Syzygy Analytics

Summary

In this study, I analyse the performance of soccer teams in different game-states (score-lines) in an attempt to answer the question posed in the title.  I show why a component of success for leading teams is low-scoring games and not just a large xG/goals differential/ratio.  I compare team performances across a large sample for tied teams and one-goal games. Teams leading by a goal after the 70th minute demonstrate weaker performance compared to their previous level at a tied state. I introduce an idea to quantify how big the role of loss aversion is by comparing the performance of teams who see a draw as a loss to the performance of teams who likely see a draw as a more comfortable result.  This leads me to the conclusion that parking the bus is a viable option.  When allocating a sizeable proportion of the weaker performance of the leading team to the loss aversion that the trailing team experiences we see that total goal suppression offsets performance degradation enough to potentially render parking the bus an EV-neutral strategy.  Isolating deliberate tactics from psychological effects and weak performance are particular challenges.

Introduction

The phrase ‘Parking the bus’ has long had equivalent phrases in Spanish/Portuguese but became popular in English speaking football due to comments José Mourinho made early in his career.

The phrase refers to a football team carrying out defensive tactics. i.e. they are so defensive and sitting so deep, it’s like a bus is parked in front of their goal. The tactic is likely performed by a ‘protecting’ team, either protecting their lead or possibly a draw.

Game state (the current scoreline) exerts various influences on football games.  Motivations shift, play styles shift, outcomes shift.  Psychological human factors can also play an important role. 

I was encouraged to undertake a study on the topic after reviewing two recent papers on the topic:

“The Impact of Formations on Football Matches Using Double Machine Learning. Is it worth parking the bus?”  https://arxiv.org/pdf/2602.16830 by Ruiz-Menarguez and Badiella

The authors analysed performance across various tactical formations concluding that there is no statistical evidence that ‘parking the bus’ offers any scoring or winning advantage.

“Parking the bus” by Guan, Cao and Swartz https://www.degruyterbrill.com/document/doi/10.1515/jqas-2021-0059/html

The study by Guan, Cao and Swartz concludes that teams sit deeper when leading and this does not help suppress chances (more shots conceded).

Is Tiger Woods Loss Averse? Persistent Bias in the Face of Experience, Competition, and High Stakes
By Devin G. Pope and Maurice E. Schweitzer https://repository.upenn.edu/server/api/core/bitstreams/db1d8c8b-b462-4150-93a7-c6c061aae605/content

These authors demonstrated that professional golfers putt better for par than they do for birdie/eagle. They considered an extensive array of possible factors that could otherwise explain why professionals are better at putting for par than for birdie. None of these factors were able to explain much of the difference between the success rate of birdie putts vs. par putts, leading to the conclusion that golfers are simply more focused when trying to avoid dropping a stroke (putting for par) than when trying to gain a stroke (putting for birdie).

These studies include some professional work beyond my own skills. However, I think I can add to the study of this area as I have worked on predictive team ratings for almost a decade and game states have been a particular area of interest for me.  I do not want game state effects to weaken the predictive nature of my ratings so I have cycled through various methods to adjust for game state.   I think this experience puts me in a good position to consider various aspects of the question posted in this piece.

Part 1 – How does a leading team maximise their win chance?

When a team holds a one-goal lead, a draw in the remaining minutes will secure a victory.  As low scoring football games have a significantly higher draw chance it follows that leading teams benefit from a low scoring game.

A lot of this study has been focused on one-goal games (either team is leading by 1 goal) from the 70^th minute until the end of the match.  This late game window and scoreline is one of the most extreme game state scenarios and comes with appropriate sample sizes.

To show that a component of success for a leading team is low scoring totals we can use the Poisson distribution. 

Firstly, let’s see how the win % of leading teams changes if we hold the expected goal (xG) ratio at 0.5 (an even share of chances) and lower the total match xG.

Figure 1

Play after the 70^th minute accounts for around 25% of the match.  The first few rows of this table are realistic, the last few less so.

Figure 1 shows if a team leading by 1 goal concedes 0.5xG and creates 0.5xG after taking the lead, they will win the match about 75.6% of the time.  However, if they could control the game such that they only create 0.3 xG and only concede 0.3 xG this win% grows to 83.0%.  The expected points growth for these same numbers (2.48 > 2.63) is worth over 2 points a season if the team finds themselves in a similar situation 15 times a season.

If a team creates more xG than they concede, this also helps them win the game:

Figure 2

For the first 5 xG scores I fixed the total goals in the game and increased the goal ratio of the leading team.  For the second five I increased the xG differential while keeping xG conceded level. 

As you might expect, adding offensive xG while leading is less valuable than reducing xG conceded.   

I included this part to emphasise a team’s objectives when leading by a goal.  As a draw is a success, a positive xG differential/strong xG ratio is only 1 component of success – a low scoring match is another.  It’s not correct to say a team is performing worse while ahead just because of a weaker xG differential, we should look at overall xG suppression as well. 

Let’s now examine how real teams are playing at tied, and how they are playing at +1, for the whole game and after the 70^th minute.

Part 2 – Game state analysis across 10 European leagues (total scoring rates)

I have analysed 10 leagues (the 2025/26 seasons of the Premier League/La Liga/Serie A/Bundesliga/Ligue 1/Eredivisie/Bundesliga II/Scottish League/Super Lig and the Primeira Liga) with data sourced from whoscored.com (powered by Opta)

The analysis in this part is split up by team ability.  Having experimented with various angles before this write-up I felt there was more to be learned by this categorisation than others.  I am using my own xG model that incorporates 33 zones and 7 shot types. 

Figure 3

This table shows the total goals per game (the numbers are standardised to represent goals per 90-minute game) in games between tied teams.  Owing to the symmetrical nature of a tied game there are only 4 different columns of data here (as every minute played at tied for a big favourite is a minute played at tied for a big underdog).

Figure 4

This graph displays the data from figure 3.
The error bars are 95% confidence intervals.
All data points represent the total goals per game per 90 minutes of play.

Each data series has a minimum at the centre and maximums at the edge.  This represents the fact that one team being stronger than the other is a component of increased goalscoring in football (one set of players is stronger/faster than the other).  Also note how much higher scoring games are from the 70^th minute onwards.  I’m confident this is a psychological effect that sees players start the game in a more relaxed fashion (a team has no strong motivation to take an early lead) before finding more motivation and urgency/intensity as the game nears its finale. 

The differences in xG and goals can be related to a specific xG metric.  The xG metric I created uses 7 shot types and 33 zones.  It does not include a big-chance adjustment for this study.  Across all minutes, figure 4 shows slightly higher figures for goals compared to xG.  This may be a small calibration issue, for example 100 xG with my xG metric may result in 102 goals.  If we look at data points for the >70^th minute however goals do have a margin over xG for all 4 bins used in this plot.  A fancier xG model may not have showed this effect.  This implies either the chances after the 70^th minute are bigger than what my xG model sees or players are better at finishing after the 70^th minute (most likely the former). 

Next, we can look at the same parameters but for play where one of the teams leads by 1 goal.

Figure 5

The data will not be symmetrical around the central column like figure 4 as a strong team leading by a goal does not create an underdog team leading by a goal.

Figure 6

The error bars are large for heavy underdog teams leading by a goal as this is a rare occurrence, especially after the 70^th minute.

