A few weeks ago, I ran 10,000 simulations of the Premier League to see where each team was likely to finish this season.

The Numbers Game

I Simulated 10,000 Premier League Seasons. Here's Who Wins.

I ran 10,000 simulated Premier League 26/27 seasons to see where teams are most likely to finish in the coming year…

Read more

a month ago · 53 likes · 30 comments · Dan Fleming

Given that the NFL Week 1 kick-off is now just around the corner, I thought it was time to try my hand at predicting how the season will play out here, too. As I noted in the Premier League piece, every analyst and podcast has an opinion on win totals, but what happens when we let simple play-by-play data run the season 10,000 times?

So, after 10,000 simulations of the 2026 NFL season, who wins the Super Bowl?

Remember, this is a project to see what can be done with open source data in a relatively short amount of time. I don’t have the resources that the NFL Media, Caesars Sportsbook, or ESPN have to run my models, so some executive decisions have to be made in order to get a representative model, without accounting for EVERY piece of news, coaching staff changes, or injuries.

Subscribe now

The Model

The first thing to decide is how we’re going to structure the model. Do we want to throw a million variables in to see what sticks, or build a relatively simple model that captures the general quality of teams over the past 5 seasons.

In using multiple seasons worth of data we, by design, begin to account for some variance in year-to-year performances that may be caused by injuries, coaching staff changes, or strength of schedule. However, by weighting it more heavily for recent years, we’re able to have the model use more recent data a little more to account for how things have been most recently.

For this model, I decided to make Expected Points Added (EPA) my guiding light. EPA helps us measure the net point value a team or player creates (or gives up) on average for every snap they take. The model for this looks at the down, distance, yard line, and time remaining before a play to assign an expected points value. The EPA is then calculated as the expected points value after the play minus the expected points value before the play. It tells us how well an offense turns situations into scoring opportunities.

Once we had calculated EPA, I used the nflfastR package in R to gather our data and nflseedR to simulate our seasons. This was set up to run 10,000 full-season simulations using the official NFL tiebreaker rules and a full postseason bracket. However, after the first run through there was a problem. When variances are so small, in a league like the NFL, the model suffered from severe regression to the mean. This left every single team falling in a 7.0 to 10.0 win bracket.

To correct this, I added a scaling factor to the net EPA differential when calculating each team’s win probability. In doing so, this expanded the net point differential from a 3.5 point spread, to a more realistic 18.5 spread for elite vs bottom matchups. This made sure that elite teams acted like elite teams, and struggling teams hit realistic floors.

What do 10,000 NFL Simulations look like?

The visualization here shows the distributions for each team’s projected regular season wins. This isn’t to say that this is the locked in, final NFL standings. In fact, it shows just how much overlap there is between teams. A couple of games going in either direction could be enough for a team like the Steelers to finish much higher than their average ~8 wins highlighted above. However, it does give us some insight into how we’d maybe expect some teams to perform more generally.

The Big Guns:

Buffalo and Philadelphia sit in a tier of their own at the top of the simulation. They are both projected to finish on 13 wins, with ceilings touching the 15-16 win mark.

The Contenders:

Baltimore, Green Bay, Detroit, Seattle, and the LA Rams form a tight, high-upside group around 11–12 projected wins.

The Middle of the Pack:

A massive block of teams (Texans through Saints) spanning 7 to 10 wins with wide, flat distributions. Small week-to-week bounces swing these teams from division contenders to top-10 draft picks.

The Bottom Tier:

The Jets, Raiders, Panthers, and Titans form a clear basement tier, centering around 3–4 wins.

It is worth remembering that in most cases, past performance is not totally indicative of future performance. However, by modelling EPA profiles over the course of 5 seasons, we do get to see more generally the quality of a team and how we might expect them to perform rather than simply mirroring what they did last time out. This is a nice way to frame the Chief’s projections of 11 wins - they’ve struggled more than usual in the last couple of years, and the model shows us that, but there is plenty of room for them to turn that around. They are not resigned to an 11 win season already.

The Numbers Game vs. ESPN

I mentioned ESPN as an organization that has many more resources than I do to predict sport performance. Fortunately for me, as I was putting together this article Mike Clay of ESPN posted his predictions for the year. I was able to work through this to plot a figure comparing my rankings to his to see how close (or far away) I am. I calculated a correlation coefficient to see how closely we match, a figure of 0.88 shows a very strong positive correlation between the two ranks, suggesting that we have a general consensus.

In the figure above, my model’s projected win counts are plotted along the Y axis, with ESPN’s Mike Clay’s projections on the X axis. By plotting it this way, we’re able to see where we agree or differ on teams. If a team falls perfectly on the diagonal line, Mike and I are in complete agreement. If a team features above the line, my model is more bullish on them to perform well. If they fall below, then Mike has them winning more games than I do.

The Bulls:

• Green Bay Packers (~11.5 wins vs. Clay’s 9.5): Strong underlying efficiency across recent seasons gives Green Bay a higher floor than consensus expects.

• Miami Dolphins (~7.0 wins vs. Clay’s 2.5): While Clay projects a total collapse, our weighted EPA signal sees Miami remaining competitive in the mid-tier.

The Bears:

• NY Jets (~3.0 wins vs. Clay’s 7.2): The model sees multi-year underlying EPA efficiency metrics as severely lacking compared to consensus public optimism.

• Chicago Bears (~6.5 wins vs. Clay’s 9.5): Clay projects a major leap; our 5-year weighted EPA engine demands sustained on-field proof before handing out 10-win projections.

• NY Giants (~5.0 wins vs. Clay’s 7.0): Model anchors heavily to multi-year offensive efficiency deficits.

The differences here highlight a difference in assumptions and information. With the time and resources available to Clay, he’s able to account for much more information than I am. However, I’m happy with how my model has performed here, especially considering it was free to fit, and done in an afternoon’s work.

The Business End of the Season

Regular season wins only get you so far, though. At some point, the attention has to turn to the post season and who’s going to secure a bye in the first round and who will secure those wild-card spots. The figure below shows teams with a high probability of making it to that stage in coral, with those in green having a moderate chance, and grey representing little to no chance.

The Buffalo Bills are set to dominate the AFC East (94.5% division probability) with a massive chance at the AFC #1 seed (54.1%) with home-field advantage through the playoffs and a 32.5% chance at winning the Super Bowl (the best of anyone in this simulation). The Eagles own the NFC East with a win probability of 96.3% with a 43.0% chance of the NFC #1 seed. The Ravens are one of the other clear standouts, with total control over the AFC North (86.4%) and a 43.0% chance of the #1 seed.

The NFC North seems to be a bit of a coin flip between Green Bay (44.8%) and Detroit (44.6%). Both have similar (~85%) playoff odds, but winning the division would provide an easier run to the main event. Similarly, the NFC West seems to be a three-way fight between the Rams (80.2% playoff), defending champs Seahawks (78.3%), and 49ers (65.6%). Whoever makes it through as champions will have truly earned their title in that division.

Teams like Tampa Bay, (72.8% playoffs, 2.1% Super Bowl) and Denver (70.5% playoffs, 2.1% Super Bowl) find themselves in a tricky position of being good enough to consistently make the post-season, but missing that next gear to take them through four straight playoff games.

The figure below gives a nice representation of team outcomes. I’m sorry if your team’s bar is largely grey, but you were probably expecting something along those lines already.

