Syzygy Analytics

Over the last month I have been doing a lot of performance reviews of my football ratings.  I wanted to prove that capping the value of an xG chance is a useful thing to do.  I had this idea from a very early stage in my football analytics journey and today I want to share it.

Normal xG is initially designed as a descriptive measure.  It uses different parameters about the shot to calculate the average chance of scoring each shooting attempt made on goal.  My goal is to approach all this kind of analysis with a more predictive mindset.  I am happy to discard any descriptive meaning for a metric that does a better job of telling us how well a team will perform tomorrow.  What do the attempts on goal today mean for the ability of a team tomorrow?  That is a different question to ‘what was the average numbers of goals a team would score from the shots today?’. 

Firstly, today I want to look at adding values for goals to the xG metric.  I speculate that adding real goals to the xG metric will improve the predictive performance.  This is because I expect adding some value for real goals will capture some above/below average finishing ability or goalkeeping ability of the team.  It may also be the case that goal difference diverging from expected goal difference is telling us that the assigned xG values were not representative of the real quality of the chances.

Afterwards I will consider the idea of capping large xG chances. 

My xG metric uses 33 different zones and 7 different shot types.  There will be more precise and complicated xG models out there but I’m not sure the precision is necessarily worth much predictive value.  I have incorporated schedule adjustments, red card adjustments and lowered the value of rebound shots. 

Why do I compare to betting odds?

To test the metric, I am comparing the metric to betting odds at kick off (thanks to Joseph Buchdahl for his footballdata.co.uk website).  I am assuming that the global betting markets are quite efficient (at kick-off) and are a good way to test the metric performance.  I have been doing similar analysis looking how xG predict goal difference (GD) as well, but it doesn’t feature today.  I preferred comparisons to the betting odds as the sample size for GD drops as we go through the season, GD is more subject to variance and may be complicated by game state.  I hope I will incorporate goal difference in the future to test predictive performance as it still has value.  There’s a lot of permutations with which I can analyse this material today so I had to settle on one of these options otherwise this article would keep getting delayed!  

This is using data from the 2021/22 & 2022/23 Premier Leagues, the 2023/24 Serie A and the 2023/24 La Liga.  These leagues were randomly selected as I was having trouble with certain datasets.

Adding values for goals

My process for testing the metric performance against the betting odds is as follows:

1. Calculate performance for each team for the season to that point
2. Price up each game using said metric and home advantage
3. Compare how favoured each team is vs. how favoured each team is according to the betting lines at kick-off
4. Calculate the average difference between how the metric would price up each game vs. how the betting lines did

Figure 1

Each series in Figure 1 compares the predictive performance of the raw xG metric with the performance performance of the same xG metric with certain values added when goals are scored (as shown by the key).  When the line drops below the x-axis into the green area, this means the added goal value is improving the performance of the rating and vice versa.  Each series uses a 5-game rolling average. 

When we add 0.25 xG for real goals (solid red line) we can see it damages the predictive performance of the team until we reach a sample size of 10 games.  For sample sizes of larger than 15 games the graph shows that we learn more about the real ability of the team by adding 0.25 xG for every goal scored.

Adding 0.5 xG for goals (solid black line) hurts the metric more for early chances but then helps a little more for larger samples.  As we are effectively adding 0.5 xG for every 1 xG scored this means this metric has a weight of 66% xG and 33% goals.  I didn’t expect the uptick on the right side of the graph, but the most likely explanation is noise based on teams having different motivations in the late season.   

Adding 1 xG (dotted black line) for every goal scored is clearly too much and even for larger samples the metric performs poorly compared to the previous smaller added values.

Overall I prefer working with the smaller added goal value of 0.25 xG.  Larger goal values than this damaged the performance for small samples more than they help over larger samples.

Capping high value xG chances

Next, I want to consider capping the value of xG chances.  In standard xG models, the highest possible xG value will be close to 1 as some shots are almost guaranteed to be a goal (e.g. shots attempted from very close to the goal line)   

Large xG chances often feel too large to me, i.e. does a team deserve 6 times the credit for creating 1 0.6xG chance vs 1 0.1xG chance.  Or similarly, should creating 6 0.1xG chances be worth more than creating 1 0.6xG chance?   