This plot illustrates plenty of differences to figure 4.  The 70^th minute onwards is still the higher scoring portion of the match, but the gap is less significant – in fact the error bars almost overlap at times.  Matches where the heavy favourite leads by a goal are not significantly higher scoring than when closely matched teams are playing a 1 goal game. 

Figure 7

Figure 8

Figure 7 made a comparison with ratios while figure 8 plots the change in difference in total GPG.  Positive y-numbers mean 1 goal games are higher scoring than tied games.

You may look at the black and pink series (goals and xG for all minutes) and conclude that one-goal games are intrinsically higher scoring than tied games.  This is arguably wrong as when we compare tied states vs. +1 states for ALL minutes, we are comparing different minutes of the game.  Tied states on average exist much earlier in games than 1 goal games, and, as shown in figure 4, football games are much higher scoring late on in games (even for tied, closely matched teams).  Comparing tied play in the 30^th minute to 1 goal games in the 70^th minute is comparing apples to oranges.

Play after the 70^th minute removes this issue as we focus in on the same portion of the match for both game states.  (I calculated that the average minute for 1 goal games after the 70^th minute is only 40 seconds later than the average minute for tied games after the 70^th minute.)  In turn we then run into sample size limitations.

The plot shows evidence that strongly favoured teams play lower scoring matches when they are 1 goal ahead than when tied.  As the team leading by 1 goal becomes less favoured the total match scoring increases. 

Figure 9 shows the sample sizes involved in these plots

Part 2 – Game state analysis across 10 European leagues (performance & field tilt)

So far, we have looked at the total scoring of teams during tied play and during 1 goal games.  1 goal games appear somewhat lower scoring for a fixed minute, but it depended on how favoured the leading team was pre-match.  Next, it’s relevant to consider the performance level of teams in these states (attacking xG/total match xG).

Figure 10

Better teams are better.

Figure 11

Similarly, strong favourites are much better when leading by a goal than strong underdogs.  However, the dominance levels have dropped relative to tied.  Using the log method I can compare these performance drops.

(I am using logs because the difference between 0.8 and 0.7 is more than the difference between 0.6 and 0.5.  An 80% share of play is 4x more of the play than the opponent, while at 70% share is only 2.3x as much.  Compare this to 0.6 and 0.5 which is 1.5x vs. 1x as much respectively.)

Figure 12

These are calculated as if the team’s play at +1 was matched up to their own play at tied.
The ratios are converted into logs using

The difference of the logs is then converted back into a ratio using

Performance drops are universal and more extreme after the 70^th minute.

xG ratio drops a similar amount for teams of nearly all abilities.  Underdog teams fare slightly better, but this is likely down to a) variance in the sample data and b) underdog teams are in a more similar position drawing or up a goal.  Underdog teams and their opponents (after the 70^th minute) may see both scenarios in a similar way (a draw is a good result as an underdog but not for closely matched teams or favoured teams).

I can use the Poisson distribution to calculate the win % change for tied play after the 70^th minute compared to +1 play after the 70^th minute.

Figure 13

I used xG ratios and total xG to calculate these figures.  For example, if a team’s play at tied (>=70^th minute) had an xG ratio of 0.6 and the game had a total xG per 90 of 4.0 this would be an xG score of 0.6 – 0.4 (>=70^th minute).  If at plus one they had a ratio of 0.4 and total xG per 90 of 3.5 this would be an xG score of 0.3 – 0.525.  The success rate is then the draw % plus the win % of these xG scores computed using the poisson distribution. 

The lower scoring nature of some play at plus 1 is not enough to counter the significant degradation of performance for teams in every category.  This means teams leading by a goal suffer worse outcomes than if their share of play at tied was replicated.

Those familiar with modelling games using the Poisson distribution should not be too surprised with these results as this is the main reason why draws are more common than the distribution predicts (all games are biased to return to a drawn scoreline).

Very dominant teams see the smallest reduction in success rate owing to their games becoming significantly lower scoring.  It may be that the strongest teams are playing the situation of a 1 goal lead the smartest but the findings in could also be due to a psychological component.  It’s possible big underdogs don’t have the belief they can equalise while when the weaker teams lead, they face a particularly strongly motivated opposition (desperate to avoid a shock loss). Struggling to separate psychological/more subconscious factors to deliberate tactics is a common theme in this study.   

Field Tilt

If we want to answer the questions of ‘should teams park the bus?’, we need to relate performance to field tilt.  A defensive field tilt (many more passes conceded in the defensive final third than attained in the attacking final third) is approximately what people mean by ‘parking the bus’.  Another factor to consider is how lower field tilt can also result from ‘bad play’.  E.g. a team is getting dominated, getting tired.    

If a team has a ‘tilt’ above 0.5 that means they are passing more from the attacking third than their defensive third.

Figure 14

In a similar vein to figure 10, good teams play more in offensive areas.  The better teams see a more attacking field tilt the longer the game remains at tied.

Figure 15

Comparing the difference using the previously described log method:

Figure 16

Relative to their play at tied, all teams see a similar field tilt change towards the end of games.  However, less favoured teams start the change earlier.  The most strongly favoured teams only changed their play style significantly from the 80^th minute on.  Underdog teams see a strong field tilt change as soon as they score, even if it’s in the first 15 minutes of the game.

Figure 17

Performance at +1 matched up to performance at tied (0.5 = play is the same, <0.5 means play at +1 is worse)

The xG ratio changes do not strongly match up with field tilt changes from figure 16.  While xG ratio deteriorates for all teams through the game (at +1 compared to tied), it does not correlate with the early style changes from underdog teams.  The xG ratio change is more strongly correlated with minute than it is field tilt. 

How do strong teams lower scoring

I found a strong correlation between total scoring and a passing metric.  The following compares the total final third passes in the match at +1 to the total final third passes at tied.  A lower total therefore implies more play is taking place in the middle of the pitch.

Figure 18 – data points below zero mean there is less play in either final third in 1 goal games vs. tied games

This is quite intuitive; more play further away from the goals = less goals.  When they lead the strong teams can control the match better in the middle of the pitch and the ball spends less time in final third.  Weaker teams lose some control, and we see more end-to-end action when they lead by 1 goal than when tied (the field tilt changes I showed previously illustrates this is not because they are defending deeper).  

We have seen teams leading by a goal suffer quite heavy performance degradation (by xG and goals) regardless of how favoured they were to win.  This is fairly in line with significant field tilt changes through the game.  Leading teams that are at least equal in strength to their opponents do play lower scoring games when leading by a goal (which helps win % on its own).  I showed pass numbers that demonstrate this is probably because they can control the ball better in the middle of the pitch.

Overall, however, teams of all abilities would have more success if they could replicate their play at tied.  Does this mean their tactics when they lead by a goal must be flawed? 

Part 4 – Loss aversion

Loss aversion is an important aspect of both life and sports.  Daniel Kahneman and Amos Tversky were key researchers in the late 20^th century credited with the formalisation of the idea.  They calculated that humans are approximately twice as motivated to avoid a loss than achieve a win.  This is key for losing game states as losing teams will naturally be able to find more motivation than the protecting/leading team. 

It is a challenge assessing how to quantify this in football terms.  We have seen in the last part that leading teams suffer a significant performance hit but how do we know whether to assign this to loss aversion or the tactics they employ.  Without being able to separate the two it is hard to answer the title of this study.