As I said earlier, this isn’t intended to be a ‘oh the season doesn’t need to be played, we’ve decided the outcome already’ type of list, but more food for thought. It should give a good idea of who we can expect to be stronger or weaker, and show just how much overlap there is in the NFL. With a short season (compared to the NBA, MLB, or Premier League), a single win or loss in a game week could be enough to tip the fortunes of your team.

What Did I Miss?

All models are wrong, but some are useful (Box, 1976).

My favourite quotation about modelling - all are wrong, but some are useful. I have no doubt that I’ll get plenty of messages highlighting how far wrong my model is, and how people disagree with it - and that is fine! It can’t be perfect, and the unpredictability is why we love sports so much. I hope this model can serve as a starting point for your thoughts for the new season!

Of course, there are impacting variables that play-by-play models can’t see, such as offseason rookie draft selections and rookie QB performance variance, free agency moves and late roster changes, or in-season injuries and coaching changes. All of these will impact the outcomes for a particular team, but none can be foreseen accurately.

With that in mind, my thoughts are this:

Buffalo and Philly will be elite, the Bills are the model’s favourites to win the Super Bowl, Green Bay are underrated on the whole (I’ll defend Aggie Alum Jordan Love all season), and it looks like a rough season for New York football fans!

The model has Green Bay significantly ahead of general consensus, do you think that’s right? Or is this a 5-year EPA signal that the off-season has already broken?

Did You Enjoy Reading?

If you’re enjoying The Numbers Game, consider upgrading to a paid subscription to support the analysis.

The Numbers Game Supporter: £6/month or £50/year

Everything in the free publication, plus:

→ Early access: read Monday’s article Sunday evening before everyone else.
→ Research Notebook: a short email each Sunday from inside the analytical session.
→ The Method: one monthly piece on a specific analytical decision in R.

The main Numbers Game articles will always remain free.
Supporting means coming further inside the process.

The main problem I have when writing these articles is figuring out how best to fit the model. The…

Read more

Once I got drawn into watching the team at Utah State play, with now NBA winners Sam Merrill and Neemias Queta on the team at the time, living just outside Salt Lake City didn't help. Before long I was spending far too much money on Jazz tickets, jerseys and sports cards, watching Donovan Mitchell, Rudy Gobert, Lauri Markkanen and Walker Kessler in person.

I now live back in the UK, making those late-night tip-offs much harder (thank you league pass), but I still follow the league (and obsess over fantasy basketball) every season.

It took a while to understand the patterns of play, offensive and defensive set ups, and how teams create opportunity in a whole new sport to me, but I got there eventually. What has interested me the most recently is just how different some teams are playing. When I was getting into basketball it was so heavily reliant on 3-point shooting, but today, there seems to be a much wider range in approaches.

So, what actually makes a great NBA offense in 2026?

Subscribe now

Defining Different Offensive Styles

To answer this question, I took five key offensive metrics that capture different ways of creating offensive opportunities; how quickly teams play, where they shoot, how they generate second chances, how they move the ball, and how often they get to the free-throw line. I ran a k-means cluster analysis across all 30 teams on these data points. K-means simply groups together teams with similar statistical profiles, allowing the data, not my opinion, to identify distinct offensive identities.

How many groups should we use?

If we had a strong theoretical reason for choosing a number, we’d use that. However, in this case, I decided to use the data to inform our decision as well. By plotting the chart below, we can use the point at which the line elbows to a different slope to identify the best number of clusters to use. As you can see below, this number was 4.

This left us with four distinct types of NBA offenses:

• Balanced Attack: These teams are flexible and use a range of approaches to scoring. Rather than specializing in one offensive weapon, they are able to spread their strengths across multiple dimensions, meaning they can break down defenses in a way that is most efficient for the current situation.

• Fast Break Machines: These teams keep the pace and tempo high and focus on ball movement. They’re quick to crash the offensive glass to create second chances. The combination of rebounding and ball movement is consistent with an approach designed to create additional possessions and attack before the defense is organized.

• Half Court Creators: These teams play at a slower tempo but really focus heavily on drawing fouls from half court play and crashing the glass. Comfortable when slowing games down, these teams rely on physicality, drawing contact, and generating second chance opportunities.

• Three Point Avalanche: As the name would suggest, extraordinarily high three-point shooting. Unsurprisingly these teams generate high shot volumes from behind the arc while sacrificing some rebounding and interior play in exchange for spacing.

What is nice to see here is that the four archetypes look very different to each other, showing us just how differently teams play across the league. This makes for a more enjoyable watch as a fan, but if you’re someone that enjoys the tactical matchup, it can be very interesting to see how teams and defense adapt to which type of offensive unit they are playing (I do believe that NBA teams play defense, unlike some…).

The interesting part isn’t necessarily that the shapes look different. Each archetype has a recognizable trade-off. The fast break teams stand out for pace, ball movement, and rebounding, but have a lack of three-point shooting and fouls. On the other hand, half court creators sacrifice tempo for physicality and second chances, while balanced attacks sit somewhere between the two.

Offensive Rating vs Possessions Per Game

If we take all 30 teams, and plot their points per 100 possessions, and possessions per 48 minutes, we get a very interesting figure. The more vivid the circle, the more wins the team got in the last season. Each circle is colour coded to the team’s archetype identified above. What is the first thing you notice?

If one offensive philosophy dominated modern basketball then we would expect to see one colour clustered near the top of the chart, but we see the opposite. The highest teams on points per 100 possessions are different colours. The Knicks, Heat, Nuggets, and Celtics represent all four archetypes, but score the most in the league on a per-possession basis.

The second most interesting thing here for me is that the high-win teams are spread horizontally. So being fast or slow doesn’t necessarily relate to the number of points you’ll score - so perhaps a system suited to your line-up, and executing on that approach, is more important than the number of possessions you get per game. Quality over quantity. This complicates the idea that simply playing faster is a recipe for offensive success. The league’s fastest teams aren’t automatically its most efficient, and some of its best offensive units operate at a much slower tempo.

One question you might be asking right now is how are Denver one of the slowest pace teams (low on possessions per 48 minutes), but classified as Fast Break Machines? The reason is that pace isn’t the only variable in the clustering algorithm. Denver generates transition opportunities through quick decision making, ball movement, and offensive rebounding. Their label reflects how they attack, not simply how many possessions they play.

My Utah Jazz… one of the most faded circles on the scatter plot, with only 22 wins last year. They scored fairly well. This shows how important both sides of the ball are. People love to say that NBA players don’t play defense, but I’d argue that this tells you that claim has little to stand on. How can one of the best scoring teams in the league finish winning just over 25% of their games? Poor defense… both matter! Offensive identity can explain how teams score, but it can’t explain everything. Championship teams still need an elite defense.

The Spurs and Pacers represent two sides of the same coin. Both teams play with a fast break style, but produced vastly different seasons last year, with the Spurs finishing on 62 wins while the Pacers ended on just 19. The style does not equate to quality. This really nicely shows us how important it is to execute the style you’ve adopted, and provides evidence for the potential for success using a range of styles. No single style is the key, you have to play well. The same blueprint produced both one of the league’s best teams and one of its worst. The difference here isn’t philosophy, but more the players executing it.

The Full Rankings

I started by asking what makes a great offense. The data showed me that there isn’t one answer.

Modern basketball is often described as becoming increasingly homogeneous, with every team chasing the same offensive ideals. This analysis suggests otherwise. The league contains several different offensive identities, and no single one dominates the top of the standings.

