024. The concentration of losses among non-white voters, particularly among Hispanic and Asian voters, as well as white people without college degrees, was a notable feature of weak runs by Eric Adams and Kathy Hochul in 2021 and 2022, and then a nationwide problem for Democrats in 2024.

While Harris trailed Biden’s support level by 6.6 points citywide, she fell about 8-9 points in Council Districts 30 and 47. And CD30 and CD47 are not among the places where Harris lost the most votes from Biden. Nonetheless, they are representative of the shift away from Democrats in 2024.

(CD47 spans most of the Southern Brooklyn waterfront while CD30 is a relatively conservative area of Central Queens.)

So this raises a question. If Democrats gained back ground unusually in two council districts which had shifted right, were Democratic gains in 2025 concentrated in precisely those areas?

A comparison of Levine’s and Williams’ vote shares with Kamala Harris’ from 2024 and Joe Biden’s from 2020 shows that this appears to be the case.

Levine and Williams saw only modest improvements over Harris among the white conservative voters of Southern Staten Island, Northeastern Queens, and Rockaway Beach / Breezy Point. But across more heavily Asian and Hispanic areas of Northern and Central Queens, the Bronx and Southern Brooklyn, their gains almost perfectly mirror Harris’ losses.

This map does not quite look like 2021 to 2024 never happened, but it is close.

The 2025 election in New York City shows a massive swing back toward Democrats in a city which had been the epicenter of Democratic losses over the last several years. This swing back did not happen because of differential turnout, despite 2025 being an off-cycle election, because Zohran Mamdani motivated conservatives to turn out to vote. It looks like somewhere between six to eight percent of voters simply changed their minds about Republicans.

The Exceptions: Upscale Manhattan and the Haredi Vote

We see massive gains for Mark Levine, and some for Jumaane Williams, in areas of Southern Brooklyn, South Williamsburg, Central Queens and Far Rockaway which are home to large Haredi Jewish communities. The vote in Haredi areas tends to move more like a bloc, with influential rabbis able to activate large majorities with their endorsements, and often happy to swing back and forth between Democrats and Republicans in local elections to ensure funding and support. The fact that Levine, and to a lesser degree Williams, were able to sway Haredi votes is a specific local dynamic best treated as an outlier.

To account for this, and to describe the political geography of New York City more clearly, I built a “Gaussian Mixture Model” based on electoral and census data to categorize the political regions of New York City. (See the companion newsletter on the political geography of New York City for a full discussion of these model results and what they can tell us about the city.)

One thing that came out very clearly from this analysis was the distinct political behavior and demographic qualities of Haredi districts. If we exclude the districts in this cluster, Levine and Williams end up almost exactly even, at 75 percent, Harris is unchanged at 69 percent and Biden jumps up slightly to 77 percent.

But there’s another clear outlier on the map. The swath of Manhattan from Battery Park and Tribeca on the lower west of Manhattan, up through midtown and both the Upper West Side and Upper East Side bordering Central Park, shows a notable trend against Democrats. This was also a region where Cuomo ran unusually strong in the general election. My model found a large cluster that maps directly to these districts: heavily Democratic, well-educated, extremely high-income, mostly white areas where Zohran had limited primary support. The cluster model also identified Riverdale in the Bronx as having similar dynamics.

In these areas Levine and Williams lost about 5 points of support compared to Harris, unlike in the rest of the city where they made large gains.

So what is going on here? I am torn between two possible explanations. To what degree did affluent white voters in Manhattan, who opposed a Mamdani mayoralty, vote Republican down-ballot in the hopes of balancing out Mamdani rather than because they have changed their political leanings? And to what degree have a minority of Manhattan Democrats reacted to the politics of 2025 by turning to the Republican Party?

If we exclude the Manhattan cluster from analysis as well, the graph looks pleasingly symmetrical.

My guess is that this is not quite right, and Democrats really have lost some percentage of Manhattan voters in the welter of 2025. It is easy to imagine the story of moderate Democratic voters, many of them Jewish, who were negatively polarized against the pro-Palestinian protest movement and then against a socialist candidate for Mayor. But even if that is what happened, will these voters continue voting Republican after 2025? I would guess there are a fair number who will swing back in elections where state and federal, rather than local offices are up for a vote.

Further, for the question of what New York City says about the country, is this then a story peculiar to New York City, a story of voters activated by the Columbia protests and the backlash to the Mamdani campaign?

If the goal is to interpret what New York City says about the rest of the country, we should not lean too heavily on the peculiar dynamics of Manhattan. I think here we have a situation where dynamics truly peculiar to New York City have been the primary driver of election results, and this is less likely to be representative of the nation as a whole.

And in that case, this data reflects a city where Republicans had been gaining ground steadily for years, in ways that were representative of movements in the nation as a whole, and where in a high-turnout election with a representative electorate, Democrats gained seven to nine points above their previous level. That is a very large shift and it was probably primarily a matter of voters changing their minds about whether voting for Republicans is a good idea.

The National Context

In a week where Democrats saw a 13-point overperformance in a special election in Tennessee with nearly midterm-level turnout, this analysis can be added to the growing evidence of a blue wave collected by G. Elliott Morris at Strength in Numbers. Further, the specific shape of Democratic gains in New York City matches with the national trend.

Nate Cohn at the New York Times identified in the special election results from Virginia and in particular from New Jersey evidence that Democratic gains were likely driven as much by persuasion as turnout. Both exit polling and voter registration data suggest that the electorate was, while Democratic-leaning, not so blue that it could account for most of the Democratic wins on the night.

Further, some of the biggest changes from 2024 occurred in North Jersey counties, many heavily Hispanic, which had flipped to Trump in 2024. An analysis by the polling expert Mary Radcliffe found that Trump’s approval rating has been dropping fastest among precisely the groups Trump won over in 2024: younger voters and Hispanic voters. The polling is not granular enough to identify shifts among Asian voters, particularly Asian voters without a college degree, which are so striking in the New York City data. But it would make sense that yet another group of voters who tried out voting Republican in 2024 have decided perhaps it was not for them.

Election Day 2026 is still far off, and as always there is no more powerful driver of political outcomes than Events. But the signal from the 2025 data is becoming undeniable. New York City offers some of the strongest proof yet of a building blue wave. Precisely because Mamdani’s candidacy activated conservative opposition and generated a highly representative electorate, the down-ballot results from this election offer an unusually clear picture of former Trump voters, numbering perhaps seven to nine percent of the electorate, actively changing their minds.

Despite these seemingly shocking numbers, Liverpool’s end of season slide was not considered anything notable at the time. The win for Chelsea made a major difference in the Blues’ chase for a Champions League spot, but it didn’t matter to Liverpool. The Reds had clinched the title the week before at Tottenham. They had nothing to play for.

Famously, the difference between European sporting competition and the formats in use in North America is that in league play in Europe, there’s almost always something on the line. The teams at the bottom risk relegation to a lower league while at the top qualification for different levels of European competition remains at stake until late into the season. Only rarely does a team have more than a few weeks of matches where the points don’t matter for something.

But perhaps because there are so few of these matches, there is broadly an acceptance of the notion that teams ease up after their fate is sealed for the next season. No one asked what was wrong with Liverpool despite the obvious change in their performance level.

So this raises the most basic sort of sports analytics question. Are the intuitions of soccer fans correct? Do teams actually stop playing their best when they have “nothing to play for”? And how exactly does their play change?

What To Expect: Jonathan Liew’s Study

In a 2014 study in The Telegraph, Jonathan Liew identified 76 matches in the Premier League between 2010⁠–11 and 2013⁠–14 in which at least one of the teams involved had nothing to play for. He found that these teams did indeed play worse, taking fewer points per match. And strikingly not only did their goals conceded increase, but so did their goals scored per match. But the increase in goals conceded outpaced the increase in goals scored, leading to an aggregate decrease in goals difference.

In this study, I wanted to update Liew’s study using a larger set of data that would allow for more fine-grained analysis. It was always striking to me how intuitive Liew’s findings were, that teams with nothing to play for just stopped defending, and maybe took some more risks to attack. These results suggested that soccer players really do prefer not to defend so much, that defending is a thankless task involving hard work, focus, and putting your body on the line which players would love a chance to opt out from. And perhaps it is even the case that without the normal pressures of league football competition, players can express themselves more in attack.

The core questions here, then, are how much worse do teams play when they have “nothing to play for”, are these changes in play different in defense and attack, and do teams which have secured different places in the table, whether that means winning a title or facing relegation, show different effects from one another?

Who Has Nothing To Play For? Building the Data Set

So the first task is identifying the relevant matches. I looked at matches played in the big five European leagues starting in the 2010⁠–11 season.1 In these league seasons, teams might have “something to play for” at many different levels.

• Title. Obviously the team that finishes first wins the title.

• Champions League group stage qualification. In different seasons and leagues, somewhere between the top two and top five sides might gain automatic qualification to the Champions League group stage with their finish.

• Champions League playoff qualification. Teams finishing 3rd or 4th, in some seasons, have won access to the Champions League playoffs rather than directly to the group stage.

• Europa League and Conference League group stage or playoff qualification. This follows the same structure as Champions League qualification for the group stages or playoffs, but takes on many different forms season to season.

• Relegation / Promotion playoff. In the Bundesliga and Ligue 1 in most seasons under consideration, the top-ranked team among the bottom three in the league goes to a relegation / promotion playoff.

• Relegation. The bottom two or three teams (or the bottom four from Ligue 1 in 2022⁠–23) are sent down to a lower division.