I had an idea to demonstrate a clearer example of loss aversion by looking at tied play.  I think we can see a similar effect to 1 goal games in tied games specifically when strong teams are playing weaker opposition.  In said situation I think strong teams will see the draw as a loss (2 points lost) and experience loss aversion as full time starts to approach.

This means we can look for a loss aversion effect by comparing how strong favoured teams play at tied compared to their less favoured counterparts.

Firstly, the following shows play at tied between teams that are closely matched:

Figure 19

Figure 20

‘Closely rated teams’ are teams that are rated within 0.2 goals per game of each other.

The 10 leagues used for analysis are the same as in the previous part (over 3,000 games).   

The table shows matches continually becoming higher scoring the closer we get to full time.  This is constrained for game state and constrained for similarly matched teams so is an accurate barometer of how goals increase through football games.  We could have hypothesised that closely related teams may be happy to settle for a draw late on in games, but this is not borne out by the data. 

Figure 21

The numbers are relative to the 90-minute average GPG for each category (figure 19).  The (avg GPG) row at the bottom shows the averages.  For example, 1.54 on the xG table (left) means big mismatches (one of the teams is expected to win by at least 1.4 goals) are 54% higher scoring after the 90^th minute than the game average. 

I am going to treat the numbers between closely matched teams as our ‘control group’.  The idea is neither team sees a draw as a loss so no loss aversion is present for these games. 

Comparing the xG and goal numbers against this control group gives:

Figure 22 – Does this figure demonstrate loss aversion?

The data shows teams that heavily favoured teams play higher scoring games as the clock nears full time.

We also should consider the performance level of the favoured team:

Figure 23

The numbers are standardised such that we could say if a dominant team (expected to win the game by at least 1.4 goals) in a tied game state after the 90^th minute played itself between the 60^th and 70^th minutes (also at tied) said team would achieve a 60% xG ratio (0.64 matched up against 0.53).  The 0.85 value for goals may be related to a small sample size.  This figure comes about from a goal difference for this data point of 20 – 1.  This 0.952 ratio is then standardised against an average performance ratio of these teams when tied of 0.800.  Using a logit transformation method this results in the 0.85 seen in the table. 

This shows more favoured teams increase their performance level very late in games, increase the goal conversion rates (compare xG and goals in the previous figures) and play higher scoring matches. 

Lastly, it’s useful to consider the field tilt changes as well:

Figure 24

The field tilt changes more through a game between increasingly mismatched teams.

I think all these tables represent loss aversion because each column represents teams with extra reason to view a draw as a loss (a single point is nearly always a bad result if you are so favoured to win the game).

In the first section, I was focused on comparing play after the 70^th minute in one-goal games to the same time frame in tied games. Using the data above I calculate the following ‘loss aversion’ factors for the 70^th minute onwards:

Figure 25 – Heavily favoured teams find this much extra performance after the 70^th minute (I averaged results for 70^th minute onwards from figures 22, 23 and 24) .  These numbers compared to their own performance for the entirety of the match and relative to how these metrics changed throughout games between closely rated teams.

I believe these values go some way to quantifying loss aversion as they denote the extra performance a team finds when they begin to fear dropping points.

An issue with this method for assessing loss aversion may be that the lower ranked team does not have the same motivation for a draw for each level of mismatch.  This is because teams that go in as heavy underdogs may be more motivated to protect a draw than closely matched teams (heavy underdogs see a draw as a greater success).  However, seeing as I used the word ‘protect’ here, the underdog team is not in loss aversion, unlike the favoured drawing team that feel a draw to be a loss.     

It now becomes more apparent that strongly favoured teams (figure 8) lower the total scoring of a match in a comparison that is not equivalent to the less strongly favoured.  After the 70^th minute when tied, these teams will feel loss aversion because a draw is a bad result.  This increases the total scoring at tied which results in 1 goal games being relatively lower scoring.

Considering how teams of differing ability profile the same in terms of pass ratio, xG ratio and (with extra adjustments) total GPG as well I am not sure separating teams by ability reveals as much extra information as I initially thought it might.

I want to compare the values from figure 25 to earlier figures 10, 16 and 17 that analysed the play of teams that lead by a goal late on in games.  Averaging the numbers across teams of all abilities gives the following stats for trailing teams:

Figure 26

xG and goal totals are close to 1 which means 1 goal games after the 70^th minute are neither higher or lower scoring than the equivalent period of tied games.  However, given  1 of the teams will be experiencing a strong loss aversion, the fact total goal rate does not increase is important.  Loss aversion numbers from figure 25 shows we would expect an increase of around 10%.  This makes sense as we would expect a more end-to-end game – the trailing team will often leave space in defence in a desperate attempt to find an equaliser.  The tactics of the protecting team must be suppressing total goals/xG. 

xG ratio and goal ratio see a large increase for the trailing team but some of this is because of improved performance from the trailing team owing to the loss aversion they experience.  The goal ratio figure from figure 25 is subject to some uncertainty.

The field tilt/pass ratio of trailing teams is far beyond the loss aversion figure.  This implies the protecting team is making a deliberate effort to defend deep as loss aversion alone is unlikely to create such a large change in field tilt.

Figure 27

If we take the xG ratio the win rate of teams drops by ~4.4% but if we take goal ratio as more accurate it drops a miniscule ~0.3%.  The calculation here is effective for a team that takes the lead in the 70^th minute.  I have not found evidence that parking the bus is a powerful and beneficial tactic but there is evidence here that the suppression of total goals (increases draw % for the remaining portion) and the relatively small decrease in performance make it a viable strategy.

Part 5 – assorted analysis angles

League by league

It could be revealing to compare team’s play in different countries to see if the philosophy changes.

Figure 28

The big 5 leagues all seem very similar play between teams.  Some other leagues see teams appearing to play slightly less defensively when leading after the 70^th minute.  I found a weak correlation that this did result in an inability to suppress goal totals but this correlation was solely down to the Scottish data point (1 goal games after the 70^th minute in Scotland were high scoring).  

The consistency across the leagues (especially the big 5 leagues) is evidence player psychology could be as much a driving factor for field tilt change as managerial choices (otherwise we might see more differing philosophies between leagues).

Team by team

I analysed team by team to see if teams that saw smaller/larger field tilt changes saw differing performance figures (again after the 70^th minute and comparing tied play to +1 play).  I found a correlation that a less defensive field tilt saw better performance (not a big surprise as an attacking field tilt is also an indicator of playing well) and a weak correlation that a less defensive field tilt saw high scoring games (bad).  I decided to stop pursuing this line of analysis as the error bar of each data point was too high to work with.  This could be mitigated in the future by a very large sample of teams.

General team play

I correlated field tilt for each team to overall ability to see which teams are unexpectedly defensive or attacking. 

I found a weak correlation (r^2 of 0.02-0.05) that the more defensive teams played lower scoring matches. 

It could be interesting to use player values to see if the teams that are more defensive/attacking (in terms of field tilt) than would be expected for their level have higher or lower than expected performance levels given their player values.  I hypothesise playing defensively may be a smarter option but maybe that is biased by high profile teams like Atletico Madrid over the last decade.

Part 6 – Conclusions

I have shown how play changes between drawn games and 1 goal games. 