Using similar methods to Figure 1 I set out to compare the performances of xG metrics with different caps.

Figure 2

Let’s consider how each cap affects the predictive performance of the metric.  All these metrics are compared to a raw xG metric with no caps and no value added for goals.  

The 0.1 xG cap is an extreme cap that removes a lot of shot quality information.  Any more of a cap and we essentially be left with a simple raw shot ratio.  As soon as the sample becomes a reasonable size this cap significantly damages the metric performance.

The 0.2 xG cap is more intriguing.  We see a clear improvement in metric performance for smaller samples but for larger samples it performs poorly.  If we are using a small sample of games to rate a team (e.g. the first 11 games or a team has been using a much changed line up for a small sample of games) then on average capping xG chances at 0.2 xG will give us a more accurate assessment of their performance.

0.3 xG and 0.5 xG caps see similar trends but with less extremes. 

There is reasonable evidence from this graph that we don’t need to cap xG chances.

Capping the maximum allowed xG value and adding values for goals the same time

So far, I have looked at adding goal values and capping xG chances separately. 

Next I want to consider if we could reach any different conclusions by capping xG chances and goal values at the same time.

I am going to use an xG metric that has an added value of 0.25 xG for goals.

The goal values are separate to the caps so you could argue the cap of a single chance is not really 0.25 xG (if an 0.25xG chance is scored it will be worth 0.50 xG in the metric).

Over the last few weeks, I have been comparing no xG caps with 0.25 xG added goal value against a 0.25 xG cap with 0.25 xG added goal value.  The sample here stands at 11 leagues (I have attached the data at the bottom of the article).  

Figure 3

Data from the 2023/24 Serie A and La Liga, the 2022/23 Premier League, and the 2025/26 Premier League, Bundesliga, Ligue 1, League 1, League 2, Bundesliga II and Championship are used for this graph.

This graph is a comparison between an xG metric with 0.25 xG added for goals scored and an xG metric with 0.25 xG added for goals scored and a maximum limit of 0.25 xG for any 1 chance. We can see a consistent improvement of around 3% for all the sample sizes pictured here.  This is a different trend to when we capped xG chances in the absence of all added goal values (Figure 2).  I think the two changes blend together well as goals helps improve the predictive power of the rating at larger samples while capping high xG chances helps at lower samples.

The average error for each game is a little under 0.2 goals.  The capped xG metric is showing around a 3% improvement in prediction which is a little less than 0.01 goals per game.  Not that high but this will amount to something significant over hundreds or thousands of games.

Conclusion

We saw today that weighting xG difference with real goals improved the performance of our metric.  I think the best weight for goals in an xG metric is somewhere between 18-33% (between 0.18 xG and 0.50 xG added for each goal).  The lower end of this (0.18 xG) will not hurt the performance of the metric for smaller samples while the higher end might be best for larger samples. 

Next, we looked at limiting the maximum value that could be awarded to individual chances.  When doing this we saw we improved the xG metric up to around 12 game samples but for larger samples we were better off just using full xG.

Finally, I combined the two ideas and capped xG chances for an xG metric that included value added for goals.  This result in a solid improvement of the performance of the metric across a large spectrum of sample sizes. 

I am planning plenty more work in this area, including some analysis with game state adjustments (I theorise that capped xG stands to benefit more from a game state adjustment than non-capped!).

If you enjoyed this piece today, please consider subscribing using the box on the top left on the home page syzygyanalytics.co.uk.  Also if you want to contact me for any reason at all you can leave a comment or DM me at x.com/samh112358.

Thanks for reading!

Unfortunately, a bad weekend last time out where a lot of my line predictions were wrong (I did a PL week 27 preview).  A lot of it was bad line-up news but that couldn’t explain all of the error.  More work needed!