A confounding factor is that qualification rules are fluid. Domestic cups in particular offer an alternate pathway to Europa League qualification, and when lower-placed teams win domestic cups, they can take one of the league’s qualification slots. Rather than slog through 15 years of domestic cup results to track the shifting Europa spots, I implemented a strict filter to avoid false positives. Based on the relevant league and UEFA rules, I found the places which corresponded to title-winning, Champions League qualification, and relegation, and identified the moment in the season where clubs mathematically locked in these places. I estimated the Europa qualification places based on the maximum number of places which might be given to Europa qualifiers and then excluded these places from consideration entirely. This ensures that I can be equally definitive that the final group below, teams which have guaranteed survival but cannot qualify for Europe, includes no one who could qualify for the Europa or Conference Leagues via their league position.

This method identified:

• 203 matches played by 63 different teams which had clinched a league title

• 249 matches played by 127 different teams which had secured a Champions League group stage place, as well as 15 matches played by 12 teams that had secured a playoff spot

• 337 matches played by 127 teams that were guaranteed relegation, as well as one team that had clinched a spot in the relegation playoff

• 1,115 matches played by 542 teams that were guaranteed survival in the top division but would not earn a place in Europe via their league position

Their results in these matches were easily distinguishable from their results in previous matches, to highly statistically significant degrees.

Read more

And it means that while I covered the Premier League earlier this week, I still have a lot more Power Rankings to share.

La Liga Team Ratings and Power Rankings

La Liga table at FBRef for comparison

Glossary

If you want to see the method and explanation for these power rankings and how they were developed, that is in my earlier post this week, along with some proposals for future study.

Team Performance: This is a measure of how well the team has played in its matches this season, based on statistics that best project future quality, adjusted for red cards and opposition quality. It is expressed in goal difference per match. You can think of this as, how many goals better or worse than the average team has this club played over the season?

Attack Performance: This is the attacking component of Team Performance. How many goals better has this club’s attack been, compared to the average team, over the season so far?

Defense Performance: This is the defensive component of Team Performance. How many fewer goals and scoring chances has this club’s defense conceded, compared to the average team, over the season so far? Note that better defensive performances here are negative, in the sense that the team has conceded fewer chances and goals than average.

Schedule Difficulty: This is the opposition quality adjustment, also expressed in goal difference per match. A team with “+0.1” schedule difficulty has played a schedule which is harder than average by a margin of 0.1 goals per match. This measurement takes into account home field advantage and the team ratings estimated quality of opposition. A positive Schedule Difficulty reflects a harder schedule, in the sense that the team’s typical opponent is better than average.1

Team Rating: This is the current overall rating for this club, representing how my system will project their future performance. It is based on a weighted average of team performance combining goals and xG, regressed to estimated team value, and for promoted teams past performance is adjusted for league difficulty. Team Rating is scaled to 1.0, with 2.0 being a team that is roughly twice as good as the average team and 0.5 being a team that is roughly half as good as the average team.

Attack Rating: In the model itself, it is Attack and Defense Rating which are used to project matches. In the newsletter below I will get into the guts of what makes these up, but the basic components are what I listed under Team Rating. Like Team Rating, Attack Rating is scaled to 1.0, with 2.0 being an attack roughly twice as good as league average, and 0.5 half as good as league average.

Defense Rating: This is the defensive component of team rating. Note that while it is scaled to 1.0, now lower numbers are better. An 0.5 Defense Rating reflects a defense roughly twice as good as league average (that is, conceding half as many goals), and a 2.0 Defense Rating reflects a defense roughly half as good as league average.

La Liga Commentary

• We covered La Liga in some detail on the most recent subscriber episode of the Double Pivot podcast, using these team ratings and our observations of the matches to break down the table.

• Probably the most immediately striking there here is that while Real Madrid rate as the best team in Spain, Barcelona do not join them in that top tier so far. Barca have struggled with injuries, but also the defense has been conceding far too many good chances. It is very, very hard to win a title when you are conceding the sixth-most expected goals in the league. My system here rates their defense a little better than raw xG would suggest, both because there are two penalties in the xG total and because Barca have allowed only ten non-penalty goals, a more midtable total. So far these performances over nine matches have only dropped their Team Rating a little behind Madrid’s. Clearly there is time for a comeback.

• Villarreal play Espanyol on November 8. I may regret it but I am planning to watch and the season data so far suggests it has a chance to be an entertaining match between two teams that have, at least so far this season, been playing good attacking football.

• But the La Liga midtable really tells the story of Spanish football right now. The dominant tactical trend among sub-elite clubs is Getafe-style negative football, a style built on preventing opposition chances with a defensive approach that doesn’t rely so much on conceding opposition possession but on blowing up midfield with physicality, longballs and aerials, and driving down the percentage of live ball play in the match. It’s not great.

Serie A

• Serie A table at FBRef.

• As it has been for several years running, Serie A remains the most competitive top league in Europe. Inter have consistently had the best underlying numbers, but not by such large margins that the normal variation of soccer2 couldn’t allow Milan or Napoli to sneak in for a title. With Inter sitting third, three points behind Roma and Napoli, such drama may be on offer again in Italy.

• If there was one thing most people agreed on before this Serie A season kicked off, it was that Atalanta losing longtime manager Gian Piero Gasperini to Roma would mean that Atalanta’s magic in developing attacking talent would leave for Roma as well, or perhaps that Gasperini had lost his touch and it wouldn’t work at Roma either. But somehow what has happened is that Atalanta remain one of the best teams in the league with a new breakout striker in Nikola Krstović and Roma have joined the ranks of Serie A’s leaders due to an elite defense, which is carrying a still sputtering attack. Football!

• Promoted Cremonese are in eighth place now but with the second-worst expected goals difference in the league, and even this model which gives significant credit for goals scored has them among the worst sides in the league. It might be they have booked enough points to stay up but it looks fluky.

• Generally the bottom of the table in Italy looks just as competitive as the top so there may be a dramatic relegation battle here as well.

Bundesliga

• Bundesliga table at FBRef.

• Bayern Munich’s 2.80 Team Rating is by a good margin the best of any team in the leagues covered here. At the same time, Bayern’s dominance over Germany is not new. Their Team Rating has run over 3.0 for periods in the last decade, in 2019 under Nico Kovač and earlier in 2017 under Carlo Ancelotti. During Pep Guardiola’s tenure between 2014 and 2015 the Team Rating almost never dropped below 3.0 and nearly touched 4.0. Bayern have unquestionably dominated the league so far but it will take even a higher level than this to match their modern heights.

• It is striking that despite RB Leipzig running more or less even on Team Performance with Dortmund and Leverkusen, the model has not lifted their Team Rating up to a similar level. It is never clear to me what rate of updating is best, especially when a team has historically been competitive but has a terrible off year. Perhaps the club mean could incorporate a larger sample of past data. In any case, Leipzig’s negative expected goals difference last season will clearly take more time to filter out.

• Gladbach have fallen all the way to last place and they rate among the bottom teams in the league across all metrics, but the one reed of hope is they have played the toughest schedule in the Bundesliga. On the other side of the coin are Stuttgart, who have played seven of their eight matches against teams in the bottom half of the table with the only exception being eighth-place FC Köln.

Ligue 1

• Ligue 1 table at FBRef.

• The most striking thing here is not so much that PSG are not running away with the league — they have been rotating the squad in league matches with abandon and are still in first place with a perfect Champions League record as well. Rather, it is the apparent stratification of Ligue 1, a league more commonly marked by parity (PSG excluded) over the last decade. The top clubs are putting some distance between themselves and the rest of the table.

• Marseille have played about one-third of all their league minutes either with a man advantage or disadvantage. They have a 24-11 goal difference overall but a 16-8 goal difference at even strength. This ends up not making too large a difference in Marseille’s overall team rating because of the way the red card adjustment is implemented, with a 40 percent credit or penalty to attacking and defending. Marseille creating about five xG and scoring eight goals while up a man over a few matches rates as such strong play, despite the favorable situation, that they still remain near the top of the ratings.

All underlying data from Opta unless otherwise noted.

On one level, this is a soccer analytics newsletter and obviously there should be Expecting Goals Team Ratings during the season. At the same time, team rating systems are one of the few areas of soccer analytics where people already have a wealth of options to choose from. Even after the tragic and still-mourned death of the FiveThirtyEight soccer model, the Elo-based Opta power rankings and the gambling-odds-based Pitchrank power rankings continue to provide insights into the competitive balance of global soccer. And at the league level, there are excellent models for the English Premier League at Analytic.football, Elevenify and the Cannon Stats newsletter.1

If there were no Expecting Goals team ratings model, people could still get insights from soccer analytics about their favorite teams. And while I have worked hard to optimize this model and find new ways to use the statistical record to evaluate clubs, these are surely marginal gains.

I decided to build this ratings system for two reasons. The first is because it is fun. And the second is that building a ratings system opens up a variety of possible new studies and new ways of approaching studies.

Here’s how this system came to be. I was working on a study on the effects of red cards and teams playing with a man advantage, and as a first step I aggregated statistics in games with red cards to find how many goals teams scored before and after the red cards. Teams playing up a man, specifically 11 against 10, typically score and concede at rates that equate over a full match to about 2.0 non-penalty goals scored and 0.7 non-penalty goals conceded. This suggests the value of a red card is worth about 1.3 non-penalty goals. However, teams that get red carded have typically been playing worse in the match up to that point as well. The average non-penalty scoreline before a sending-off , scaled to a full match, favors the team that would later get a man advantage by about 1.3 to 1.0.2 So perhaps the value of the red card is less than it seemed.

The question this raises is, why do we find that teams which eventually get a red card are already being outplayed before the red card? Is it because they are, on average, worse teams? Is it because teams that are losing tend to take risks that can lead to a sending-off? Or might the sample of games with red cards select teams that simply weren’t playing well that day, including but not limited to committing a red card offense? This question cannot be answered simply on the basis of what happened in those matches. It requires an objective measure of team quality that can be used to project game outcomes before the fact. It requires team ratings.