A common theme of the study has been psychology.  A lot of the results we have observed for the different game states for teams of different ability can be explained by well-known psychological effects such as loss aversion.

I have found limited evidence of managerial tactics as I believe psychological dynamics explain most differences between play at different score-lines.

I have not found evidence that parking the bus helps teams achieve victories, but I also showed that it is very close to not hurting the performance either.  Quantifying the loss aversion effect is a challenge so this conclusion could be subject to change in the future.

It could be naïve to look at data and say what players should/shouldn’t do.  Telling players not to sit deeper with a 1 goal lead could force unnatural behaviours that harm performance.  

Managers should be flexible, consider their team’s strength and learn from past results.  If a team is strong at defending crosses (most strong defences are) then intentionally sitting deeper could be very valid (crosses are more frequent in this situation).  This is something that could change on a team-by-team basis.  Managers should be familiar with loss aversion and incorporate it into tactical decisions.  For example, how motivated are we to take an early lead?

If you have enjoyed this study, please consider subscribing to my blog using the white box on the top left of the main page at syzygyanalytics.co.uk.  If you have any ideas, suggestions, further alternative methodologies or criticisms of what I have written please get in touch!  You can message me on x.com/samh112358 or syzygyanalytics.bsky.social

References

1. Guan, C., Cao, J., & Swartz, T. (2021). Parking the bus. Journal of Quantitative Analysis in Sports, 17(3). Link
2. Ruiz-Menarguez, M., & Badiella, L. (2026). The Impact of Formations on Football Matches Using Double Machine Learning: Is it worth parking the bus? arXiv pre-print. Link
3. Is Tiger Woods Loss Averse? Persistent Bias in the Face of Experience, Competition, and High Stakes By Devin G. Pope and Maurice E. Schweitzer* Link
4. Opta match data via WhoScored.com

The 2026/27 season is almost here so, just in time, I have some 2026/27 predictions.

I’ve analysed the last 20 games of 2025/26 using a) an xG model (I call it the “0.25 0.25” xG model, the first 0.25 refers to all chances having a maximum possible value of 0.25 and the 2^nd refers to a bonus xG value for a goal being scored) and b) market inferred ratings from the closing betting lines.

I added some adjustments for the team’s motivation in those last 20 games, compared how much European distraction they had last season vs. this season and taken an educated guess at the positive or negative influence of player transfers. 

Arsenal lead the way in my predictions, but Manchester City follow closely behind.  I think Arsenal are a known quantity, and 80.6 points is quite a standard prediction.  On the other hand, I am more positive about the chances of Manchester City than most people I did not measure a massive impact for Rodri last season when comparing City’s performances with him on versus off the pitch; plus, Andersen looks like an excellent replacement.  Furthermore, I am not overly fussed about Guardiola leaving.  I have posted studies on the impact of managers changes before on this page and could not find evidence in the data that these changes had an important impact.  Admittedly, those studies were for mid-season manager changes at teams more likely to cycle average managers.  However, without any empirical data to demonstrate Guardiola’s ability as a manger, I remain sceptical.  I have not factored in any potential punishment for Manchester City’s ‘115 charges’ case, though looking at their relegation odds of 12/1 it looks like we should.

Chelsea and Liverpool are closely matched in 3^rd and 4^th but sit quite a long way behind the top two.  Without putting weight on market-inferred ratings from last season, Liverpool would be even lower, as I did not rate their performances very highly.  Unless some big transfer activity happens (Barcola?), I don’t see a reason to be too bullish on them.  Chelsea have made plenty of signings (Rogers and Lacroix being the biggest) and may well be poised to take advantage of being the best Premier League team with no European competition.   

I have Manchester United as underdogs for the top four.  I think Casemiro is an important loss, while Sancho and Tielemans don’t move the needle much.  United fans may end up fairly disappointed once again.

At first, I was surprised to see Tottenham at 1.2 (1/5, -500) for a top half finish, but I have decided they are indeed likely to make a big jump.  Van Hecke and Tonali are likely to make a big difference and no Europe should help them have a less chaotic season than the previous two.

Next comes the Brighton andBrentford pair.  Owned by sports bettors, these clubs may well have by far the best analytics in football.  Brentford have made one key signing – Sangare – who may have been a big factor in Lens making Ligue 1 a title race last season.  I have both teams pushing 80% to fit into the top 10.

Newcastle have lost a lot of key players, but I think they can do better than many expect.  Nearly everything went wrong for them last season, including plenty of injuries, yet they still were far from bad.  I think they will surprise if the new signings have a good impact.

I think Villa’s key departures means they will struggle to meet their standards of last season.

Despite liking Leeds last season, I am not on their side this time (relative to betting odds).  Generally, I’m not convinced goalkeepers are one of the most important position. Even if that’s wrong, it is incredibly hard to rate a goalkeeper accurately enough to call a goalkeeping signing ‘key’.  In a comprehensive study I’m working on about ‘parking the bus’, I noticed Leeds standing out for some reckless and poor play while leading games. 

Palace/Forest/Fulham are all fairly close, but there’s plenty of scope to be wrong here.  Forest feel like the most likely of the three to have the highest ceiling and Palace feel like they have a higher ceiling than Fulham.

Sunderland occupy 17^th in my rankings so it will be interesting to see how they follow up their brilliant 7^th place from last season. 

I rated Ipswich and Coventry quite equally in the Championship last season.  I prefer Ipswich’s signings plus they have a recent season of experience in the PL so they are my pick to finish higher.  I rated Hull like a bottom half team last season so it’s no secret they might struggle.  The 22.7 pt total I have above gives them about a 77% chance of finishing rock bottom.  Given the strength in depth of the Premier League it will not be a massive surprise if they flirt with Derby’s 11pt ‘record’.

That’s all on the PL for now.  I have been working a lot on a large  study on game state tactics and ‘parking the bus’ over the last few weeks so look out for that one coming soon!

Let’s review the performances & MIRs (betting market inferred ratings) of Spain and Argentina this World Cup and use them to preview one of the biggest games in football.

Figure 1

The first two rows show the performances of each 2 finalists for each of their World Cup games so far.  It’s expressed as a ratio (adjusted xg for divided by total xg in match) and is schedule and game state adjusted (I see a larger game state effect in the World Cup than in any domestic league.)  The average value at the end doesn’t include performances in their 3^rd group game because of questions over motivation.

The 2^nd group of rows (GPG) just shows the total adjusted xG in each game.  As I’m using ratios, I think it’s relevant to show that the xG score wasn’t something stupidly low like 0.3-0.1 (as this looks very strong for the 0.3 in ratio terms).

The 3^rd group of rows show market inferred ratings, i.e. what the expected goal ratio that the betting markets implied at kick-off.  I use 3 decimal places here and 2 for the other categories because of the accuracy I think each measure holds.  These figures are schedule adjusted so how I rate each team is a potential source of error.  The average only uses MIRs from the last 3 knockout games as I think that is a substantial sample for this measure (and the more recent the data, the better.)   

Spain’s performances have been relentlessly strong (apart from Uruguay where they had limited motivation).  Argentina have also been strong but not quite to Spain’s level.  It’s an appropriate Final as I rank them the 1^st and 3^rd best teams of the tournament (France 2^nd).