In today’s quick post we are going to consider whether teams can perform different shot types more repeatably. 

I have split shots up into

• Foot shots assisted with passes
• Shots assisted with crosses
• Shots assisted by a throughball
• Counter attacking shots
• Unassisted shots

I am totalling the xG in each category for each team across 16 leagues and calculating the R-square (Rsq) value for these totals between the 1^st and 2^nd half of the season.

Figure 1

Interpreting these numbers

Foot shots assisted with a pass are the most common shot.  The proportions are approximately 1 : 0.5 : 0.06 : 0.12 : 0.59 for passes : crosses : throughballs : counter attacks : unassisted.  This is relevant because smaller sample sizes affect how variable the rsq value will be.  I hope that taking the average across many leagues has helped smooth out these anomalies. 

Passes (foot shots)

Pass assisted shots see the highest correlation.  This implies they are the bread and butter of good teams and may be the most skill dependent category of football.  The correlation is higher on the attacking side than the defensive side – this could imply teams are generally more in control of their attacking performance than their defensive performance.

Crosses

Cross assisted shots see lower repeatability between half seasons than pass assisted.  I think this makes sense – finding a teammate with a cross is more subject to randomness than shorter passes.  We do not see the same attack/defense split as we did for pass assisted shots.  I also find this intuitive – crosses are a weaker form of attack where defensive strength/ability often is an important factor.  Strong defences consistently snuff out crosses; weaker defences do not. 

Throughballs

The large attack/defence split for throughballs is very interesting.  It means throughballs may be largely to the credit of the attack than a result of poor defending.  (Remember the low Rsq value on the defensive side means teams conceding a lot of xG via throughballs in the 1^st half of the season do not go onto concede a lot of xG via throughballs in the 2^nd half of the season).

Counter Attacks

Correlation here shows as low on both sides.  This implies counter attacks are situational and not a play style teams can rely on too strongly. 

Unassisted shots

Unassisted shots being a little luck based is an idea I’ve (rightly or wrongly) used for a decade now so not too surprised to see a low correlation.  This is the only category where the defensive side shows a significantly higher correlation which makes sense owing to the fact unassisted shots often come about from defensive errors (and not attacking skill).

I hope you found this short piece interesting today.  The idea was to help people think more about things that could help drive predictive analysis.  Please subscribe using the box on my home page if you want to read more!

Last time I did this for week 25 my predicted prices were a little less accurate than the early prices.  The biggest miss was Manchester City as money came in for Liverpool.  By kick off these teams lined up as approximately equal strength.  It seems like Doku & Gvardiol are seen as important and maybe Liverpool were seen to be finding form again vs. City losing it. City played well in the game but I don’t really believe closing prices are beatable (making money while backing at kick off).

All the prices are derived from season performances (in depth metrics) and expected lineups for the game.  I will discuss points of interest/contention for each game.

Aston Villa vs. Leeds

Market price (home -draw away) (Wednesday 18^th February)  1.915 3.725 4.85

It looks like Stach and Struik are fit for this game (this is very important).  Villa have a question mark with Matty Cash.  I have Villa priced up right around the evens mark here.

Predicted 2 3.8 4.25

Brentford vs. Brighton

Market price (home -draw away) (18^th February)  2.05 3.925 3.875

My biggest position of the week.  Brentford still seem a little unfashionable but should be back to full health (with Schade back) and Brighton are quite ordinary.  Brentford should go odds on here; this is one I will be disappointed to get wrong. 

Predicted: 1.9 3.9 4.5

Chelsea vs. Burnley

Market price (home -draw away) (18^th February)  1.245 7.9 14.75

My spreadsheet is saying Chelsea 1.3 just based on 25/26 performance.  However, once you add Chelsea being relatively fresher (no Europe/cups this week compared to the rest of the season), Burnley having poor form (since Cullen injury seemingly) and Chelsea’s lineup between stronger than their average (Palmer, Caicedo) this season I agree with the current price.