As I worked on this new model, I found that a ratings system opens up many more avenues for possible study. Some of them were pursued in building the current system, and others will become later newsletters and then additions and revisions to this model.

So, at least before these later revisions, what does the current team ratings model say and how does it work?

Premier League Team Ratings

Glossary

Team Performance: This is a measure of how well the team has played in its matches this season, based on statistics that best project future quality, adjusted for red cards and opposition quality. It is expressed in goal difference per match. You can think of this as, how many goals better or worse than the average team has this club played over the season?

Attack Performance: This is the attacking component of Team Performance. How many goals better has this club’s attack been, compared to the average team, over the season so far?

Defense Performance: This is the defensive component of Team Performance. How many fewer goals and scoring chances has this club’s defense conceded, compared to the average team, over the season so far? Note that better defensive performances here are negative, in the sense that the team has conceded fewer chances and goals than average.

Schedule Difficulty: This is the opposition quality adjustment, also expressed in goal difference per match. A team with “+0.1” schedule difficulty has played a schedule which is harder than average by a margin of 0.1 goals per match. This measurement takes into account home field advantage and the team ratings estimated quality of opposition. A positive Schedule Difficulty reflects a harder schedule, in the sense that the team’s typical opponent is better than average.

Team Rating: This is the current overall rating for this club, representing how my system will project their future performance. It is based on a weighted average of team performance combining goals and xG, regressed to estimated team value, and for promoted teams past performance is adjusted for league difficulty. Team Rating is scaled to 1.0, with 2.0 being a team that is roughly twice as good as the average team and 0.5 being a team that is roughly half as good as the average team.

Attack Rating: In the model itself, it is Attack and Defense Rating which are used to project matches. In the newsletter below I will get into the guts of what makes these up, but the basic components are what I listed under Team Rating. Like Team Rating, Attack Rating is scaled to 1.0, with 2.0 being an attack roughly twice as good as league average, and 0.5 half as good as league average.

Defense Rating: This is the defensive component of team rating. Note that while it is scaled to 1.0, now lower numbers are better. An 0.5 Defense Rating reflects a defense roughly twice as good as league average (that is, conceding half as many goals), and a 2.0 Defense Rating reflects a defense roughly half as good as league average.

Commentary

At this point in the Premier League season, the actual table bears shockingly little resemblance to the Expecting Goals Ratings table. While Arsenal have been as dominant statistically as their points total suggests, the second-place team in these ratings stands in fifth place, and the next two in twelfth and ninth place respectively.

Some of these divergences can be explained simply by going to FBRef and sorting by expected goals. Bournemouth, Tottenham and Sunderland all have negative expected goals difference despite their table position, and any model which is heavily based on xG will downrate them just as this one has.

But there are several other notable results here which are peculiar to this ratings system and may help explain exactly how it works before diving into the full methodology.

• Chelsea are not only ninth in the table but also tenth in expected goals difference. Their high rating here is a function of red card adjustment. Chelsea’s goal difference at even strength is plus-10 and their expected goals difference about plus-3. The Blues have struggled badly playing with 10 men for extended periods against Manchester United and Brighton, and this system accounts for that by scaling 10v11 and 11v10 performance by a factor of 1.4. That is, teams which are down a man are “credited” with about 40 percent better attacking and defensive performance than they actually managed in those uneven periods. With this adjustment, Chelsea move up the table.

• Fulham sit in 17th place in the table with the 16th-best xGD. However, my ratings estimate that Fulham have played the toughest schedule in the Premier League, rated at about 0.18 goals per match more difficult than average. While Fulham have not faced every one of the best teams in the league, they have yet to face any of their direct competitors at the bottom of the table. With matches against Wolves, Everton, Burnley, Forest and West Ham to come before year-end, Fulham will have every chance to move up the table and out of the relegation fight.

• There is not much in Aston Villa’s numbers here that would be surprising if you look at the xG table, but nonetheless it is worth emphasizing how strange their season has been. Aston Villa finished sixth in the PL in 2024–25, qualifying for the Europa League, on the back of a hot stretch run that also helped them establish one of the top six team ratings in my system at the beginning of this season. Since then, Villa’s attack has utterly disappeared. This season they are 19th in xG, 16th in shots, 19th in xG per shot, and tied for 16th in penalty area touches. While two recent victories on the back of three converted shots from range have improved their table position, Villa will need to start creating scoring chances at rates more similar to last season if they hope to compete for anything more than survival.

• The relegation chase. Sunderland have obviously shot to the top of the table, but by these ratings Leeds’ performances have been even better. Partly this is because Sunderland have played the easiest schedule in the league, rated at about 0.13 goals per match easier than average. But both teams are playing better than any promoted side in years. Leeds also have a roughly average team rating, in line with their Team Performance. This is because Leeds had dominant underlying statistics in the Championship, as well as a number of Premier League quality players, and so already coming into the season they were projected as a roughly average team. Sunderland, who were fortunate to be promoted, had an extremely poor Team Rating to start the season. Even nine matches this impressive have only been enough to pull their rating up to the level of some of the worst teams in the league.

• The Liverpool Team Rating. It should not be too surprising that Liverpool’s Team Performance has them in sixth place after a four-match losing streak in the league. But how has the Team Rating of one of the projected title favorites already fallen below Newcastle’s?

Read more

Every individual statistic in soccer carries within it the dynamism of the game. The work of soccer analytics is finding ways to pull apart and measure the various, overlapping relationships that are encoded in every measurement of team and player actions on the pitch.

Over the next several newsletters,1 I will examine new ways to account for these contexts and possibly offer revisions to previous results. As a first step, I have taken some of the results of the previous study and re-built models around them to look for places where contexts may be showing through. This newsletter will dig into the player-position data from the last study to offer a few new ways to identify the dynamic, tactical effects, and then examine the influence of larger competitive environments on player longevity.

The Positional Aging Spectrum

Players typically get on the ball further and further away from goal as they age, but this finding about touch profiles can be deepened with positional data. I took the playing time by position records of players from 2010-2025 in the big five leagues and built a generalized additive model to predict playing time by position in the next season with age as a factor. This creates an estimate at each age of how much a player retains their playing time at a particular position or is likely to move to a different position.

As in the previous studies, I have grouped positions together in large buckets: center back, fullback, central midfielder, wide midfielder / winger, central attacking midfielder, and striker. To begin, this is the model output for the positional aging paths of strikers.

This shows that players who get minutes at striker at a young age are often converted to wide attackers, but the tendency to move to the wing dissipates by the early 20s. There is over time a small increase in movement from striker to the 10, but ultimately strikers do not move off the position commonly after the age of 26 or so. This model suggests that after peak age, strikers tend to stay at striker with a retention rate around or above 90 percent.

If you add up the model outputs for all the different positions at age 26, you get 0.04, or 4 percent to WM, about 2 percent to AM, and negligible numbers elsewhere. That means that maybe 7 percent of minutes at striker are not retained but move to another position, and a player who plays striker at age 26 should be expected to play about 93 percent of their minutes at striker in the next season.

Although it is the case that players tend to move further away from goal as they age, this effect does not seem to be reflected in the data on positional changes among strikers. And it is also hard to find among wingers and wide midfielders.

Read more

The project I began in a previous newsletter, a “Marcel the Monkey” system for player projections in soccer, offers one way to get more clarity on what the age curve means. The most basic statistical projection system would take a regressed, weighted average of past performance, then adjust it for context effects and for aging, and provide a simple baseline expectation of player production in the next season. The previous study looked at different methods of regression and weighting for different statistics. I can use those results to aggregate players’ expected levels of performance and compare it to real performance at different ages, in order to gain a more fine-grained understanding of aging in soccer and to begin identifying the aging paths for different skills rather than for a player as a whole.

Age Curves By Position

Much existing work on aging has already approached this topic by narrowing in on age curves by position. I wrote about this a decade ago,1 and a recent study by Tom Worville at The Athletic broadly confirmed these findings. Both studies are based on the idea that total minutes played offers a useful proxy for the age curve. Assuming that clubs and coaches make decisions at least somewhat rationally, choosing the best players they have available to play, the players getting first-team minutes in top leagues should offer a reasonable sample of the best players available.

So who plays those minutes? While there are differences by position, a clear pattern emerges of minutes played by age peaking between 23 and 28 years old.2

This graph shows the results of a Generalized Additive Model3 applied to total minutes played by age and position in the big five European leagues since the 2010-11 season.

What stands out, beyond the fact that decline clearly begins at most positions before players turn 30, are the distinct curves for wingers and center-backs. Wingers peak a year or two before most players. Teenagers are more common in wide attacking positions than elsewhere, while players over 28 are rare in those roles. The peak is more concentrated, higher than all other positions, and earlier. Center back, by contrast, has the longest peak running from about 23-30 and declining much slower.4

It’s easy to offer hypotheses to explain this based on the structure of football. Wide attackers have to do the most running, and in particular the most sprinting and quick accelerating of all players on the field, and their value is the most closely tied to those fast-twitch athletic skills. Center backs, by contrast, have to cover less ground and depend less on speed and acceleration than players at other positions.

But then why do all the other positions, which also demand different and varied skills, mostly just bunch up in the same curve?

Part of the answer here is that further distinctions can be found at the very extremes of these curves.

Strikers become overrepresented in the player pool in their mid-30s, while fullbacks disappear from the player pool at the same time. Central attacking midfielders make up a significant percentage of minutes played both among very young players and also among older players.