This is fairly rare in my opinion, with teams in other sports quite often identifying a new style that creates opportunity and then the rest of the league rushing to emulate that success. The top end of the league is represented across all four archetypes, the differentiator is execution, roster depth, and the other half of the court - which I deliberately left out here.

There are plenty of other offensive metrics that could have been included here. I deliberately kept the model simple enough to capture broad stylistic differences rather than every nuance of basketball. Different approaches to clustering could also produce different archetypes, but the central finding remains consistent: elite offenses come in many forms.

The two highest-win teams in the 25/26 regular season were OKC and the Spurs and had two very different playing styles (Half Court Creators vs Fast Break Machines). Is that just a coincidence, or does having a clear offensive identity, whatever that may be, matter more than the identity itself?

Every week, The Numbers Game goes deep into sports data like this. If you enjoyed, feel free to subscribe or send me a question for me to try answering using open-sourced data in the comments below!

As research, and data, involves so many decisions, I wanted a way that I could be a little more candid and transparent about why I made some specific choices along the way and what came to mind while I was writing. That’s what the Research Notebook is for.

While writing this week’s article, I was surprised by how difficult it was to obtain relevant data for the WSL. Given that the league has grown dramatically in the last 10 years or so, I was half expecting data to have been built into the system and be relatively easy to find. Perhaps this was foolish of me. With that being said, it means that we don’t have access to some data for the 2025/26 season just yet, so can’t get the complete picture. I think what I’ve done gives us the idea, but it goes to show how relying on open source data can limit your ability to answer the questions you’re asking.

I was also glad that I took the analysis one step further. Originally I was going to end the article at the league-wide trends, but decided that I needed to go a little deeper and see if things varied by team. Fortunately I did - the final figure reveals the most interesting finding of all in my opinion. The league as a whole isn’t changing too much, but some teams (Everton, Brighton, and Leicester) are gaining rapidly on the big three, while others (Man United, West Ham, and Liverpool) are falling behind. It shows just how much the management and administration of a team can impact their success against the big three and long-term development. I’m excited to see how this continues to play out over the next few seasons.

This is also an important lesson on taking your time and just a second to really think if you’ve fully answered your question. Data analysis is slow, and we shouldn’t be rushing it to get half of the answer!

That's the Research Notebook. A short email to paid supporters before I clear R each week. This is what it looks like. From next Sunday, the Research Notebook goes to paid supporters only. This is the last free one.

If you want it every week, you can upgrade below.

Over the course of the league’s 16-year history, it has been dominated by these three teams. In fact, one of them has won it every year in the last 11 seasons. The last team outside of those three to win the league was Liverpool in 2013 and 2014, then before that it was Arsenal again. On their own, Chelsea have won exactly 50% of titles, finishing top in 8 of the 16 available seasons - good news for Chelsea fans like me!

The Two-Tier System: Calculating the ‘Elite Tax’

This dominance of three teams has led to a two-tier system in the WSL. The big three fight each other for the title each year while the rest play for 4th. So, playing one of these big teams is obviously difficult for other teams, but just how much of a tax does playing them come with?

I calculated the average points scored per game for every team in the league when playing one of the Big Three, and when playing the rest of the league in the last five seasons. The differences are probably bigger than you’d expect.

When playing one of the big three, you’d expect to come away with somewhere between 0.28 and 0.44 points. Not a great return compared to the average points per game of between 1.35 to 1.42 when playing anyone else. Looking at the PPG gap over time shows no sustained convergence. After narrowing slightly in 2022/23 (0.97) and 2024/25 (1.01), the gap has widened to a 5-year high of 1.12 PPG in 2025/26. On paper, the league seems to be just as top-heavy as it was five years ago.

Subscribe now

Under the Hood: Actual vs. Expected Goal Difference

This is less than ideal for a competitive league that is trying to promote a more competitive product. However, this could just be due to some very unfortunate variance and luck, right? I explored this by examining the comparison between actual goal difference (GD) and expected goal difference (xGD) in these seasons.

This side-by-side will show us what the historical goal difference between big three and non-big three clubs has been, and what the expected goal difference between them would be. If the league is in fact becoming more competitive across the board, the xGD should at least be closing the gap a little - hinting at better chance creation against the big three, but just some unfortunate finishing.

In 2021/22, non-big-three teams were outscored by an average of 2.67 goals per game, despite expected goals suggesting a deficit of -1.64, a 1.03 goal overperformance by elite finishers. By 2024/25, that actual deficit narrowed to -1.85, while the expected goal difference shrank to -1.39. The earlier gaps could be explained by goalkeepers under-performing, the big three being particularly clinical with their chances, or the more than occasional wonder-goal that, by definition, is a low-quality chance and against the expectation.

The Breakdown: Individual Club Movers

This raises the question, is this trend uniform across the league? Or as independent entities are WSL teams making progress at different rates, or in different directions when it comes to performances against the big three? I looked at how this average goal difference has changed in the two three-season blocks (2021/22 to 2023/24 and 2023/24 to 2025/26), it is easy to see how some teams have improved, while others have gotten worse.

This makes sense if we look at some of the transfers these teams have made during this period. Man United sold Alessia Russo to Arsenal right at the start of the later period here, while teams like Everton have added England internationals like Ruby Mace to strengthen their line-up.

The Risers: Everton (+1.56 GD/game), Leicester City (+1.50), and Brighton (+1.17) have all shown improvements in the most recent three years.

The Fallers: Liverpool (-1.11), West Ham (-0.50), Manchester United (-0.39), and Tottenham (-0.36) have all fallen away from the big three, highlighting the impact of key departures (Russo leaving Man United, for example) and high squad turn-over.

This provides some promising news for fans of the WSL. The competitive nature of the league is somewhat conditional. Some teams are making big improvements, while others are getting left behind. The aggregate of this is that the league looks to be in a fairly similar position to 5 years ago, but when you examine individual teams, this is a little more promising.

To remain competitive teams are required to spend and focus on squad building more than ever. The London City Lionesses haven’t held back on this front since joining the league, and signing Alexia Putellas, Mary Earps, and Mapi Leon this summer transfer period has only strengthened the signal of their intent. More of this will only be good for the league, which will likely be cyclical in nature, allowing the WSL to bring over some of the best talent from the NWSL (like Chelsea did with Alyssa Thompson last summer) to strengthen the league even more.

What is Next for the WSL?

I’m looking forward to seeing how this develops over the next few years. This isn’t the type of systemic change that happens overnight - this will take years to play out with support from the league itself to allow clubs to make the moves they need to. A more competitive product will only lead to more eyes, more transfers, and a continually developing league.

We’re still watching the league grow, and as it grows the data will become more readily available and the sample size will grow with it. While we’re just looking at trends and patterns here, there seems to be a fairly clear signal. Multiple clubs are showing improvement above 1.0 goal difference per game, while most of the ‘losers’ are declining by less than 0.5, it isn’t revolutionary, but heading in the right direction.

What do you think the number one thing the WSL could do to promote a more competitive league-wide experience would be? I’m interested to hear your thoughts in the comments!

This was fueled in large part by the mandatory factory shutdown as the paddock takes a break. The time off in the summer gives us time to sit and take stock of the current situation. Which teams are in with a shot? Which teams have a driver in contention for the WDC? Which teams have lost all hope and should begin turning their heads towards the 2027 season instead?