Argentina posted very consistent MIRs through the tournament until the semi-final in England.  In my semi-final preview I showed how I rated Argentina favourite for the game and was slightly conservative in predicting closing odds that rated both teams equally.  The market did drift out a few ticks (which gave me a moderately successful pre-game trade) but I still don’t quite understand why Argentina’s MIR for that game seems so out of line.  Spain’s MIR did not actually clear Argentina’s but has trended upwards and will continue to do so after their performance against France.

Figure 2

The first columns denote the weight put on MIRs and performances as methods to rate each team. These differing weight correspond to the implied match odds you can see on the right hand side. I’ve fixed the match GPG at 2.2 which corresponds with the current market line (as I haven’t focused on predicting total goals).

I generally feel 60% MIR – 40% performances is an appropriate blend and that matches with the current betting market.  The more weight I put on performance the more favoured Spain become.  The MIRs doubt they can really be this good and regress their level somewhat (a 5-game sample is no guarantee of anything).  I can see a case siding with Spain, owing to their consistent performance level.  It is, however, very debateable whether ‘their performances are consistent’ is a valid reason to put more weight on the average performance level.

Predicted closing line 2.32 (Spain) – 3.98 – 3.125 (Argentina)

This is it for the World Cup but please subscribe to this site using the white box on the home page at (syzygyanalytics.co.uk) to be notified whenever I post.  I am planning my own version of this paper “Should teams park the bus?”! https://www.sfu.ca/~tswartz/papers/bus.pdf

Introduction

When considering why I predicted the closing line wrong for Mexico/England and Spain/Belgium it dawned on me I *may* have had a flaw in my approach to football betting for the last 9 years.  I’ve put a lot of work into developing my ratings with numerous adjustments, but it was the conversion of the ratings into match odds that I have been somewhat overlooking (which relates to the format of the rating).

I was using team ratings in terms of (adjusted) xG difference.  If team A had average scorelines of 1.4-0.6 and team B had average scorelines of 2.0-1.2 I’d view both teams as equal strength given their shared goal supremacy of 0.8.  I’d work out the goal line (total goals expected in the match) separately.  I gravitated to ratings in terms of (adjusted) xG difference because it seemed convenient to work with.  Schedule adjustments, game state adjustments, pricing up a match between two teams, it all worked well.

The problem is there may be a flaw as 2 teams posting similar xG differences may not always be of the same skill level.  Team A scores 70% of the goals in their games while team B only scores 62.5% of them (Team A’s matches are much lower scoring).  If, instead of expecting an average supremacy from team A of 0.8 goals per game we expect them to score (on average) 70% of the goals this is a dramatic switch-up of approach.

Consider pitting team A against team B.  We’ll say they’re both from a league that has an average of 2.6 goals per game.  From team A’s perspective, team B will add 0.6 goals to this game compared to team A’s average game.  Using my original method, these 0.6 goals will be distributed equally to team A and team B.  This means we rate both teams equally and treat the higher game total as separate.

However, now consider team ratings as the ratio of goals scored.  Team A is rated better (0.7 vs. 0.625).  This means when we match them up, team A will be the favourite.  They’ll be even more favoured when they are awarded more than a fair share of the 0.6 goals added to the game as well. 

So, is team A the better team in this scenario?  I think we can say they are definitely *some* amount better.  An average scoreline of 1.4-0.6 is worth almost 3 expected points more than an average scoreline of 2-1.2 over a 38-game season.  Their defence is twice as good, but their attack is more than half as good.  Their games are not so low scoring that they will suffer too many draws.

I need to put more effort into thinking about this.  Is the expected scoreline for this game 1.3-1.3, is it 1.53-1.07 (using ratios ) or it is somewhere between the two?  Do teams who score 70% of the goals in their low scoring games still score 70% of the goals in higher scoring games?      

Today I’ll use ratios to preview the two semi-finals. 

France vs. Spain

Figure 1 (‘Perf1’ ‘Perf2’ ‘Perf3’ in the top table should say Last 32, Last 16, Quarter-Final)

All performance and MIR (market inferred performance) figures are schedule adjusted.

For the 2^nd table I only used the last 2 games for MIR and excluded the 3^rd game for performance (motivation questions in that game for both teams).

The more we weight towards performance the further things swing towards Spain -Spain’s performance have bene consistently strong.  The markets opened with France around 2.4 and they have now drifted out to 2.63.  France are coming off a very strong performance vs. Morocco.  My draw price is a bit off; I will work on that.

We’re quite close to kick off now as I write this so I’m not going to predict a move from now as between 50% and 75% MIR weight seems reasonable.

Closing line 2.62-3.325-3.175

Argentina vs. England

Figure 2 (‘Perf1’ ‘Perf2’ ‘Perf3’ in the top table should say Last 32, Last 16, Quarter-Final)

Ok I have Argentina as the favourite here across all MIR/Performance blends.

As I write at 6:18pm UK time on Tuesday the market is 2.81 – 3.025 – 3.175

I’m going to predict a closing line of 3.0-3.0-3.0!  If I’m wrong, I’ll hopefully learn why.

Thanks for reading, subscribe using the white box on my home page to be notified of any posts.

I am writing previews for the last 3 quarter finals.  First up, Spain vs. Belgium then the next two will follow very soon!  I will review the betting odds implied levels of the teams and as well as World Cup performances so far.  I want to compare what I think the betting line should be to its current number.

Figure 1

The MIRs (market-inferred team rating) contain various adjustments.  For example, for Spain’s RD3 game against Uruguay, Spain were not strongly motivated as they had already won the group.  This will result in the MIR pushing down Spain’s overall rating when it is not deserved.  This can be corrected by adding some small amount.  Also, I improved the schedule adjustment for each team by varying the MIR rating of their opponents for each game.  For example, Portugal’s lacklustre performances reduced their rating as the tournament went on so Spain are considered to play an opponent with a rating of +0.94 instead of, for example, a rating of +1.16 had they played Portugal in their opening match.

These adjustments help us make a more direct comparison of how the global sports markets rate each team.  These MIRs will be very accurate as it’s only professional syndicates placing very large bets that significantly move the markets.  We can see Spain are rated consistently highly across the tournament.  It appears there may be a small drop off since the first two games but as they were so strongly favoured in those games (against Cape Verde and Saudi Arabia) it stresses various parts of the process (e.g., are they more favoured against these teams because there’s less likely to be a game state effect where they hold on for a narrow victory).  Similarly for their round 3 game against Uruguay, it’s hard to quantify the impact of reduced motivation.

Spain’s MIR for the 2 knockout ties so far is +1.58.  This means they would be expected to win by an average of 1.58 goals against an average world cup team (24^th best team of 48).  Spain’s performances have been consistently solid, between 1.5-2.0 goals better than their opponents each time (except for Uruguay, which has been rated upwards but even so, I think I should discount that match).  Spain have not spent a minute losing so far and plenty winning which results in a game state adjustment of around +0.27.  I am using a simple game state adjustment where I expect leading teams to underperform by about 0.55 goals per game.  I am intentionally using one decimal place for performances and 2 for MIRs as I think it represents how much information each rating system contains (over these sample sizes).

Spain’s strong performances push their overall rating into the 1.67-1.83 region.