The striker and AM effects point to the importance of disaggregating skills. The strikers who maintain their level into their 30s tend to be penalty box finishers, who can provide value with their understanding of space and their strength. Likewise early career attacking midfielders tend to have much more winger-like skills and while those later in their careers tend to be more exclusively passers. My hypothesis is that the reason the curves for FB, CM, AM and ST overlap each other so much is that there are many different ways to play these positions and so certain skills associated with young players will predominate at one time and other skills associated with older players will predominate at other times. Using position as a proxy for the complexity of player aging does offer some insights but it is not fine-grained enough to get into a number of important issues.

This study offers a deeper view into soccer player aging and the effects on players statistics than can be found just by looking at age and minutes. I have modeled aging tendencies over many of the key, modern on-ball statistical metrics, and this newsletter will present and reflect on those model results.

Aging Curves for Statistics

The modeling process here is ultimately similar to what created the first set of age curves, and so the graphics will look similar as well.

If you prefer to see the results of a study rather than read through its technical method, feel free right now to scroll down to “Shot Statistics” where I will begin to talk through all the results.

The process begins with the 2-1-1-5 regressed average for statistics that I found was the overall best method of predicting future performance. I created a projected per-96 statistical projection by taking a weighted average of total production, in which the most recent season has double weight (2), the two preceding seasons have single weight (1). This weighted average was regressed by adding to it 480 minutes (5 full matches of 96 minutes) of league average production for players at the same position.

By aggregating every player where it’s possible to create a regressed, weighted average of past production and comparing that projecting to the player’s real production in the following season, it is possible to search for patterns by age. This is done again with a Generalized Additive Model based on projected production, real production, and age. Are players more likely to decline or to improve on their projected levels in a particular statistic at age 24, age 25, and so on?

The underlying results of the GAM only show year-to-year change and, for instance, look like this for shots attempted.

These are then extrapolated into a curve to make the results easier to interpret, showing peaks based on when players cease improving and begin declining.

These extrapolated aging curve graphs, then, chart the yearly tendency for players to improve or decline in a certain statistic, linked together to create a graph scaled to average player production across the whole population. Average player production is represented by the dashed line. The standard decision of the model is represented by the shaded area, with larger error bars visible earlier and later in players’ careers where the samples are smaller and noisier.

Most of the results are reasonably intuitive, but there are a number that require deeper consideration and several that we may puzzle over together. I will start with the most basic statistics: shots, goals, and expected goal, and then continue to look at statistics for ball-carrying and receiving, for defensive ball-winning, and finally passing and touch location.

Shot Statistics

Shot attempts and xG track each other pretty clearly without significant apparent difference in xG per shot over a player's career. The peak runs pretty early, between about 22 and 27. There is a slight tendency for xG to peak later than shots, but only by about one year.

The non-penalty goals graph is undoubtedly puzzling. Why do goals not show the same early improvement curve and then decline a little more slowly than shots and xG?

Read more

I’m pretty happy with how the essay came out, and you can read it here.

In this newsletter, I wanted to expand on a few points from our conclusion. The core of the piece is an overview of the many ways that Barcelona and Chelsea have attempted to hide their year-over-year losses via financial tricks and accounting maneuvers, and how they have ultimately been successful in evading consequences from their leagues’ financial regulators. The essay ended up growing into a larger disquisition on the state of football as a business and I think summarized the basic ways in which Mike and I have been thinking about football business and discussing it on the Double Pivot podcast.

Chelsea’s and Barcelona’s financial schemes raise an obvious question. Why would anyone want to get involved in a business where you lose so much money every year that you have to deploy financial engineering to hide some of your losses?

Obviously, some people would do it because they want to see a particular football team win matches and they’re happy to see the wealth they’ve accumulated over their lives diminish to make it possible. And some people would do it because they controlled the sovereign wealth fund of an oil state and that state wanted to establish a soft power foothold in Western Europe. And some people would do it because the money they looted from the collapse of the Soviet state made them incredibly rich but the new Russian dictator they helped bring to power could threaten their wealth and safety if they couldn’t move it out of Russia, and buying a football team in London could help establish themselves and their portfolio securely in a society outside that dictator’s reach. There are real exceptions, which I will come back to.

But most of the people and organizations that buy and own football clubs articulate some business reason for purchasing a club. Certainly Clearlake Capital, the new majority owners of Chelsea, are not acting out of generosity toward the people of West London. So what is the business of football at the level of club ownership?

These two paragraphs from our article form the core of my understanding of these dynamics.

The first key point is that the business case for owning a soccer team has never depended on profit-making. Premier League clubs reported 777 million pounds in losses in the 2023 season despite revenues greater than any other soccer league in the world. In North American sports, most owners can expect regular profits because of the cartel system of fixed leagues. Collective bargaining in these systems means that wage costs can be set as a percentage of revenues league wide. Competition between leagues in Europe, and competition to avoid relegation to lower divisions, has led to significantly higher relative overall labor costs. Wages in the Premier League in 2023 ran at about 67 percent of total club income, while in North American sports this number typically runs between 45 to 55 percent, and with the added certainty that comes from collective bargaining agreements.

Where owners have made money in soccer is through asset price appreciation. When Roman Abramovich was forced to sell Chelsea in 2022, he was paid 2.5 billion for the club. Abramovich purchased Chelsea in 2003 for 140 million, which calculates to roughly 16 percent yearly growth, or 13 percent adjusted for inflation. Other sales tell a similar story. In 2005, the Glazer family completed the purchase of Manchester United in a series of deals finally valued at 790 million pounds, and they sold a 25 percent stake in the club to Jim Ratcliffe in 2023 for 1.3 billion. That calculates to eleven percent yearly growth, or eight percent adjusted for inflation. Interested owners might plausibly look at both of these growth rates as undervaluations of their expected returns. Abramovich took those profits in a forced sale in which he did not even legally control the entity at the time it was sold. The Glazers loaded Manchester United with a half-billion in debt in their leveraged buyout takeover and have overseen some of the famous side’s worst-ever performances, yet the club’s valuation still grew more than six times over in 18 years.

Football clubs consistently lose money. Even the most successful league in the world sees year-over-year losses in the aggregate running toward three-quarters of a billion pounds. And they lose money for extremely straightforward reasons. While North American sports leagues have mostly capped the percentage of revenues that go to player salaries through collective bargaining,1 the competitive structure of European football has prevented any such limitations on labor costs.

But despite this, the value of Premier League clubs at sale keeps going up at eye-watering rates.

Football clubs, for the last couple decades, can be understood as financial assets which consistently lose money year over year but appreciate at a rate that typically swamps these losses. This is the structure of a bubble, more or less. But it has been a couple decades, and the bubble has not popped.

One trend which has clearly powered the continued growth of football club valuations has been the spectacular growth of football revenues. Across the affluent world, people have been spending more and more money on sports for a few decades. This trend may finally be levelling off, as can be seen in the English Premier League’s domestic television revenues.

The 2016 domestic TV deal still stands as the best deal the Premier League has negotiated in yearly revenue terms, even without any adjustments for inflation. Grace Robertson wrote an excellent essay on the history and current state of the Premier League’s television revenue, making the case that we have probably hit a market top. Overseas revenues do continue to grow, with the most recent set of deals estimated at a 30 percent increase, but even that is significantly less rapid growth than was common over the last few decades.

There has been a coherent argument that it makes sense to buy into a growth industry like sports in the belief that eventually these revenues can become consistent profits. But despite the argument, that has not happened in European football. Labor costs remain elevated and these seem to be few paths to the consistent profits of North American sports teams which typically pay about 50 percent of revenues in wages.

One thing that Mike pointed out to me as we were writing the article is that the structure of football club ownership, where you pay some year-over-year losses to lock in longer-term asset valuation gains, also looks more appealing in the low or zero-interest rate environment of the 2010s. Once government debt offers substantial, safe yields, a football club starts to look like a less appealing place to park money for a while.

I think this situation probably leaves club ownership at a crossroads. They can either figure out a way to drive down labor costs as a percentage of revenues or they can see their glorious asset appreciation curve level off while still paying down yearly deficits. This is the core of the European Super League drive, and why it was supported by the ownership groups of Liverpool, Manchester United, Arsenal and Tottenham. A fixed league offers the opportunity to collectively bargain with labor rather than constantly competing across national borders for talent. Removing the threat of relegation excises from the game the death spiral logic of running losses in order to avoid running even larger losses the next season in a lower league. The Super League push decisively failed, however, and little in the game over the last few years has suggested any trend toward serious wage limits.

Now, one of the reasons the Super League failed, along with the fan protests from below, was the refusal of Paris Saint-Germain to go along with the plan. There is no need for a club owned by the sovereign wealth fund of Qatar to get into a fight with UEFA just in order to limit its losses. Those losses are a small price to pay for the soft power gains of club ownership. Manchester City, owned by Sheikh al-Mansour of the United Arab Emirates, was then among the first clubs to back out of the Super League.

This can be read in two ways. First, the existence of nation-state ownership may fatally undermine the class solidarity of club owners. As football clubs and leagues have continually sought out new sources of wealth to cover their losses, they found massive gains in bringing in oligarchic and nation-state actors who would happily eat losses while providing huge payouts to exiting ownership. But now these folks have seats at the table and they do not prioritize driving down labor costs as remotely the same kind of existential goal as their fellow owners.

I remain deeply skeptical of the business logic of club ownership, and in particular of the possibility of the kind of systemic reform that would be needed to lock in lower labor costs. The ownership class lacks the kind of solidarity that has benefited owners in North American sports leagues, and this sort of solidarity is difficult to build. Without any path to controlling labor costs, the continued growth of football club valuations still looks like a bubble. This is an economy where people lose money every year and then make it back by selling off the money-losing assets to new buyers. If I were advising a wealthy consortium I would try my best to steer them away from club ownership.