This final question is particularly important, because if you’re one of those teams, you are basically driving what I like to call a ‘zombie car’ for the rest of the season.

What is a zombie car? I hear you ask…

A zombie car is one which has no hopes left for the season, it is already destined for the scrapyard. The only problem is that the team has to continue racing it at the remaining rounds of the championship. It is essentially a dead car walking (or driving).

So, at what point is the Constructor’s championship sealed? When should we be expecting teams to give up on this year’s cars and begin looking ahead. Or, when will the top teams be pushing until?

People often say that the constructor’s title is won by the summer, but this doesn’t seem to be the case in most seasons. The figure below shows the gap between first and second in the constructor’s title as a percentage of remaining points. It answers a simple question: if second place won every remaining point available, could they still catch the leader?

At the start of the year, round one, they are all grey, as there aren’t any gaps between constructors. As the season goes on, this gets darker and darker, until the leading team has mathematically secured the title - marked by a coral diamond. One complication is sprint weekends. A standard race weekend has a maximum of 44 points on offer for any single team, but this jumps to 59 on a sprint weekend. We account for this in the background here, so everything is accurate to the remaining points available.

Historical data were gathered from the open source Jolpica API and the retired Ergast database for this project. For each race in every season, I calculated the maximum number of constructors’ points still available and compared that to the gap between first and second as a percentage.

Perhaps the most striking thing here is that the title has only once been decided before the typical summer break period (between round 13 and 14), back in 2020. In every other season in the last 15 years, the battle has gone on much longer than this, sometimes all the way until the final round as in 2024. However, you will see that in many years, the gap starts to grow long before this final date.

It is important to remember that this figure shows the gap between the 1st and 2nd teams. What we don’t see is how far behind the bottom ranked teams are. So, if the title is clinched by round 17 or so, how long ago were the bottom teams driving these zombie cars? Probably long before the summer even started.

This comes with some problems. First, this is a major hit for the team morale. Engineers and drivers have to continue to show up each race weekend knowing that the main prize is already out of reach. While some will be able to direct that energy to next season, it could be difficult to deal with for others. Perhaps more directly related to the team itself, it also triggers a factory resource shift. When the data signal that the team is effectively out of contention, the incentives change. Teams might begin shifting wind tunnel time, budget, and R&D resources away from the current chassis and into next year’s car. This has the potential to widen the gap even further, creating this split in the pack between those battling for constructors points and the rest of the grid.

If the championship is effectively over for much of the grid before it’s mathematically decided, should the FIA do more to keep more teams in contention for longer? Or is it simply a consequence of rewarding excellence?

Let me know what you think in the comments!

The Premier League season is just around the corner, and by now I’m sure you’ve heard every opinion under the sun on who’s going to go down and who will be sneaking into the top four this year. Meanwhile, I’ve been thinking, what can we do with some simple data to help our opinions be a little more informed?

Over the last three weeks I’ve been posting a report card everyday for each team and their prospects for the season ahead. This article outlines the process I used to produce those report cards, and gives a little more insight into the findings of my modelling.

Subscribe now

Big companies like Opta collect thousands of data points every single premier league match day. However, football - at its core - is a relatively simple game. Each match produces two numbers, home goals, and away goals. Most matches produce between 1 and 4 total goals. This quite neatly falls into what we call a Poisson distribution. Over the course of 1000s of football matches, this is fairly replicable and the rarity can be estimated quite accurately. I’ve plotted the expected Poisson distribution alongside the total goals per game in the Premier League last year to show just how well the pattern holds.

What varies between teams is the rate at which they score and concede goals. Arsenal and Man City score at a much higher rate than Leeds. Liverpool concede at a lower rate than Everton. We can quantify this rate of scoring and conceding separately for each team, when playing both at home and away. Using this, we can calculate the most likely goal expectation for any given fixture.

There are 380 matches in a Premier League season. I built a model to calculate two numbers for each match, expected goals for the home side, and expected goals for the away side. It then simulates an actual scoreline by randomly drawing values from a Poisson distribution centered on these expectations. If we run that process 10,000 times for each fixture, aggregate all 380 results into a league table for each simulated season, and we have 10,000 Premier League seasons to make some predictions from.

The primary weakness with this type of model is that it slightly underestimates the number of goalless draws and close 1-0 matches when both teams are expected to score very little. There is an adjustment (from two statisticians, Dixon and Coles, in 1997) we can do for this, which counteracts that problem. This results in slightly more draws and 1-0 results than the original model would produce, which we know is more true to the real world.

For this project, I gathered the last 5 Premier League seasons (2021/22 - 2025/26) results from football-data.co.uk. To make sure that we reflected the real development of teams over those five years, I assigned a different weight to each season. For example, the 2025/26 season contributes 100% of its weight, while 2021/22 contributes just 15%. A team that dramatically improved last year is rated on that improvement, but their form of years past still plays a role, just to a smaller degree.

Now, for the promoted sides - Coventry, Hull, and Ipswich - who have little recent Premier League data. I assigned each of them a penalty relative to the league average. Their attack strength was set at 85% of the league average, while their defensive vulnerability was 118%. This means they should likely fall towards the bottom of the table and will be involved in a relegation battle - which most of us are expecting anyway.

What 10,000 Simulations Look Like

After 10,000 simulations, here is what I found.

The top three are in a league of their own but can barely be separated. Man City, Arsenal, and Liverpool’s distributions overlap massively between the 75-95 point range. Any of those three could win the league based on these predictions. Chelsea seem to be peaking at around 60 points and finishing in 4th place to round out the Champions League spots. The main thing to notice here, is just how big of a gap there is between the first three, and fourth place. This is the main structural finding of the model.

Man United and Brighton are statistically identical. Both average 53.8 projected points. Their distributions sit almost exactly on top of each other suggesting the model genuinely cannot separate them, which is itself informative - those fixtures could be key clashes to watch out for if they end up in a battle for Europa League football.

The mid-table cluster is difficult to separate. Brentford through to Crystal Palace have wide, flat distributions spanning 20+ points. This isn’t necessarily a modelling failure, it reflects the genuine unpredictability in that part of the table. Small differences in underlying ratings produce large variance in final positions when teams are closely matched.

Leeds - at the bottom? The model has a full season of their data from last year, but their distribution peaks at around 30-32 points with a ceiling of 42 points. This would just about keep them safe in years gone past, but would rely on them performing very well compared to what we’re expecting most of the time. So this pessimism around Leeds is data-driven, but they’re not necessarily resigned to relegation before the season has started!

The Numbers Game vs The Bookies

I run The Numbers Game in my free time outside of my day job, so I have limited time and resources to run models like this. I don’t pay for data (because I like to see what I can do with open-source data), and I just don’t have the time to create full, bespoke, datasets like Opta or bigger publications can.

In light of this, I was keen to see how my model correlates to the market. Specifically, how closely this model I put together fairly quickly using only publicly available data, compares to the betting companies (who are in the business of making money by finding edges in the data).

The first thing to note here, the two are very highly correlated, with a Spearman rank correlation of .93 (where 1.00 is perfect). This was reassuring to see, we agreed on a lot of teams, only differing massively on 3 and the relegation battle (this is where my limited ability to model the promoted teams comes into it).

The market suggests that Everton will be finishing in 10th. However, my model has them averaging 15th place, with a 28.9% relegation risk. Based on this analysis, their underlying goal metrics over the five-season window are weaker than their recent league positions suggest. If you think that Everton’s points totals have flattered their underlying quality - you might be in agreement here.