Belgium have had a bumpier ride.  They were very ordinary against Egypt initially but bounced back with stronger performances against Iran and New Zealand.  This pumped up their rating, and they went off even-money shots to win in normal time against a strong Senegal side.  A poor performance (and an unlikely win) followed, resulted in a much lower MIR against the USA.  The performance against the USA was a return to perform suggesting Belgium’s MIR will rise once more.

I don’t think there are any significant injuries or suspensions.  This is entirely relatve – we’ve trained our ratings purely on the World Cup so far so it’s only of interest if a player who has been missing is now returning or vice versa.  Amadou Onana picked up an unfortunate potentially long-term injury against the USA but had not featured heavily this World Cup regardless.  Lamine Yamal missed minutes in Spain’s group games but similarly Doku (a player I rate highly) and De Bruyne have both missed games and could be used as fresh options for this game.  I don’t think the potential lineups are a mitigating factor. 

My ratings here show Spain as favoured by between 0.63 – 0.67 goals.  Currently at 11am BST on Friday the market shows Spain as 0.85 – 0.95 goal favourites.  I’m not sure how to explain the difference so I am siding with Belgium here.

Edit: I think there’s a problem with my approach properly respecting differing goal lines and total goals per game. Spain are posting these numbers on a lower total goal count than Belgium. I’ll work on that

Last 16 remaining previews

I had a good day yesterday predicting the correct direction for price moves in both games.  (Both prices moved a bit more than I expected though.)

Let’s look ahead to more Last 16 games.  A lot of betting lines have already moved but there is still potential for more movement.  Check the previous previews for more information on the graphics shared in this post. 

Brazil vs. Norway

Brazil put up good numbers in a 1 goal win against Japan.  Norway’s performances really haven’t been all that special except for the game against France.  That’s an odd one as it gets up-rated quite significantly because Norway played a weakened team for that game.  It seems like a negative that their performance rating is propped up by that game.

Brazil, meanwhile, has lineup concerns.  Raphinha won’t be fully fit, although he has been injured for the last 2.5 games anyway.  Casemiro is a small doubt and Paqueta is also out.  Norway’s performances want to push me more towards the 1 goal fav side of things, but Brazil’s injuries pull me back.  Brazil are currently 0.75-0.8 goal favourites which I agree with for the moment.

Predicted money line:   1.88 (Brazil) – 3.8 – 4.8 (Norway)

Mexico vs. England

This is an interesting one.  England are currently rated as 0.2 goal favourites which is a long way off what the MIRs and performances would imply.  Of course, the home advantage and altitude of the game is a huge deal, but it’s already included in all the above data – Mexico have played ¾ of their games so far in Mexico City (and the other at 1500m in Guadalajara).  England have some health concerns – Declan Rice has been playing through a lot of pain and Reece James’ world cup is over.  However, there is a chance Saka could make his first start of the tournament.

Mexico have spent a lot of time winning but that’s represented by their large GS adjustment.  I just can’t get all the way from the data above to England only being favoured by 0.2 goals.  I’ll be aggressive here:

Predicted closing price:  2.28 (England) – 3.15 – 4.1 (Mexico) 

It’s unlikely the line will move this much but we will see.

Portugal vs. Spain

Another interesting dilemma as the two rating methods say different things.  Portugal are rated like a strong team but except for against Uzbekistan they really haven’t played like a strong team.  There is of course a massive question over whether Christiano Ronaldo is still good enough to play for Portugal.  Spain are currently around 0.55 goal favourites which requires a high weight on the performances of Portugal so far and a low weight on the betting markets.  If Ronaldo didn’t start, I think Spain would be worth imposing.  I can’t put complete stock in Portugal being as poor as some of their performance, so I’ll respect the market rating of Portugal so far and predict:

4.04 (Portugal) – 3.75 – 2.06 (Spain)

USA vs. Belgium

Belgium closed an even money shot against Senegal in the last 32, something many were surprised to see.  This was in part because of strong performances against Iran and New Zealand.  They were very ordinary against Egypt and Senegal so it feels tricky to gauge which Belgium will show up.  The USA miss Balogun after his controversial red card.  The markets certainly like Belgium and I think they will again. 

Predicted closing line:

2.9 (USA)– 3.55 – 2.68 (Belgium)

Argentina vs. Egypt

I ‘m pretty in line here with the ~1.4 goal current line market line.  Argentina have spent a massive amount of minutes winning so far but are coming off a difficult game against Cape Verde.

Predicted closing line:  1.405 – 5.15 – 11.25  (no change)

Switzerland vs. Colombia

No disagreement here either, I think around 60% weight on MIR is about right which puts me right in line with the betting market. 

Predicted closing line: 3.925- 2.31 – 3.225

Today I have 2 match previews for the last 16.  I will post the rest in due course.  I am previewing the games trying to predict the betting odds at kick-off as I don’t currently have a strong belief that bets can be good if they are not ahead of the closing line.

Canada vs. Morocco

Figure 1

MIR = Market Implied (team) Rating

GS ADJ = Game state adjustment (I’m seeing quite a large game state effect in the World Cup.  Leading teams so far have performed at a level of around +0.2 adjusted xG per game which is massively offset by leading teams being significantly better to the tune of around +0.8 adjusted xG per game.  This makes performance in leading game states around -0.6 xG per game.  In domestic leagues the largest game state effect I’ve seen is around -0.4xG or -0.5xG per game.)

BC ADJ = Big chance adjustment (An adjustment for whether teams have logged an unexpectedly good or bad big chance differential relative to their underlying rating).

HOME ADV CORRECTED = As Canada have played 3 of their first 4 matches in Canada and this match will be played in Houston, Canada get downrated by 0.75 (3 of their first 4) x 0.3 (estimated home advantage for games in Canada) = 0.225 (see Home Adv Corrected) column.

My player-based team ratings (from my World Cup preview) are now ditched and I’m using a combination of previous market closing lines and performance ratings for the 4 World Cup games each team has played thus far.  Previous market closing lines are adjusted if I think teams had differing motivations or a team began to under/over perform expectations enough across the World Cup such that the rating of each opponent is not correct.  For example, in this case Morocco’s MIR vs. Haiti (RD3) is rated down a small amount because of waning motivation for Haiti.  (As an example for the second point, Paraguay now appears to have been rated strongly against the USA but that’s because the rating of the USA has since risen.  In that case Paraguay’s market inferred rating against the USA would be rated down.)     

Both these teams have impressed this tournament – you can see by the MIR values increasing each round.  Canada’s is increasing it just appears to have decreased in the last 32 because they lost their home advantage!

Whatever weight of MIRs and performance I use here doesn’t matter – I’m seeing Morocco as ~0.75 favs.  Chadi Riad is a major doubt which is a small negative for Morocco, and they also have had 1 day less rest and played for 120 minutes against the Netherlands.  Alphonso Davies might be fit enough to start for Canada which really could be a boost to Canada’s hopes.    

The market has been up and down a lot but now sits with Morocco around 0.8 goal favourites.  12 hours ago, Morocco was closer to 0.65 (if the Morocco -0.75 line is decimal odds of 2.2 that’s equivalent to being 0.75 goal favourites).  I think we might see Morocco drift out a little again before kick-off if my analysis is accurate enough.