But the same nation-state ownership that seems to be undermining the case for cost controls also might point to the source of new funds to keep the market whirring. What if three exceptions I outlined at the beginning of the newsletter can serve as a source of asset appreciation? That is, perhaps there are still a few more state-backed ownership groups still to buy into the top divisions of European football. Perhaps some new class of international oligarch will emerge in the coming years looking to move wealth across borders. And perhaps the appeal of owning a football team will continue to convince wealthy individuals who should know better to commit large bids at the next auction. There is probably still more wealth out there to buy into club ownership for non-business reasons. There are surely still new sources of irrational money, wealthy individuals reading to buy into bad arguments for these purchases because they’re talking themselves into buying a new toy.

The reason to bet against the bubble popping is a deceptively simple one. People love sports. Football is the most popular sport in the world. Owning one of the most famous football clubs in the world offers a path to recognition at the highest levels of society for people and states who would pay anything for the privilege. And owning a football club offers a possibility of excitement and joy that more rational investments typically cannot. The market for football club ownership then should be understood a kind of perpetual motion machine, driven by the irrational attachments of sports fans from every level of society.

Now, the underlying statistics suggest that their recent struggles are not a fluke. While City are averaging on the season about 0.5 more expected goals per match created than conceded, the balance has recently gone negative. These numbers include both Premier and Champions League matches.

Of course, a few matches with an aggregate negative expected goals difference and a passel of dropped points do not usually constitute a crisis. But Manchester City manager Pep Guardiola has set quite an incredible standard for himself. His teams have won six of the last seven Premier League titles. Before that he won three Bundesliga titles in a row with Bayern Munich and three out of four La Liga titles in his last four years at Barcelona. Further, these teams have consistently had the best underlying numbers in the league to back up their table position.

Pep Guardiola and His Remarkable xG Difference

The one title in the last seven that City did not win, they still had better expected goals difference than Liverpool. His side also failed to take home a Premier League title in Guardiola’s first season, 2016-17, but even from third place they still won the Expected Goals Title with a plus-45 xGD.

As Chris Glover of the PL Fantasy newsletter has documented, this run in the last few months stands out as the worst of Guardiola’s time at City. Because I have data going back a few more seasons, I wanted to expand the sample to see if there were any comparable points.

Read more

Two weeks ago, I shared my Premier League Power Rankings as well as the methodology by which they’re calculated. We have had another round of matches in the Premier League and so the Power Rankings have updated.

But the same method can be easily applied to other leagues around Europe as well, and so we will also be looking at Serie A, La Liga and the Bundesliga by the same method. I want to emphasize that these are still toy power rankings, they are a method I think is tested well enough to be interesting, but not optimized to a point that I would send it out there against anyone’s best model.1

Still, I have tested the model enough over the years to believe that it is providing real insight. And in the Premier League we do have some real continuing movement week to week.

For the power rankings, these are what the six columns mean and how they’re calculated.

axGD/M: My 24-25 estimated expected goals difference for the team at even strength, adjusted for strength of schedule, includes a small adjustment for actual non-penalty goals scored and conceded. The weight on actual goals will increase over the season.

Att Rating: Adjusted attacking performance in 2024-25 above or below league, scaled to league average = 1 (so 143% is 43% more expected goals production than league average)

Def Rating: Adjusted defensive performance in 2024-25 above or below league, scaled to league average = 1 (so 67% means conceding about two-thirds as much expected goals production as an average team). Lower is better here.

SoS xGD: The expected goals difference per match that an average team would be projected to have against their schedule based on my team ratings. So +0.11 means an easier schedule, an average team would be projected to have a +0.11 xGD against this schedule

Proj G/Mt: These are my underlying team attacking ratings for the projections, based on 23-24 and 24-25 data, a projection of goals per match. Because they incorporate 23-24 data (with a weight that decreases over the season), these are much less responsive to performance in individual matches. Note how Chelsea rank well above Manchester City in performance this season, but still have notably worse Projected G/Mt and GA/Mt going forward.

Proj GA/Mt: These are my underlying team defensive ratings for the projections, based on 23-24 and 24-25 data, a projection of goals conceded per match.

For paid subscribers, below is my commentary on the Premier League power rankings table as well as the tables for Serie A, La Liga and the Bundesliga with commentary.

Read more

In this newsletter, I want to expand the scope of this task a little bit.1 I want to do the descriptive work of explaining what happened in the election, but I also want to use this data to intervene in a few discussions that have circulated in left, progressive, and moderate Democratic discourse since November 5th. These are, in a quick sketch, the interventions I hope to make.

• Claims that the 2024 election was driven by “turnout” or by “persuasion” alone are both unconvincing, and rather than arguing for one or the other, we need a richer understanding of what it means when voters stay home or switch allegiances.

• Our best data suggests that the Harris campaign ran stronger in swing states than elsewhere, and so we should look deeper than questions of campaign tactics and policy messaging over the final months of the election for our answers to what happened here.

• A uniform swing against Democrats can be seen at nearly every level and among every group of voters, but there are also particular shifts which were most pronounced among Hispanic and Asian voters. Drilling down to the precinct level in New York City allows us to find trends in Asian voting in particular that are difficult to identify nationally.

• The relatively consistent trends among the (very different!) people lumped together in the census as “Hispanic/Latino” and “Asian” suggests we should not look to more granular causes among particular communities but to questions of what unites larger groups of people to understand what happened.

The Shrunken Electoral College Gap

So at a first glance, what happened is that Donald Trump won the election. He is likely to end up with a roughly 1.5-point popular vote victory. The Republicans will control both houses of Congress with a 53–47 advantage in the Senate and, depending on a few slow-counting races in California, between a 222–213 to a 220–215 advantage in the House.

These results point to the first unexpected note about the 2024 election. The Electoral College bias, which gave Trump his victory in 2016 and allowed him to come shockingly close in 2020 despite losing the popular vote by over four points, disappeared this year. Democrats somewhat outperformed their baseline in the House. Harris lost the tipping point of Pennsylvania by a margin of 1.7 percentage points, nearly identical to the national popular vote difference. What happened in the swing states and in the non-competitive states to bring about this outcome?

The analysis here is necessarily preliminary. The best studies of what happened in an election are not released until many months later, when large surveys can be conducted based on verified lists of voters. The Democratic data firm Catalist will produce a 2024 equivalent to their study of 2022 some time next year. Until we have large-scale survey data, the best material to work off is the election results themselves.

For a variety of reasons, people in the United States do not live in perfectly representative communities. Neighborhoods where particular ethnic groups predominate are common, as is sorting by class and education and even political affinity. This means that we can look at election results by county and precinct to learn about how people voted. If we can identify areas with very high rates of college education, we can make some best-guess estimates of the voting patterns of college educated voters.2

If we break out completed county returns nationwide, there is roughly a six-point shift in margin from Biden to Harris. Biden won the popular vote by about 4.5 percentage points, while Harris lost it by about 1.5. Democratic losses were found everywhere across the country, in cities, towns, suburbs and rural areas, but were most concentrated in large urban cores and their immediate suburbs.

While there were shifts of between three to eight points everywhere, the causes of these shifts vary. In suburbs, smaller cities, and rural areas, Harris lost vote share while the number of votes cast remained mostly steady. In the largest cities, typically stronger areas for Democrats, turnout collapsed at the same time as Republicans made persuasion gains.

It’s worth pausing here on questions of “persuasion” and “turnout” to discuss what these terms mean. In the abstract, a “persuaded” voter is one who was definitely going to vote and your campaign convinced them to vote for you instead of your opponent. Successful voter “turnout” refers to finding a voter who will definitely vote for you, but only if they make it to the polls, and ensuring that they cast their ballot. Even in a wholly-imagined story, it is clear the terms are vague. No campaign can know for sure that one person’s support is guaranteed if they turn out, or that another person is certain to vote and merely their choice at the ballot box is up for grabs.

But we will return to that problem later. When we are working from county and precinct data, we have a basic problem of measurement. All we know are totals. We can look at two indicators: total votes cast and vote margin. But they are not equivalent to “turnout” and “persuasion”. Even if one candidate exactly matches their vote total from the previous election, and the other sees their vote total decrease by 100, we do not know that this was all “turnout”. Possibly both candidates lost 50 voters from the previous election, but 50 voters switched their preference. Or likewise if total number of ballots cast is exactly the same and only the margin shifted, it’s possible one candidate had a bunch of their voters stay home and the other candidate turned out an equal number of new voters.

Still, these numbers are not meaningless. Both of those extreme versions of “persuasion” masquerading as “turnout” or vice versa are possible, but unlikely. Because “turnout” and “persuasion” are abstracted categories from the more complex process of how people decide to vote, all vote outcomes are some combination of the two. If we find consistent trends across multiple similar counties and precincts, we can draw tentative conclusions. We cannot say one situation was all turnout or all persuasion, but we can make reasonable guesses at the mix of persuasion and turnout depending on vote totals and margin across a large set of data.

Perhaps the best example of large-scale trends across different places in which we can see these overlapping effects is that voters in swing states appear to have behaved differently from those in non-competitive states.

There is a notably uniform swing in competitive states against Harris. Turnout was still comparatively worse in urban areas than elsewhere, but it was up slightly rather than down drastically. Harris lost vote margin on average compared to Biden, but only about three points of margin rather than over seven in non-competitive states.

The simplest explanation of what makes swing states different from non-competitive states is that campaigns spent money in them but not in the non-competitive states. In areas where the Harris campaign focused both its spending and its volunteer outreach, Democrats did significantly better than elsewhere. These effects served to boost the number of voters going to the polls and improve Harris’ standing with those voters. Turnout and persuasion seem to have gone hand in hand, and shifts in both were more concentrated in urban areas.