On the other hand, my model was more optimistic with regards to Fulham. Their home attack ratings are stronger than their recent performances might suggest. However, with plenty of transfers being made (6 departures over the summer, including Harry Wilson), the bookies might be accounting for this more heavily than I am.

Another divergence from the market comes in Aston Villa. The market has them around 6th, while my model sees them at 8th. Last year’s European competition likely impacted their ability to produce high-quality metrics in the domestic competitions which could mean my model is underrating them. However, in recent news the departure of Morgan Rogers might play into them finishing lower than 6th if they can’t replace his output - although again, my model doesn’t account for this.

Finally, the unpredictable mess. Coventry, Ipswich, Hull, and Leeds all sit close to the diagonal but at the bottom end of the table. The model and the market are aligned in that, but my model seems to have the order different almost every time, with some the average finishes being somewhere closer to 17th for each team - suggesting that they swapped orders quite a lot in the 10,000 simulations. My model can’t tell you which order these four finish, but it looks like all four are in serious trouble.

The Breakdown

So we know who the model likes and doesn’t like, and we know how that compares to the bookies. All we need now is to understand what the model actually says happens to each club.

The title race is one of the three-horse variety. Man City hold the highest title probability (by quite some way), but Arsenal and Liverpool emerge as genuine contenders across the 10,000 simulations. The coral segments are all visible and meaningful across all three, so this isn’t a foregone conclusion.

Chelsea seem to be back in the mix! Well, for the final Champions League spot at least. Their top 4 probability is high, but their title probability is basically 0%. We’ll see how Xabi’s men do after some big changes in lineup this year.

Leeds look to be in trouble. Their route to survival seems to be outdoing one of the newly promoted teams, but I’m not sure who that will be. From what I’ve seen, Hull seem to be everyone’s easy pick for relegation, so that leaves either Ipswich or Coventry, both of which will be tough challengers. Who do you think will be the bottom three this year?

The Europa League spots are a bit messy. Most teams have some probability of finishing there, which is the model essentially admitting that predicting 5th-7th is very difficult to do. A small rating difference will produce a large variance in finishing position after a full 38 games in this range. Talk about marginal gains.

What Did I Miss?

All models are wrong, but some are useful (Box, 1976).

Like I said, I don’t have the time or patience to gather information on every transfer, player level modelling, or managerial tendencies to account for those in my predictions. As far as a predictive model goes, this was fairly rudimentary. We can’t account for injuries, transfers, European competition impacting league form, domestic cup runs, or tactical changes. That doesn’t mean it is useless, though. Here are the strongest predictions - Man City 1st, Chelsea 4th, Everton 15th, and Leeds most likely to finish in 20th.

Each gameweek, I’ll be able to tell you if reality is running hotter or colder for each team than the model expected. We’ll find out together.

Enjoying The Numbers Game?

This simulation took around 6 hours to put together, gathering the data, building the model, running 10,000 simulations, designing the charts, and writing the article.

Every article takes some time, but the analysis stays free for everyone, always.

If you’ve enjoyed reading The Numbers Game and would like to help me keep building bigger models, creating more visualizations, and publishing articles like this every week, you can become a Supporter.

Become a supporter for £6/month or £50/year, helping me dedicate more time to putting together these more in-depth articles. There’s absolutely no pressure. Reading, sharing, and commenting on my articles helps enormously too.

Either way, thank you for being here and let me know what to explore next!

This year, I decided to go particularly deep into FPL and ended up winning my work league and finishing in a respectable 558,626th (top 4.3% of managers) - not too bad out of over 13 million managers in just my second season.

It goes without saying, I still had some very rough weeks. Raging on a Monday night when a player I sat on my bench drops 13 points. Or refreshing my FPL app to check if that player with a rumoured injury is going to start while I’m out with friends. Now we’re only a month or so away from the new season starting, and with the new game just launched, I thought I’d use this as an opportunity to apply some data storytelling to the art of FPL, because one of the keys here is to see where luck ends and skill begins.

So, who got lucky last year? Who was unlucky? Are they likely to do similar this year?

To answer these questions, we can use xG and xA to see who performed better (or worse) than we might expect them to. Before that, we need to establish what these things mean. Expected Goals (xG) measures the probability that a shot will result in a goal. Expected Assists (xA) measures the quality of the scoring chance a pass creates, basically, the xG of the shot that followed. They take into account things like player position, actions leading to the shot, and patterns of play. Each individual chance is scored from 0 and 1, where 1 would equal a goal 100% of the time. These can be added together across a season to give the total amount of expected goals.

We can use these statistics to see how players perform based on what we’d expect them to accomplish. The first figure below, shows how forwards and midfielders (with at least 5 xG and 900 minutes in the 25/26 season) scored relative to their xG.

In the top right corner, we see Haaland. His expected goals was just over 25, and he scored 27, so he just outperformed what we’d expect him to do. Thiago was a similar story, slightly outperforming his xG with actual goals. Those that rise above the line scored more goals in the 25/26 season than we’d expect them to. Those below the line underperformed. The over performers worked well with the lower quality chances they were afforded. This isn’t too much of a problem with Haaland or Thiago, who only slightly outperformed their xG, but for a player like Semenyo, who scored significantly more than we’d have expected him to, we might need to consider a regression to the mean this season.

On the other side of that line are the players that underperformed. The question here is if they were unlucky and faced some exceptional defending/goalkeeping, or were just out of sorts when it came to finishing. These players could be undervalued FPL assets this year, about to bounce back after some poor variance - but they could also be on a downtrend and might need to be avoided. This is where you need to have some sort of strategy ahead of time to avoid sinking weeks and weeks into an underperforming asset that you’d be best to pivot away from earlier.

Assists are another key point scorer in FPL (3pts), only one less than a goal scored by a forward. To optimise our point potential, we should be trying to identify players with high assist threat to complement our goals. The standout, and huge outlier, here is Bruno Fernandes. He massively overperformed his xA with his actual assists, along with (but to a lesser extent) Bowen, Saka, and Cherki. This positions him as a solid option for a set-and-forget style player, who will produce and outperform many others. This is true even if he doesn’t perform quite as well as last year. You can see below that he not only had the highest number of assists, but also the highest xA - so if he does get a little unlucky this season, his xA numbers are still good enough for us to expect a pretty good return on assists (just probably not as many as last year).

With FPL being a game about probability and expected value, I created a regression risk index by calculating a ‘luck score’. This figure captures how much a player outperformed their xG + xA (of course this isn’t always purely luck, but they’ve been on the side of favour in some way). Those with a positive score (coral) indicate players who dramatically over performed last season and present as the biggest risk for regression and a step back this year. Those in teal are players who underperformed, and therefore could provide sneaky value in 25/26.

Players in the positive might still perform well, but I’d be wary of expecting them to produce at the same level as last year (i.e. Fernandes, Bowen, and Semenyo). So select (and triple captain) at your own risk!

The final thing we need to understand before we make any big decisions on what to expect this year is if history tells us to expect this regression and fall from the over/under performers.

The figure below shows the last 3 pairs of seasons (23/24, 24/25, and 25/26) and plots player’s goal over performance in one year against their performance the next year.