My predicted closing odds are 5.5 – 3.575 – 1.86

Paraguay vs. France

Figure 2

Paraguay meets France in Philadelphia for what is expected to be a one-sided affair.  Paraguay went to penalties against Germany but have had 1 more day rest than France.  As you can see in these tables, Paraguay’s group stage performances were seemingly poor and their MIRs suffered as a result.  Interestingly, it’s going to be extremely hot in Philadelphia today which could be a useful confounding factor for Paraguay. 

Paraguay’s performance against Germany will have stopped their MIR slide although France looked electric against Sweden.  Diego Gomez returns from suspension, but Omar Alderete is injured.  France has lost Aurelien Tchouameni to injury.

The market is currently at around France 1.94 goal favs.  I have a slight interest in Paraguay because I think the heat might blunt France’s attack enough to make 3 or 4 goal victories less likely. 

My predicted closing odds for the 2-goal Asian handicap market is

Paraguay +2 1.83, France -2 2.2

Finally, the group stage has finished, one of the 495 possible permutations of 3^rd place finishers that could happen, has happened, and we arrive at a 32-team straight knockout.  Let’s preview the last 32!

• All ratings are in the unit of expected goals per game
• Betting market inferred ratings and 2026 WC performances are schedule adjusted to the best of my ability.
• Check out my World cup preview for a lot more information on the player-based team ratings

My odds vs. Market odds

Figure 1

Game by game

Brazil vs. Japan

I rate Brazil as 0.78 goal favourites here which is currently quite close to the global betting markets.  As you can see by the performance ratings, Brazil have been mediocre at times during this World Cup so far.  Brazil likely miss Raphinha (a player I rate very highly) while Ko Itakura will likely not be fit to start for Japan.  Raphinha played 1.5 games in the groups stages which influences all 3 constituent parts of my overall team rating (which have a 2x Betting Market inferred : 1x WC Performance : 1x Player Rating weighting).  Therefore, no Raphinha means Brazil should be downgraded a little bit.  Itakura is a loss for Japan also but not quite as significant.

I think a small lean to Japan can be justified.  Just as I’m writing this Brazil have drifted from 1.78 -> 1.81.  Just putting the finishing touches to this post and lineups are out and Brazil are now 1.84. Hopefully this is not the only correct lean I have in the post!

Germany vs. Paraguay

This is right in line with the betting market again for Germany vs. Paraguay.  Germany are strongly favoured here to secure a last 16 spot.   Paraguay’s group performances were not very impressive at all. 

Netherlands vs. Morocco

2 solid teams face off here in Mexico.  Morocco’ s performances have been at least on a par with the Netherlands so far, but other factors give Netherlands the edge here.  

Ivory Coast vs. Norway

There’s a bit of a disagreement with each rating system but overall, I think Norway as half goal favourites is reasonable. 

France vs. Sweden

It’s interesting how little France rotated in the group stage – not what you’d expect from a team with a squad like theirs!  Sweden has some players to be weary of (I have had a slight Sweden lean this tournament) so if I did lean a side it’s definitely Sweden.

Mexico vs. Ecuador

I have this as an equal matchup, which is very interesting.  Mexico are once again playing in Mexico City (over 2000 metres above sea level), something which is key to consider.  The first two parts of the rating already include said home advantage (markets obviously consider it, and on pitch performance boosted by the home advantage) but the player-based team rating is not.  If Mexico’s home advantage is worth 0.5 goals, then this adds ~0.13 on to their rating.  It could be argued Mexico’s home advantage is slightly lower in this game as Ecuador are used to playing games at high altitude (because of their own high-altitude stadium in Quito).

Why are the betting markets so much higher on Mexico?  If I’m wrong, I could be undervaluing just how much home advantage there is.  Even the ball flies and bounces different in the high altitude so maybe Mexico are able to enjoy a huge advantage in this way (an advantage that will keep increasing as they play more games in Mexico City).  My ratings favour Mexico by only 0.08 goals after the adjustment which does make me keen to get Ecuador on side we just have to be wary of how large the home advantage could really be.  I’m disappointed to read Ecuador are only travelling to Mexico City the day before…

England vs. DR Congo

England are strong favourites here, I agree with the market rating of 1.7 goals.  Going just be performances so far this tournament it could well be another grind for England.

Belgium vs. Senegal

An intriguing matchup takes place in Seattle on Wednesday.  I have it as a bit closer to a 50-50 match than the betting markets currently depict.  Belgium played fairly well in their final 2 group games and missed Jeremy Doku (a player I rate highly) for parts of that.  If my betting market inferred ratings are correct, I think the betting lines will move a little bit towards Senegal

USA vs. Bosnia & Herzegovina

The USA are rated very strong favourites in the betting market, and, like Mexico, I don’t have see them as quite such a strong favourite.  Home advantage is only worth around 0.1 goals on top of the overall team ratings shown here (as the player-based rating is the only rating to not already include the home advantage).  Pulisic will be fit which is very important for the USA so I only have a slight preference to Bosnia.

Spain vs. Austria

This is the last-32 match with the biggest divergence between my opinion and the current opinion of the global betting markets.  I have been low on Spain since pre-tournament (too many players who play for mid-table Spanish sides) and a bit high on Austria.  Lamine Yamal did not feature too heavily in the group stages which is something that is not properly factored in here.  My player ratings measure the difference between Yamal and Ferran Torres as about a quarter of a goal per game. Yamal only played half the minutes in the group stage meaning Spain’s performances were hurt by around 0.13 xG diff per game. As I’m 0.6 goals away from the betting line, this isn’t enough to explain the 1.5 goal line.

Portugal vs. Croatia

This is close to what the betting markets say but I can see a case for leaning towards Croatia when considering performance data only.

Switzerland vs. Algeria

There’s quite a satisfying parity between Algeria’s metrics.  Switzerland have been drifting but still look a bit over-valued by my calculations. 

Australia vs. Egypt

Another really close game!   Again I think it’s slightly closer to 50/50 than the market currently thinks.

Argentina vs. Cape Verde

Despite not being particularly high on Argentina this tournament I am surprised to see them as mere 2 goals favourites for this game.  Argentina rested and rotated a fair bit in their group games as well. A potential issue with my approach is lack of focus on the total goal lines. Is this match low scoring enough (both sides have low scoring tendencies for their level) to mean that Argentina are not 2.6 goal favourites after all.

Colombia vs. Ghana

 The rating components are quite inconsistent here.  I rate Ghana’s players quite highly but performances relative to Colombia leave a lot to be desired.  I could see a case for backing Colombia here as it is only my player ratings keeping my ratings as close as they are (just over 1 goal apart).

For a short time today every World Cup team has played 2 games.

These are my team ratings, currently using a 2:1:1 weight for the average of the 3 different metrics!

Introduction

Over the years I have focused on team performance metrics using passes/shots/goals to make predictive team ratings.  International football is a different beast to domestic football because nations play far fewer games than domestic teams.  On top of this schedule adjustments are much harder than domestic leagues as between world cups nations nearly only compete within their own continents. 

Towards the end of the most recent domestic season, I started developing some individual player ratings by measuring team performance when each player is on-pitch or off-pitch. I began to feel there was more potential in this method than I had previously realised.  This encouraged me to try and build individual-based team ratings for the World Cup. 

The following outlines my process in order to come up with team ratings and a simulation for the winner of the world cup.