If we want to look for reasons why the Harris campaign lost the election, we should look elsewhere than the messages with which they bombarded voters, via paid media and volunteer canvassing, in swing states in the final months of the election.

Demographic Analysis: Turnout and Persuasion

Of course, we can do more than divide up the United States by population density. Cities, suburbs, towns and rural areas may be inhabited by a lot of different people. The 2020 census offers a number of useful demographic categories, and suggests a few different things were happening in these different areas. And here we begin to see clearer evidence of distinct turnout and persuasion effects.

We see, in particular, different dynamics in areas which are more heavily white or black. There is a small but significant shift toward Trump in both more and less educated white counties, and a relatively small change in total ballots cast. This looks like a small but meaningful persuasion shift. In heavily black counties, it appears to be a turnout drop outpacing margin shift, suggesting here the problem was primarily turnout.

If we look at the swing state / non-competitive state split, we see something similar.

This data highlights the difficulty in 2024 of telling a clean story about turnout and persuasion. The midterm election of 2022 offers a useful counterpoint. In 2022, Democats voted at much lower rates, but the Republican vote share increased by only a small margin, and in fact did not increase in the swing states. This is a persuasion election, where gains by convincing voters to switch sides clearly outweigh the effects of getting your base to the polls.

There are no such easy stories to tell about 2024. In swing state counties with high concentrations of white people without college degrees, where Trump is strongest, there were larger increases in votes cast than anywhere else in the nation. But in swing state counties with high levels of college education among white people, where Harris expected to run even or slightly ahead, turnout increased and Harris’ share decreased. And in swing state counties with high black populations, while there was a small increase in total ballots cast, this increase was smaller than elsewhere and so cost Harris votes at the margin.

Outside the swing states, black turnout appears to have collapsed, leading to significant losses in vote margin for Democrats.

Everywhere we look, we see the hallmarks of both persuasion and turnout gains favoring the Republicans.

Hispanic Vote Share: Even Further Complexity

I could not include the Hispanic vote analysis in the above section because almost all of these heavily Hispanic counties are in non-competitive states. We would expect to see somewhat better Harris numbers in swing states, but it cannot be observed from county-level data.

Further, and even more strikingly, the huge Hispanic county turnout drop is concentrated in New York, New Jersey, and California.

It is oddly consistent that Harris lost about 10–15 points in heavily Hispanic areas, but in some places it coincided with collapsing turnout and not elsewhere. It requires a different kind of analysis to understand what happened in Imperial County in California compared to Cameron County in Texas.

Both are border counties with close connections to Mexico, one at the furthest eastern point of the border on the Gulf of Mexico and the other almost all the way to the Pacific. In both the population is supermajority Hispanic or Latino (89 percent in Cameron County and 85 percent in Imperial County), and they live mostly in small and medium-sized cities. In both counties Trump improved his margin by close to 10 percentage points from 2020. But how he did it was very different.

Cameron County, TX

2020: Biden 64,063 to Trump 49,032
2024: Harris 54,156 to Trump 60,925

Imperial County, CA

2020: Biden 34,678 to Trump 20,847
2024: Harris 23,229 to Trump 22,433

Trump added fewer than 2,000 votes to his total in Imperial County and over 10,000 votes in Cameron County. But he actually improved his margins by slightly less in the latter because of a decrease in votes cast in Imperial County that crushed Harris’ totals.

The paradox here is that two demographically similar groups of voters, in relatively similar areas, split their votes by a similar margin, but the way they split their votes was almost a total mirror image of the other. One offers a paradigm of a persuasion shift and the other an equally paradigmatic turnout effect. And despite that, both counties ultimately shifted toward Trump at similar aggregate rates.

While I cannot speak to the local concerns in either county that could have driven these effects, I suggest that if we stick with thinking of “persuasaion” and “turnout” as wholly different things, we will struggle to understand what happened here.

One important piece of context for these shifts is that they did not begin in 2024. Rather, after some regional swings in Hispanic voting patterns in the Rio Grande Valley and South Florida in 2020, the 2022 election saw large shifts in Hispanic and Asian voting patterns spring up across the country and most notably in New York City.

New York City: We Matter

While I produced this analysis of the 2024 election in New York City for a local Brooklyn political club, and for them and for me the interest in New York City is obvious but parochial, there are good reasons to look to New York City regardless.3

Kathy Hochul in 2022 only won the two-party vote in New York City by a 69–31 margin, after Biden won 76–24 two years earlier. This year, Harris not only couldn’t claw back any of Hochul’s losses but in fact ran slightly behind her in the city, taking only about 68 percent of the vote.

And it was mostly the same votes lost in the same places, despite turnout being much lower in 2022, as is common in midterm elections.

These maps show swing from Biden to Hochul/Harris in darkening shades of red and swing to Hochul/Harris in blue (if you squint).4

The Republicans picked up votes in conservative Staten Island while Democrats held vote share but could not increase it in their strongest areas of Manhattan and Brooklyn. Democrats took massive losses in Southern Brooklyn and most of Queens besides the areas near the East River. The biggest difference between these maps is that Harris lost further vote share in the Bronx.

Overall, the swing went against Democrats at similar rates in similar places.

Further, if you know the geography of New York City, you will know that the areas where both Hochul and Harris lost the greatest vote share — the reddest areas — are some of the most heavily Hispanic and Asian areas of the city. The neighborhoods where Democrats maintained their 2020 vote share are neighborhoods with either large black populations or high rates of college education among the white population.

Breaking up the city by these five categories can explain a huge amount of the vote shift. These five maps highlight precincts in New York City by these demographic characteristics.

The nationwide county-level data could not identify this Asian vote swing because there are few counties in the United States with such predominant Asian populations. But by zooming in to the precinct level in New York City, we can see this vote shift happening and that it had begun by 2022.

Black and college-educated white precincts appear in very pale colors because of the lack of shift there compared to heavily Hispanic and Asian precincts.

With these five demographic categories showing such visibly clear effects on the maps, I tried a multiple regression by demographics against Biden→Harris vote shift, and found a significant effect.

The R2 of 0.43 means that just under half of the variance precinct-to-precinct in vote shift can be explained by just these five demographic categories.

The massive vote shift in Hispanic precincts matches the nationwide county-level studies. It is in turn now matched by even a somewhat-larger effect in heavily Asian precincts.

One thing that is somewhat different between New York City and the nation is the non-college white effect. Across the country we see only small persuasion effects among white voters. But in New York City there is a much larger swing. I have two hypotheses here. First, these are somewhat different population effects. There are more than 1,000 counties with (adult) populations over 70 percent white-non college degree, and only a handful of precincts in New York with such a concentration. We are looking at just 40 percent and above in this map.5 Second, there are very high rates of evangelical religion and southern regional conservatism in non-college white areas nationwide, both of which are rare in New York City.

If my hypotheses are correct, it suggests that the Democrats’ problems with Hispanic and Asian voters run in parallel with problems with (Northern, non-evangelical) white voters without college degrees.

Now, so far this has just been looking at vote swing in New York City. Because of the methods I had to use to estimate voting precincts — New York City, very frustratingly for anyone trying to do election analysis, changes its precinct boundaries every year — it’s not possible to get really granular turnout numbers for regression or mapping. But I can aggregate precincts as I did with counties.

We see total ballots cast dropping everywhere except areas with highly-educated white populations. There is an even larger drop in vote share in more heavily Hispanic and Asian areas. You can see, I think the outline of a story here where Democrats struggled with turnout among almost all demographic groups, lost some vote share to persuasion among almost all demographic groups, and in particular lost tons of votes to persuasion among Hispanic and Asian voters. This possibly also applies to white voters without college degrees outside the South.

So What Have We Learned?

The main intervention I wanted to make here was against easy claims of a turnout election or a persuasion election. It is clear that turnout and persuasion had a complex interplay here, and while we will be able to say more once comprehensive surveys has been done, I think the biggest question this raises is exactly what “turnout” and “persuasion” imply. Why does it seem to be the case, for instance, that there were collapses in turnout among Hispanic voters and also collapses in Democratic support among those Hispanic voters who did get to the polls? Why were these drops arrayed differently across different geographic regions?

There has been a fair amount of polling on the sort of disaffected, weakly-attached, infrequent voters that may well have turned against Democrats in this election. This overview by Nate Cohn at the New York Times found that Democratic-leaning voters who skipped the 2022 or even 2020 election were the crucial voters for Democrats to turn out and persuade in advance of 2024. These voters tended to be younger and non-white compared to the population at large, and they were less likely to have a college education. They were less motivated by the defense of democracy or abortion rights, they were much angrier with the leadership of Joe Biden personally, they got their news from social media and were often mistaken about who had been responsible for overturning Roe v. Wade, and they were broadly much more unhappy about the state of the country. These voters, crucially, were at risk of either voting for Trump or choosing to stay home. This article reads in retrospect like a description of the voters we have been finding either dropping out or flipping to Trump in the demographic analyses.

The sorts of voters who stayed home and the voters who switched allegiances may have more in common than the contrasting categories of “turnout” and “persuasion” might suggest. The voters who were unlikely to make it to the polls for Biden or Harris, and those who were at risk of switching their vote to Trump, were not obviously two different groups of people. The interplay of turnout and persuasion in the data might be less confusing if we are seeing different effects on the same kinds of people.