You’ll see that the line is fairly flat, but does tick slightly upwards. This would indicate that just because a player had a standout over performance season one year, doesn’t mean that we can expect that they’ll go either way (up or down) in the season that follows. If anything, the slight uptrend suggests that elite players have a slight tendency to repeat their over-performance, which is exactly what you’d expect from players who are simply better than the model assumes, but this isn’t a steep line by any stretch. So one of our ‘lucky’ players last season might not regress, but we also might not want to expect them to continue over performing to the same degree year after year.

So, don’t avoid a player just because they performed well. After all - the elite are the elite for a reason. BUT, remember the volatility, don’t blindly expect past performance to predict future performance, and be critical when a player is vastly outperforming their xG and xA consistently. The goal for me this year is to identify players that fall into one of these categories.

First, the under-performers from 24/25 - these will likely be cheaper assets this season who could bounce back towards what we’d expect from them. They could also be some nice differential picks. Second, Haaland and Thiago. Yes, Haaland sits in the regression risk column, but he outperforms the rest of the league so clearly in both xG and goals, that even if he does regress, he still provides more value than most other players. Third, players who had some fortune last season but still created enough underlying quality that a return to their expected output still makes them strong scorers - Fernandes and Cherki being the obvious examples.

I know I’ve made FPL seem straightforward here, it’s not. It’s a game of probability and making the best decision available with the information you have at the time. Your job as a manager is to put your team in the best position to score lots of points, you can’t get wrapped up in the week by week variance and luck. Don’t keep a player in your team because you’ve had them in for a while and it is their turn to pay off. That screams sunk costs fallacy - where we’re reluctant to pivot because we’ve invested so much already. Sometimes the best move is to cut your losses and move on.

It will take some luck to win and perform well, but consistency over 38 game weeks is the most important thing.

If you’re playing FPL this year and enjoyed the article, join The Numbers Game mini-league here and get active in the chat!

May your green arrows be plentiful, and bench points remain minimal 🫡.

Who will you be choosing in your starting 11 this year? Any nailed-on-picks? Any crazy differentials? I’m keen to hear in the comments!

With that said, this means the pool for potential talent to come from is huge. Of course, different countries have differing levels of investment into infrastructure and youth development, but it is not uncommon for some of the world’s best players to be born somewhere without a traditional powerhouse sporting set up. However, if you take a moment to look back at the history, you see World Cup domination from two areas in particular.

Mapping Confederation Dominance

Using the historical World Cup dataset from Joshua Fjelstul, I explored which confederations (UEFA, CONCACAF, CONMEBOL, CAF, OFC, and AFC) tournament semi-finalists have come from each decade. Taking a decade by decade approach helps us to see the shifts in power over a longer period of time while accounting for the one-off Cinderella stories, groups of death (see here) or lucky/unlucky routes to the latter stages of any particular tournament.

Making the semi-final would be a tournament run that most nations, regardless of pre-tournament expectations, would be fairly happy with. It marks a level of quality that only the best teams will attain on a consistent basis. It is also worth noting, this figure is plotted at the confederation level, so if any team makes it to the semi-finals, the entire confederation is coloured in.

Subscribe now

The immediate stand out is the dominant dark red of Europe (UEFA) and orange of South America (CONMEBOL). This makes sense, given the dominant forces of France, Germany, Brazil, and Argentina falling in those confederations. This highlights a geopolitical issue for FIFA - if football is truly the world’s game, then why is it dominated so heavily by two regions? The game has been globalized in terms of viewership, revenue, and player migration, but this doesn’t seem to have disrupted the geographical regions that consistently produce the strongest teams when the World Cup comes around.

From Regions to Countries: A Sea of Grey

There are some exceptions that prove the rule visible in the figure. The first ever tournament in 1930 saw the United States make the semi-final in Uruguay. South Korea disrupted the scene when hosting in 2002. More recently in Qatar 2022 we saw Morocco surprise the world and make a run deep into the tournament - the first time an African nation broke into the final four. However, these are, in general, the fairy tale stories.

To show that more clearly, I also plotted the same figure at the country level, highlighting only the country itself that made the final four. It can only be described as seas of grey, with blocks of red through Europe and South America.

The 2026 World Cup presents a major structural change to the tournament structure by expanding the tournament to 48 teams. Since 1998 (the last expansion), the tournament has featured 32 teams. The additional 16 slots provided by this expansion have offered more guaranteed qualifications to Africa (CAF), Asia (AFC) and North/Central America (CONCACAF). This seems to be a strategic move from FIFA to challenge the dominance of Europe and South America by expanding pathways for other countries and confederations to be competitive on the global stage.

Take a look at the 2020s block on the map. The current tournament wraps up that block, so we had a front seat to see how this story ended. We know the final four was made up of England, France, Spain, and Argentina. Yet again, Euro and South American dominance. However, all but one of the African teams that qualified progressed from their group, marking a shift in the powers at the early stages of the tournament at least. As with most structural change, this will take years to play out, potentially needing us to wait until the 2030s or 2040s to see the results of the changes being made today.

Will the 48-Team Expansion Change the Picture?

Do you think the 2026 World Cup expansion to 48 teams (and rumored 2030 expansion to 64) provides an opportunity for other confederations to make it deeper into the tournament, or does it just dilute the quality of the tournament, giving the old guard easier group stages to navigate?

Every week, The Numbers Game goes deep into sports data like this. If you enjoyed, feel free to subscribe or send me a question for me to try answering using open-sourced data in the comments below!

While we hear statements like this thrown around all the time, especially during the World Cup, it raises the question how common are upsets in the World Cup and when do they usually happen?

Subscribe now

In this article, I take a look at the rate and occurrence of upsets in the World Cup between 1998 and 2022. I chose these dates because they all had a consistent tournament structure, with 32 teams, which means that structurally, there should be no impact on the number of upsets. An upset here is defined fairly simply as a game in which the lower-rated team won the match (be that at 90 minutes, extra time, or penalties).

To quantify rating, I’m using football Elo (similar to that used in chess). This is a metric that accounts for multiple data points, including the kind of match, goal difference in matches, home team advantage, to calculate a team’s rating that updates after every game they play. To quantify the size of an upset, we are simply looking at the Elo difference between the two teams when the underdog won.

Are Upsets Becoming More Common?

First up, we’re looking at if upsets are becoming more common. It looks as though we are in an era of football where upsets are more common than in years past. The 1998 and 2002 tournaments saw relatively few, whereas in more recent years we’ve seen upset rates be much more sporadic. Russia vs Spain in 2018 springs to mind pretty quickly here as a standout upset example, but the rate of upsets are scattered from year to year so we don’t see a neat and tidy trend. This speaks volumes to the unpredictability of football, especially in an international tournament where every game is all-or-nothing. So it looks like the rate of upsets is volatile, and will differ from tournament-to-tournament. This is interesting in the sense that it is so unpredictable - we could get 11% upset rate like in 2002, or nearly a 50% rate like in 2018. The beauty of the World Cup.

Which Stage Delivers the Most Drama?

With that in mind, which stage of the tournament provides us with the most upsets? This is where we have some interesting findings to break down. Below you’ll see that the final typically results in an upset 29% of the time, but there are a couple of caveats to this. First, you can see the error bar is HUGE. As there is only one final per tournament, the sample size for the final is especially small, meaning one upset can carry a huge pull in the results. Secondly, if we’re defining an upset as the underdog winning, this needs some thought. Both of these teams have made the World Cup final - it is rare for there to be a true, massive, underdog at this stage. So teams may be more evenly matched, but the higher ranked team still wins the majority of the time.