If you are not interested in the method, you may wish to skip to ‘Team Ratings’.

Methodology

Domestic Team Strength

Take the 15 players most likely to feature for each nation in the world cup and calculate the average strength of the domestic teams they play for.  One of the challenges in calculating the strength of domestic teams is it requires knowing the strength of the domestic league they compete in.  For the leagues that have teams in the Champions League/Europa League I can assess league strength by the performances of teams from each league in these competitions.

Calculating league strength from the CL/EL

For each match in the CL&EL I use the following formula:

Expected home team performance = Domestic league rating of home team – Domestic league rating of away team + Home advantage +/- domestic league strength of the away team.  (All in units of expected goals per game)
The domestic league strength is the output we are working towards, but we also need it as input.  This is typical of the nature of schedule adjustments.  For this first iteration I used league ratings based on the average player market values of each league.

This adjustment is necessary as different leagues will have teams with skewed schedules.  E.g. strong leagues with teams progressing deep into the competition will on average play teams from stronger leagues than weaker leagues whose teams exit earlier.

The league rating implied for each game is then Observed performance – Expected home team performance.  Complete this for every match in the EL and CL will give us the league strength.  I then use these improved league ratings to do a 2^nd iteration of this calculation to get a final set of league ratings.

I can correlate these ratings to the average market value of the players in the league. This is shown below in figure 1.

I used the linear correlations displayed into the key in order to create the schedule adjustment for the 1st iteration as just described and also to estimate the league ability of all the leagues that do not play EL/CL.

Figure 1

Individual player impact

The average level of domestic teams that the national players compete for will give us a rudimentary ability level, but we can vastly improve it by considering the impact of each player on their domestic clubs. Some players may be far above or below the average level of the domestic club they play for. One of the most clearcut examples of why this would be necessary is Saudi Arabia.  Most of the team plays in the domestic league of their home country.  Some of the teams in that league have paid big money for foreign imports and therefore rate quite strongly on a global scale (by my reckoning they would not get relegated in the Premier League).  This means the Saudi Arabian players are generally far from being key players for their domestic sides.  If we rate the Saudi Arabian national team based only on the ability on the domestic sides the squad plays for, we will significantly overrate them.

I have a few approaches for calculating the player impacts.  Arguably the best/most unique one is what I briefly mentioned at the start.  I have calculated player impacts based on how their teams perform when they are on/off pitch.

This is fairly involved and includes a performance metric focused on removing noise and multiple adjustments to cover different circumstances for when players are on/off the pitch (for example average minute on/off, gs adj on/off, schedule faced on/off.)  I reduced the on/off comparison by a factor equal to the variance of the performance differences.  This variance will be large if either minutes on the pitch or minutes off the pitch is a small number.  I also weight the calculation more to attack or defence depending on whether attacking contributions imply the player’s position is based on attack or defence. An adjustment I wanted that has proved extremely challenging is considering how strong the rest of the team was when the player is on/off the pitch. For example, a player may coincidentally miss the same games that other key players missed.

This kind of player impact analysis is only possible (for me) for leagues where advanced data is readily available online (such as whoscored.com).

Player Impact Calculations B

To calculate the impacts players had on their domestic team for leagues where less data is available, I rely on market values.  I adjust these market values by age to try and correlate them more strongly to player ability. The more different the player value rating the larger the adjustment to the domestic team ability rating will be.

Initially put 75% weight on the on/off pitch playing impact and 25% weight on the player market value adjustment. This then shifts on favour of the player market value adjustment the lower the % of players each nation has for whom I have calculated on/off pitch impact.

Team Ratings

Figure 2

Here you can see Germany’s world cup squad play for the strongest average domestic teams (first data column).  This is partly because quite a few of them play for Bayern Munich.  They aren’t however Bayern’s best players (a lot of competition there) and you can see this in the 2^nd column – they have one of the lowest on/off player impacts of the top teams.

You can see from the 2^nd data column how many sides have positive player impacts.  This is what we’d expect to see – the international players all improve their domestic sides.  When there is minimal player impact data the player’s impact counts as zero. 

Saudi Arabia’s squad containing players who are far from key players for their domestic squads is visible by the red cell in the 3^rd data column.  Qatar has a similar thing going on.    

Figure 3

These are the most finalised ratings I have so far (tweaks always ongoing).  The unit is such that if for example, France played Spain I would have France as 0.4 goal favourites (around 11/10 shots). The ratings have been edited such that the average team has a rating (expected goal difference) of zero.

Simulation Results

Next, I used some python script (with credit to google gemini) to simulate the entire world cup.  One shortcoming of the simulation so far is I don’t account for high/low scoring games.  Each game is assigned around 2.7 total goals.  I don’t think this majorly affects the results, but it is something I want to incorporate soon.

Figure 4

• Compared to the global betting markets I’m low on Spain.  Players from Sociedad and Athletic Club knock them down their team ability portion of their rating and I’m not measuring a fantastic impact for players like Rodri and Laporte in their clubs this season.
• Brazil and Germany also somewhat favoured.  I have Argentina as overrated, alongside Spain.
• I have a lot of underdogs with realistic shots.  The USA, Turkey, Austria, Ivory Coast and Senegal are all favoured in my simulations over global sentiment. 
• Most of my disagreements are quite likely due to actual performances that the teams have put out in World Cup qualifying.  For example, I am low on South American teams like Ecuador and Colombia but they looked exceptionally solid in South American qualifying.  I give Ghana almost a 1% chance of winning the tournament despite them not managing a friendly win in 6 attempts.  Time will tell how severe the shortcomings of my player-based only rating system are.

Betting

Outright winner

I have backed the following sides (decimal odds)

Brazil 11.0

Germany 17.0

Austria 220

Sweden 210

Ivory Coast 250

USA 90

Turkey 95

Some of these odds have changed (mostly for the worse…).

Individual matches

My current group game bets (all based on my team ratings) are as follows:

USA vs Paraguay – USA win 2.02

Australia vs. Turkey – Turkey win 1.755

Sweden vs. Tunisia – Sweden win 1.96

Ivory Coast vs. Ecuador – Ivory Coast +0.25 1.93

Ghana vs. Panama – Ghana win 2.16

As usual I expect reviewing the closing odds will be key to determining the strength of these bets.

Fifa Golden Ball

Here are my ‘highest rated players’ in the World Cup.  The calculation here is the ability of the team they play for vs. their impact on/off the pitch.  It’s quite a crude calculation but still a potentially useful one. To describe the meaning of the rating it’s something along the lines of ‘This player would be an average player in a team of this level (expected goal difference per game)”. Arsenal/Bayern/PSG/Man City have an approximate skill level of 1 on this scale.  One issue is the players who have played almost every minute for their clubs can only have an impact rated close to zero.  

Figure 5

Going off these ratings I am backing the following players to win the Golden Ball:

Raphinha 34.0

Jeremy Doku 81

Bukayo Saka 170

Conclusion

I have studied an international tournament like this before, so I expect to learn quite a few things during the tournament.  I look forward to considering how much a team’s true level can differ from the sum of its parts.

It is very likely that I will learn the outright market and betting lines are very accurate already, more accurate than my ratings at least.

Thanks for reading today, I hope you enjoy the World Cup 2026.

Please get in touch in the comments below or at x.com/samh112358 about anything you’d like.