The notion of finding, underlying all this variance, “the same kind of people,” is also what I am taking away from much of this demographic analysis. In particular, I would highlight the finding in New York City that areas with larger populations of white people without college degrees have shifted against Democrats in ways that seem to run parallel to somewhat larger shifts among Hispanic and Asian voters. While there are certainly specific questions of ethnic identity and experience, and quite possibly crucial questions of language in messaging, the similar trajectories of these voters from 2020, to 2022, and then on to 2024 suggests we should be looking for commonalities more than differences. What would cause Democrats to lose vote share among such disparate groups at similar rates?

Of course, this data also reveals some real demographic divergences. Democrats appear to have lost some ground among white voters with college degrees, but with steady rates of voting and no more than uniform swing margin loss, it does not appear Democrats have any deeper problems here. And with black voters, the lack of extreme swings in voter preference despite large drops in turnout is notable. The traditions that have made black voters, in the words of Chryl Laird and Ismail White, “Steadfast Democrats,” seem to have held even as turnout declined.

And this brings me to my third takeaway, that 2024 did not come out of nowhere. The losses that Democrats took nationwide, with notable concentration among Hispanic and Asian voters, were presaged by New York in 2022. The losses among white voters without a college degree follow from an even longer-running political realignment. This loss was driven by an apparent nationalization of trends that were merely regional in 2022. And it is hard not to see the exception, college-educated white voters, as indicative of the larger trend. Most black, Hispanic, Asian and white voters do not have college degrees. If patterns of Democratic voting among ethnic and racial minority populations are weakening, while only the too-small group of people with college degrees flips the other way, the future of the party is at risk.

If that is true, and if it was a set of longstanding political traditions and institutions which prevented black voters from turning against Democrats at the same rates as most other voters, is this a bulwark the party can rely on for the next few elections? It worries me deeply that Harris’ loss could have been significantly worse if there had been more significant shifts of black voters against the Democrats and not merely a drop in turnout, which in turn was concentrated outside the swing states.

A broad set of national trends are breaking against the Democratic Party right now. These trends appear in national voting patterns and in greater detail in New York City local elections. But we should not view them as the only factor in this election or elections in the future.

The International Context: Feel the Dialectic

I have so far split time between looking at national and local trends in the US election. Taking a wider scope and looking at elections worldwide tells a different story. As John Burn-Murdoch at the Financial Times has pointed out, incumbents have lost every election in 2024 worldwide. Of course, it is unlikely that it is something specific to the calendar year 2024 that is driving these results. Voters across the globe, and particularly in affluent countries with little experience of inflation in recent decades, rebelled against the rises in price level caused by the disruptions of the pandemic and by oil and gas supply shocks following the Russian invasion of Ukraine. If we look at elections in 2023 and 2024, in such affluent countries, we find a similar although not entirely unanimous trend.

Pedro Sanchez of Spain’s PSOE breaks the trend. But the average is terrible. Inflation has uprooted governing coalitions around the world. And this average also cuts against peculiarly ideological takes on the effects of inflation. Centrist coalitions and right-wing parties6 have been swept out of office just as easily as more left-leaning governments.

It seems likely that the 2024 election in the US was fought on a battlefield that favored the Republicans by much more than one would expect any future election to tilt. The trends that seem like they doom the Democrats probably would not have peaked in a Trump victory in 2024 if it weren’t for inflation, and so these trends are probably more likely to regress temporarily in 2026 and 2028 than to continue at the same pace.

There should still be every opportunity for the Democratic Party to recover from this defeat and to address head-on the demographic realignments that put the future of left and center-left politics in this country at risk.

Permit Me One Last Take

I have sought to avoid arguing about policy messaging in this piece. I know everyone wants to know if the Democrats moved too far to the left in their platform and messaging, or perhaps if they moved too for the center instead. In part I have avoided such topics because these questions will only be answered by larger-scale survey methods and are nearly impossible to grasp from precinct data.

But I have a strong hypothesis that focusing on policy messaging narrows the scope of political analysis far too much. In Cohn’s polling, the uncertain voters lacked particularly clear ideological viewpoints and tended both toward a rejection of leading progressive issue positions and a desire for a vague but more radical shift in the direction of the country that would not be answered by moderate, “commonsense” policy alignments. And more than anything else, they really hated the specific people leading the Democratic Party.

This is not to say that Democrats should not pore over the data on voter preference and make sure their messages resonate with majorities of voters. But I worry very much about a party that only seeks to recalibrate its policy messaging.

What I think is most indicative about these voters is not their ideology but their weak social and political attachments. These are voters whose preference and likelihood of voting are equally uncertain because there are few ties of any sort binding them to a political coalition. It seems like a mistake to reach out to disaffected people who do not trust traditional institutions or the fairness of the modern world, with just a better set of messages delivered through traditional channels.

These deeper and richer connections with voters can only be built through institutions, and the current Democratic Party profoundly lacks such embedded institutions. For most of the last decade I have been organizing with New Kings Democrats to build a better Democratic Party locally. In 2024, I volunteered with New York for Harris/Walz to connect Democratic volunteers with needs in swing states and run buses to towns across Pennsylvania. This organization was expanded suddenly by a group of volunteers to meet the volunteer energy that followed on Biden’s stepping down. There were no existing institutions in Brooklyn or New York City to handle this labor, despite all the elected Democrats here. This is an indication of what Daniel Schlozman and Sam Rosenfeld call “Hollow Parties.” Modern party organizations have an incredible capacity for fundraising and ability to collect the votes of millions of Americans in regular elections, but they maintain only a minimal presence in American communities beyond those strengths, and because of that the parties lack the ability to enforce discipline and set goals for their ideological coalitions.

Even if the whole problem were policy messaging, and someone figured out the perfect policy messaging to win elections, how would the Party build consensus for this new direction and then enforce discipline to maintain it?

The answer to the profoundly worrying trends in the 2024 electoral data must be institution-building and not mere revision of policy messaging. This institution-building must lift up new leaders who can differentiate themselves from the Democratic leadership that proved so unpopular in recent years. There is still time for the party to adjust and to build an apparatus that can win elections. But we must confront 2024 as not only a profound loss but an indication of continuing trends against the Democratic Party that must be stopped if the party means to regain majorities and make a better world.

As I’ll explain below, this will be a regular addition to your Expecting Goals content. When I have some stuff to say about soccer that is not a fully worked-out study, I will post it here. As the season goes on, we can track the power rankings in the Premier League, and perhaps other top European leagues, and see how they hold up and what they have taught us.

Power Rankings

These power rankings are based on a combination of expected goals and actual goals, only considering 11v11 minutes in the Premier League, and adjusted for schedule difficulty. They are based on ongoing, not-completed work on team valuation and I present them as a toy, not a fully worked-out model of true team quality.1

There are six columns.

axGD/M: My 24-25 estimated expected goals difference for the team in 11v11 situations, adjusted for strength of schedule, includes a small adjustment for actual non-penalty goals scored and conceded. The weight on actual goals will increase over the season.

Att Rating: Adjusted attacking performance in 2024-25 above or below league, scaled to league average = 1 (so 143% is 43% more expected goals production than league average)

Def Rating: Adjusted defensive performance in 2024-25 above or below league, scaled to league average = 1 (so 67% means conceding about two-thirds as much expected goals production as an average team). Lower is better here.

SoS xGD: The expected goals difference per match that an average team would be projected to have against their schedule based on my team ratings. So +0.11 means an easier schedule, an average team would be projected to have a +0.11 xGD against this schedule

Proj G/Mt: These are my underlying team attacking ratings for the projections, based on 23-24 and 24-25 data, a projection of goals per match. Because they incorporate 23-24 data (with a weight that decreases over the season), these are much less responsive to performance in individual matches. Note how Chelsea rank well above Manchester City in performance this season, but still have notably worse Projected G/Mt and GA/Mt going forward.

Proj GA/Mt: These are my underlying team defensive ratings for the projections, based on 23-24 and 24-25 data, a projection of goals conceded per match.

Commentary

When you look at just 11v11 minutes, Arsenal stand out as the best team in the league. The gap to Liverpool is not so wide that it is impossible Liverpool could overtake them in time, but on the matches we have seen so far, Arsenal have done the better job creating chances for themselves and preventing chances for the opposition. They have, obviously, received several red cards this season and less obviously failed to take points from close matches where last season the bounces tended to go their way. Liverpool by contrast have been much more efficient at translating modestly superior play into points, and this is the source of their large lead in the table.

At some point I hope to build a projection system that I like enough to share here. Based on the toy projections I have at the moment, Liverpool would be solid favorites for the title despite landing second in these power rankings.

For Manchester City, the result against Brighton seems to have finally focused the media on a clear finding in the underlying numbers. Not since Pep Guardiola’s first season in charge has Manchester City not been 1 or 1a in any advanced statistical power ranking of the Premier League. But with Rodri and Kevin De Bruyne out and Julian Alvarez gone and inexplicably not replaced, City have been playing merely solid top-four football. We are all waiting for Pep to conjure up some new tactics and turn it around but he has not done it so far.

Chelsea have clearly been the best team from outside the expected top three, and their performance is even more impressive because they have put up these numbers against one of the toughest schedules in the Premier League so far.

If the Premier League gets only four places in the Champions League, we have increasing reason to think the chase will not be terribly exciting. But a big pile-up of good-not-great teams outside the top four suggests that if this season, teams from the PL can do well enough in Europe to gain a fifth qualifying place, anything is possible.

Manchester United. No one who follows soccer through an analytics lens could possibly have been surprised by their poor start. I think one notable fact here is that their problems appear to be concentrated on the attacking side. The signings of Matthijs De Ligt and Noussair Mazraoui stabilized the back line, but there just are not enough shots in that attack. I am skeptical that a new manager will be able to find a lot of new attacking firepower in a collection of players who rarely attempt three or more shots in a match.