The second most common upset stage (although only 1% different to the final - so within margin of error) is the quarter finals. My thought here is that this is the point where the weakest teams have been eliminated in the round of 16 with just a 21% upset rate, which brings teams into the quarters with a closer gap between them, which might lead to more upsets as the underdog is much more closely rated compared to the favorite. However, the quarter finals are still early enough in the tournament for us to see one or two Cinderella runs, before they get eliminated at the semi-finals. Think Morocco in 2022.

The Anatomy of an Upset: The 250-Point Window

The figure below shows the pre-match rating difference between teams at each stage of the tournament, with upsets shown in red. In general, you can see the rating difference shrink as the tournament goes on, lending to the idea above that upsets are less dramatic as you progress through the tournament, as favorites are less established (e.g. France v Argentina). The stand out here is Spain and Russia back in 2018. The largest pre-match rating difference that ended in an upset ever, although typically we see upsets happen when teams are within 250 Elo points of each other. While not causal, it is good to know that the huge upsets are few and far between, especially when we get to the semi-final and final, you can see there have been no upsets when the Elo rating is more than 100 points different. This is a lovely visualization of variance in action - the earlier stages, with more teams, have a much greater potential for upsets, but this disappears as the tournament progresses.

The Biggest Knockout Upsets in Modern History

So, what are the biggest knockout stage upsets in modern World Cup history? Well, as I just mentioned, Russia beating Spain in 2018 is the number 1 with an Elo gap of 364 points. This is closely followed by Senegal beating Sweden in 2002 (348 point gap) and Morocco beating Spain in 2022 (283). All three of these upsets came in the round of 16 which is interesting, especially when we now have a round of 32 in the 2026 tournament, I wonder how that will play out as far as upsets go. Croatia feature in three of the 10 games, beating England in the 2018 semi-final, Brazil in the 2022 quarter final, and Germany in the 1998 quarter final. I wonder if Croatia have been chronically underrated - especially in the Luka Modric era.

Perhaps this comes down to a measurement issue - defining performance in football is difficult. I wonder if the Elo system isn’t quite dialed in if it keeps underrating Croatia - maybe there are some other teams that would fall into this underrated category. I suspect that Morocco may fall into this group at the close of the 2026 tournament. This would mark two tournaments in a row where they’ve at least made the round of 16, well outperforming general expectations.

I personally think Paraguay v Germany might make it onto the top 10 list once this tournament is complete. I don’t think people will be forgetting that one for a long time.

A final note to consider, as usual with World Cup data, the sample size is small, especially as you go deeper into the tournament. After all, there is only one final every four years. So, we have to take these findings with a pinch of salt. As I mentioned earlier, there is also some caveat to the idea of an underdog in some finals - like France v Argentina. Any ‘underdog’ in that situation is so by a marginal difference in rating, not really a true upset as most of us would think of when we hear the word.

Where on the top 10 list would you place the big upsets from the 2026 tournament so far?

Building on last week’s post, this article is my biggest yet. More figures, more data, and a more detailed narrative. If you’ve enjoyed it, please consider subscribing to make sure you get future editions right to your inbox.

Subscribe now

All content here will always remain free. That said, if you’d like to buy me a metaphorical coffee and support the work behind The Numbers Game, a small paid subscription option is available and very much appreciated.

The 2026 World Cup feels like a high scoring tournament, but is it really?

To find out, I set about gathering open-source data. This dataset comes from the Fjelstul World Cup Database. Rather than going back to the tournament’s start in 1930, working from the 90s onwards felt like a more honest baseline to capture ‘modern’ football. To smooth out noise between tournaments, I used a rolling average across five consecutive World Cups rather than treating each one in isolation. This helps avoid the dramatic influence of one particularly high or low scoring tournament. 2026 data is Group stage only!

The trend is clear here. Goals per game peaked at around 3.1 in 1998 and has generally fallen since. There has been a slight uptrend since 2014, but we’re still well below where we started. The 2026 group stage is sitting a little higher than recent history, but it’s recovering from a long decline, not breaking new ground.

Here’s where it gets more interesting. The downward trend in the World Cup has happened at the exact same time that club football is seeing a boom in goals per game.

The contrast here couldn’t be much clearer! While the World Cup has been on the decline for the last 30 years, European club football has been on the exact opposite trend. The Bundesliga now averages close to 3.25 goals per game. At the club level, football is seeing more goals than ever, so why is the World Cup so different?

Subscribe now

Before we get to why, it’s worth making sure that this difference actually means something, and we are not just comparing competitions that started at different levels. To check, imagine we set every competition back to zero at the start of the 1994 season and then track how much each one has scored more or less goals since. That’s all the next figure does. A score of 100 means that goals per game is exactly where it was in 1994. A score of 110 means 10% more goals per game than in 1994. Below 100 means fewer goals per game.

This figure tells the same story. The World Cup line is jagged compared to the leagues, but that’s what happens when you only have one data point every four years, so a single high or low-scoring tournament pulls the rolling average more aggressively. Even then, the direction is clear enough: most European leagues are scoring more than they did in 1994, while the World Cup is running about 10% below its own baseline. The only club level competition that has less goals per game in 2026 than 1994 is Serie A - but even this is only by a couple of percent. The 2026 World Cup is on track for a slight improvement, but we’re only looking at group stage data, so there is plenty of room for some cagey 1-0 knockout matches to pull that average down.

So, where have all the goals gone? The final figure breaks this down at match level. The most telling change here is the share of games with 4 or more goals. This shift, from 35% to 25% tells us where goals have disappeared from. It isn’t that we are getting more and more 1-0s, but just fewer thrillers with 4 goals or more. This isn’t to say that they don’t happen, we all remember Brazil vs Germany in 2014, but they are the exception rather than the rule.

So, we are seeing less goals, that much is true. This might explain why games like England/Croatia, Canada/Qatar, and Germany/Curacao this year have stood out so much - because they are genuinely rarer.

What we don’t know, is why this shift is happening. Is it that tournament football is structurally different to club football? The immediacy of the payoff (elimination vs progression) might put teams off pressing continuously if they have a 1-0 lead in the last minutes. It could also represent a shift in defensive organization and risk taking behaviors in international football. When the stakes are higher in the 80th minute with a small lead, the backwards pass to protect possession makes more sense than pushing on for another goal. Maybe it shows progress being made in squad depth at the international level which is creating a more balanced playing field - compared to club football which relies more heavily on transfers and money.

It will be interesting to see how the 2026 tournament shakes out. I hope we continue to see plenty of goals. I think we’ll slow down as we progress through the knockout rounds. What explanations for this difference in growth have I missed?

This piece took considerably longer to put together than anything I’ve written here before. Two datasets (comprised of too many CSVs to count), multiple figures, and much more time spent writing. If you find value in this kind of analysis and want to support the time it takes to produce, the best thing you can do is share it with someone who’d find it interesting. Oh, and subscribe for free if you haven’t already.

Before a tournament starts, analysts usually reach for the FIFA world rankings when trying to identify the Group of Death (or as ESPN did, the world football ELO ratings). This makes sense when we don’t have any match data and this is the best indicator of team quality. However, since we’re looking at this retrospectively, we can use actual performance to quantify which group was the most deadly.