Nottingham Forest look like a classic good version of a Nuno Espírito Santo team, similar to his Wolves sides that rattled off a run of top ten finishes. The attack has been juiced by some hot finishing, but this is a team that succeeds because they are exceptionally well-organized out of possession and can spring a few counter-attacks. It is difficult to collect enough points to finish in the top four or five with such a profile, because highly-defensive teams tend to achieve more draws and fewer wins.

Southampton have been one of the most tactically distinct teams in the data era, not just in the Premier League but across the big five leagues. I do not mean this as praise. But I am considering writing up a profile of this team because it is still underrated how strange it is that the worst team in the league plays such an aggressively possession-based style.

Methodology and To-Dos

Here’s how the method currently works. It strips out all minutes played not at even strength. It takes non-penalty expected goals (based on my xG model) and actual non-penalty goals in those minutes. The weight of xG to NPG begins at 90/10 and progresses to 50/50 over the full season. This is because there is real signal in goal-scoring, but it is swamped by variance in small samples. This aggregate creates an initial score for all teams.

To adjust for schedule, I use past-season data as well as current season data. The weight of current season to past season data increases every match until after the 25th match the past season data is no longer part of the team quality estimate.

Then I take the aggregate of opposition strength, weighted to 11v11 minutes, and adjusted for home field advantage. This creates the schedule strength metric. Team performance is multiplied by the schedule strength inflator (or deflator) to create axGD/Mt.

There is a lot here that could be improved. This is just a toy model. These are my own criticisms of this model.

• The lack of other potentially-useful data. Proper power rankings would include performance outside the Premier League, and perhaps payroll and player quality factors.

• The use of simple aggregates without regard for the game-state contexts of these numbers. To what degree is performance in game states where the outcome of a match is less in doubt as indicative of a team’s quality? To what degree do changes in play style when teams are leading or trailing artificially inflate or deflate production?

• Throwing out minutes at uneven strength. Would it be better to include those minutes but with a deflator based on the change in expected performance? Here also a game-state adjustment would probably have interaction effects.

• The lack of time-series adjustments in the weighting. Do performances in more recent matches better project future performance? Does performance in the second half of the preview season better project future performance than how a team played in the first half?

• Penalties and other very high-xG chances. One of the reasons that goals is a less useful statistic than xG in small samples is simply that goals are all-or-nothing. Penalties are too. What about extremely high xG chances? A team creating a 0.90 xG chance has probably not played twice as well in attack as a team that created a 0.45 xG chance. Would these numbers be more predictive if very high xG opportunities were regressed somewhat?

• The exact rate of decay in projections. How much should past performance be weighted against current performance over time, and how much should goals and xG be weighted against each other?

• Finally, if I could build power rankings that are better than just a “toy”, it should be relatively easy to build a projection system and talk about how likely different teams are to win the league, get relegated, quality for the Champions League, and so on.

Any of these could be a proper Expecting Goals post on its own, and it’s work I am following up on.

Why Post This?

I am thinking about how to use different platforms to talk about stuff. All questions of politics aside, Twitter has become an utterly dead space for sharing ideas because it is impossible to direct people off-platform to any longer-format writing. I built my following through Twitter and never got into Instagram or YouTube or other platforms, and I find myself a bit stuck.

I am going to try out using Expecting Goals and the Substack platform for sharing my thoughts about football beyond the specific work of Expecting Goals studies. I look forward to seeing how you all feel about it and what direction this work can take.

Let me know in the comments if you have thoughts about the kind of posts we can add to Expecting Goals.

However, the game did have one unusual statistical characteristic, identified by the data and analytics company Opta in a tweet shortly after the match.

That’s pretty crazy! With 20 years of data, 380 matches per season, and 11 players per team, over 150,000 players had the opportunity to pull off this feat. But it wasn’t until Sunday that it happened — and twice in the same match.

It seemed like such an oddity that I had to investigate it, and that’s exactly what I’ve been doing for the past two days.1

Before getting started, there are a couple of important notes about the data. First, exactly how you’re defining a “pass” matters. Timber did in fact complete one of his three throw-in attempts during the match. But this tweet ignores throw-ins, which I think is reasonable enough. Second, and more consequentially for the post, this statistic was noted immediately after the match. Data companies typically provide a real-time feed which delivers data to clubs and media as the game is ongoing, and a second validated feed which is produced in the hours or days after the match. Presumably, somewhere in the process of producing the validated feed, an error was identified in the real-time feed.2 Timber did in fact complete one pass from an inbounds position and he is correctly credited with that pass in the database.

So, does that mean this match was not particularly notable? Or is a player completing no passes and their teammate completing only one still a pretty wild outlier? Also, the post mentioned only the Premier League. Have there been matches in other leagues where a player didn’t complete any passes?

Wait Is That What This Post Is About?

Yes. Yes it is.

I got interested in this odd question and I spent two days wrangling data to answer it. Perhaps this is not the most consequential topic in the world, but what is the point of having a newsletter and a database if I can’t spend a couple days playing around with some weird statistics and come out the other side with a post?

The “No Passing Score”

I took my database of matches, which dates back to the 2010–11 season, and I looked at the “Big Five” leagues across Europe. This means I have about 25,000 matches to consider and over 500,000 opportunities for players to complete a match without completing a pass.

I created a very simple score to rank them.

So here are all the players who were most unsuccessful at completing passes to their teammates in the Big Five leagues since the 2010–11 season, ranked in descending order by No Passing Score.

A couple things stand out. First is that my minutes calculation includes added time, and this match had a lot of it. Second, Timber’s match with just one pass completed in nearly full time is still among the 30 matches with the lowest recorded No Passing Score. Even with his single completed pass added back in, the match was still an extreme outlier. But I also found some patterns here, some types of pass-light matches, players, and situations, and I think they tell us a few things about what happened in this match.

Read more

The most notable finding to me in that study was that a regressed, weighted average of shots has a significantly worse correlation with non-penalty goals in the following season than a regressed, weighted average of non-penalty goals does. That is, if you want to know how many goals a player is going to score in the next season, you would be better off basing your projection on how many goals they have scored in the past than how many shots they have taken. Despite all the randomness in goal-scoring, it still has more “signal” of future goal-scoring than a count of shots attempted does.

I also found that expected goals was a better predictor of future goal-scoring than goals were.1

As I went over this finding on a recording of my podcast, the Double Pivot, I realized that this is probably a finding about expected goals per shot. If xG does “work” in this sense but shots does not, then it should follow that crucial information is contained in expected goals per shot. It is a player’s tendency to take higher- or lower-quality shots, and the ability of expected goals as a statistic to capture that tendency, which allows us to project goal-scoring using more advanced statistics.

But I had not run these correlation tests on xG per shot. So, today, I did that.

Here are the results.

Read more

There are many different ways to start to make sense of this data — this is what the newsletter is about! — but as someone who came into sports analysis through baseball, I have always wanted a projection system. If a player attempted three shots per match one season after averaging under two per match the season before, what should we expect them to do next season? What about a player who has a big season in assists or take-ons or interceptions? If a player has won a high percentage of their aerial duels in 1,500 minutes, does that indicate they will likely continue to win those duels at a high rate? 

The question of projection sidesteps the question of value. I don’t need to know whether a player’s pass completion rate increased or decreased the likelihood of a team scoring goals to estimate their most likely pass completion rate next season. There are ways into the value question from projection, but they don’t need to be followed in order to learn more about the game. At the same time, statistical projection raises many other complicated questions.

League translation, for one. If a player completed six progressive passes per match in the Italian Serie A, would we expect them to complete more or fewer in the Premier League? And of course this question dovetails with team and tactical context considerations. If the Serie A team had 44 percent possession and played a low-tempo defensive style, and the new Premier League team plays a high-pressing and high-tempo style with more possession, how should that affect our projection of their passing statistics? Even season to season in the same league, on the same club, a player’s team could change managers or tactical styles and the governing context of their statistics would be different. 

League, team and tactical contexts are huge questions in soccer analytics and this series will get to them in time. The problem is, you cannot identify the league or tactical context for statistics if you don’t first estimate what the baseline is for projection. That is, say we take that player who completed six progressive passes per match in Serie A. Say they completed five per match on their new Premier League team. To what degree could that change be attributed to contextual factors, or would it be indicative of a more typical level of regression to the mean that would happen in any player’s numbers? To understand league and team contexts, we need to have a handle on how to do projections in the first place.

So this is where I’m starting. In baseball analytics, we had something called the Marcel projection. That’s a Friends reference, which apparently I do not need to explain because Friends is the most popular tv show among Zoomers worldwide (or something). But I do need to explain why it has anything to do with sports analytics. The simplest form of statistical projection is a weighted average of past performance. Weight more recent production more heavily and production further in the past less heavily, add them all together, and include some regression to the mean component as well. That simple result, merely a regressed and weighted average, is the sort of statistical projection system that a monkey could build.

Then if you add just a few bells and whistles to it, some context adjustments and age adjustments, then it’s smarter than what a typical monkey could make. Marcel from Friends was smarter than the average monkey, so that would have been Marcel’s projection system.

In baseball, “some context adjustments” mostly means park effects, which are relatively trivial to calculate. In soccer, context is everything. Context is team and tactical effects, league effects, the aging curve, and more. Extracting a statistical signal of real player tendencies from all of these contexts is one of the holy grails of soccer analytics. Getting all the way to Marcel will be a journey. Just making the monkey smarter would be a major advance in soccer analytics, at least in the public sphere. 

So we will begin with the monkey. This involves a few distinct questions. What is the proper weighting of past season performance? How is the regression to the mean component calculated and what should its weight be? Should different weights and approaches be used for different statistics?

Read more