Traditionally, ‘deadly’ is used to describe the group where good teams might go home early. I’m defining it differently here: the groups whose teams collectively performed best across the whole tournament. The two aren’t always the same, but that tension is exactly what makes this interesting. In looking at it this way, we see which group of teams performed the best during their tournament runs, not based on history as world rankings are.

Using data sourced from The Fjelstul World Cup Database we can create a score for each group based on traditional football point scoring. For each group in World Cup history, I assigned teams 3 points for a win and 1 point for a draw in every game they played through their entire tournament run, from the group stage through to whichever round eliminated them. By taking the sum of these scores, we can see the overall quality of a group. One team winning the tournament means little if the other teams didn’t win a game, or got eliminated in the first knockout game. To trim this visualization down I’ve selected just the top 10 and bottom 10 groups in World Cup history.

Unsurprisingly, the bottom two results come from the first ever World Cup in 1930 in which groups only contained three teams - limiting the amount of points teams could score by the end of the group stage to 6 rather than the 9 of today, and with fewer games across the tournament, there was simply less opportunity to accumulate points.

More recently we see the 2018 Group B (Spain, Portugal, Iran, and Morocco) in the bottom 10 along with the 2010 Group C (United States, England, Slovenia, and Algeria). These two groups ended with teams winning a maximum of one game, with a lot of draws. Not only this, all four of the teams that progressed (Spain, Portugal, England, and the US) lost in the round of 16, leaving them with just 5 points each. Tournaments to forget for those nations, yet I still remember watching England with my pals and leaving rather deflated after that Frank Lampard ‘goal’.

At the top of the list, the most deadly group of World Cup history, we have Group C from 2002 (Brazil, Turkey, Costa Rica, and China). Carried largely by Brazil winning the tournament, this group also saw Turkey achieving their highest placed finish ever - coming in at third place. Costa Rica also managed to amass 4 points in the group stage, adding to the group total.

So, the title has been claimed. Group C of 2002 - the deadliest World Cup group of all time. I’ll have to come back and take a look at this once the current tournament has finished, to see which group scores the most points and earns the 2026 Group of Death title. Of course, with third place teams now able to qualify to progress from the group stage and the addition of a new knockout round, this will dramatically shake up the number of points a group could score (more teams can now score way more points).

Do you think the 2026 format changes make it harder or easier for a single group to claim the all-time title?

Of course, it could be pure chance - things often are - but it could also reflect a whole host of other factors. Home support springs to mind as one explanation, perhaps it is due to the hosts naturally having less travel to do right before the tournament, or maybe national federations are particularly strategic in the years leading into a home tournament. This final idea caught my interest. After reading Das Reboot a few years ago, I was fascinated by the DFB’s approach to player development in the early 2000s and 2010s. Similarly, the English FA have been very vocal about their England DNA development plan - which seems to have worked nicely given the recent tournament runs of the Three Lions. So, I decided to see if there was some credit to the idea that host nations typically have years to prepare for the tournament, perhaps they are able to develop and build a core group of players ahead of time.

So thanks to Josh Fjelstul’s brilliant World Cup dataset, I calculated a squad continuity statistic: the percentage of players who overlapped between their nation’s hosting tournament and the one prior. I then compared that figure to the same team's typical squad continuity in non-hosting tournaments, to see whether hosts behave any differently than they normally would. Now, it is important to remember that sample size is relatively small here - even more so than usual because some hosts (like Qatar) don’t have a previous tournament prior to hosting, so take these findings with a pinch of salt. This analysis also spans across eras, so the hardcore developmental approach that most federations adopt these days might look a little different to those who hosted back in the 50s/60s.

Unlike last week’s data visualization, the picture here doesn’t point in a single direction.

The most striking cases are England and Germany, both of whom show a very large positive gap in the figure. This indicates that they both arrived at their home tournaments (1966 and 2006 respectively) with a notably higher squad continuity than is typical for them - of course, England won and Germany finished third in those years respectively. Russia also followed a similar pattern when they hosted in 2018, outperforming their historical performances. Whether the continuity was the cause or simply a byproduct of systems surrounding these teams is unknown and impossible to say from these data alone, but for anyone who has read Das Reboot, this might be an interesting finding.

Then there comes France ‘98, sitting firmly in the negative. Deschamps lifted the trophy with a squad that had significantly less continuity than Les Bleus typically manage across tournaments - so much for a neat narrative in this one.

On the whole, hosts are roughly spread evenly across the three camps of higher, the same, and less continuity than normal. There’s no version of this figure that screams out ‘federations systematically build stability ahead of a home tournament’ - at least not in a way that shows its head in squad continuity. A few of those toward the bottom of the figure deserve a little asterisk. Brazil, Chile, and France’s lower continuity figures are also partly a product of long gaps since their previous World Cup appearances. These include the cancelled 1942 and 1946 tournaments. I’ve flagged this in the figure note, but it’s worth keeping in mind when reading those lines.

So, where does this leave us? Hosting the World Cup seems to help - that much is pretty clear. However, if the mechanism was as straightforward as ‘build a settled squad’ then we would expect to see a much clearer pattern here. We don’t. That doesn’t mean preparation isn’t important, it simply suggests that squad continuity isn’t where the story lies. Perhaps home support, reduced travel, familiarity with conditions, or broader investment in infrastructure matter more. Or maybe the mechanisms are too subtle for this type of analysis.

We'll find out soon enough whether the USA, Mexico, and Canada continue to outperform their historical performances, but I'm still no closer to knowing why hosts tend to do so.

I’d love to hear your theories.

To answer that, I found some open source data covering World Cup tournaments from 1930 - 2022, thanks to Josh Fjelstul’s fantastically maintained github repo. For every host nation since 1930, I took their finish as the host and compared it to their average finish across every other World Cup they’ve qualified for (both before and after hosting - so the full career average not just recent form coming in). The idea is to understand whether something about hosting leads to reaching a further stage in the competition. For countries that have hosted multiple times (like France, Mexico, and Brazil) I used their average finish as host.

The result: 16 of 19 hosts outperformed their career average, several by a wide margin. Uruguay (1930) and England (1966) both jumped the equivalent of 5.5 tournament stages compared to their historical norm - of course, both went on to win their respective tournaments. Chile (1962) and Argentina (1978) weren’t far behind in improvement either.

Only one host got worse relative to their average, Spain in 1982. South Africa (2010) and Qatar (2022) were the only two who landed exactly on their typical level - likely due to typical group stage exits in the case of South Africa, while Qatar’s time as host was their first qualification.

However, most countries have only hosted the World Cup once, so we’re working with a very small sample which means a single data point (one great or terrible tournament) carries a lot of statistical weight. I wouldn’t run a regression on this and call it proof, but with 16 of 19 cases pointing in the same direction, it’s a pattern worth considering. With this in mind, I’m keen to see how this year’s three host nations continue to progress through the tournament. Will the home advantage continue to play out in this way, with teams over performing? My money’s on yes - but I’d love to hear your theory: crowd support, lack of travel, or federations simply building stronger squads around a home tournament?

This is the first post on The Numbers Game, if you’re into this kind of historical sports data digging, I’ll be doing one of these most weeks. If that sounds like your thing, subscribe below and you’ll get the next one straight to your inbox. While the World Cup is ongoing, I’ll be focussing on that, but will open up to other sports such as basketball and baseball as those seasons continue and pick back up! Feel free to reply with ideas or questions you’d like me to try answering, or other sports/datasets you’d like me to take a look at!