Predicted: 1.245 7.9 14.75

West Ham vs. Bournemouth

 Market price (home -draw away) (18^th February)  2.52 3.825 2.92

A tricky one in my opinion.  West Ham’s recent form is substantially different from the whole season so it’s a question of how much weight to put on this.  2.7 was available early in the week on West Ham which I did not take.  Bournemouth’s performances were expecting to drop more since they lost Semenyo and had other injuries, but they have been playing respectably.  I’m sticking with the current market price.

Predicted: 2.52 3.825 2.92

Manchester City vs. Newcastle

Market price (home -draw away) (18^th February)  1.45 5.85 7.4

Newcastle will be injured and tired for this game.  This is already well factored into the price as City are only a tick or two longer than their starting price vs. Fulham.  Their performance against Fulham was ok, but you can’t ignore the chances they conceded just because they were 3-0 up. 

EDIT:  Thursday 19^th.  1.515 5.35 6.7.  City have drifted after Newcastle’s performance mid-week in Azerbaijan.  I’ve been surprised about the money against City recently, I suppose poor form and the absence of Doku and Gvardiol.  City get quite a large positive game state adjustment in my ratings – they have spent so many minutes ahead.  Their performance at tied is comfortably top of the league.  Haaland is a fitness doubt which is a big issue as well.

Predicted: 1.515 5.35 6.7

Crystal Palace vs. Wolves

Market price (home -draw away) (18^th February)  1.64 4.35 6.3

Wolves must be close to checking out.  They were dreadful against Forest and are potentially not that fresh for this game after an FA Cup game on Sunday and Arsenal on Wednesday.  Palace have a UECL match on Thursday so I will want to see how many players they rest before being confident of a Palace shorten.

Predicted: 1.55 4.59 7.3

Nottingham Forest vs. Liverpool

Market price (home -draw away) (18^th February)  4.55 4.15 1.855

Forest are in Turkey on Thursday night which really feels like not what they need right now.  It complicates this game as Liverpool will be fresh and raring to go.  I don’t know how seriously Forest will take the EL game so I’m not making a play for now.

Predicted: 4.235 4.15 1.87

Sunderland vs. Fulham

Market price (home -draw away) (18^th February)  2.79 3.375 2.87

Sunderland will still be without Xhaka but otherwise both sides are quite healthy.  Sunderland look big currently in my eyes.  I rate the performance of the 2 teams this season about equally.  The situation then tips in Fulham’s favour because of Fulham’s fresh lineup and Sunderland missing Xhaka.  They both missed players for AFCON so really the difference is just Xhaka and I can’t get all the way from Sunderland 2.25 (given my comment on the teams being equal strength) to 2.8 so I’m siding with Sunderland here.

Predicted: 2.5 4.2 2.76

Tottenham vs. Arsenal

Market price (home -draw away) (18^th February)  6.9 4.35 1.605

I agree with this line

Predicted: 6.9 4.35 1.61

Everton vs. Manchester United

Market price (home -draw away) (18^th February)  3.825 4.02 2.03

2 healthy sides except for Grealish.  I have a small lean towards Everton.

Predicted: 3.67  4.02  2.09

Introduction

This week I’m want to test a variety of different metrics to compare their predictive power.  I have always focused on the predictive and I think providing value as an analyst is almost solely a function of predictive analysis than descriptive.  For this reason I am less interested in a descriptive idea such as xG (a measure of how often a chance is scored) but rather what is the correct value of a chance today to predict the success of the team tomorrow.

Most xG models use complex calculations to arrive at values for each shot.  A lot of data is used such as shot type, distance from goal, distance from centre line and even parameters such as speed of attack or defensive coverage.  This could present a problem for analysts – this data may not be readily available, or they may not have the expertise required to turn the data (e.g. X/Y coordinates) into information (xG value).  In this post I am going to be comparing the predictive performance of some simpler shot count metrics with some xG metrics.

Today the the metrics I will be comparing are as follows

• Understat.com npxG difference (non penalty expected goal difference)

• Understat.com xPts

(Expected points is how many points a team would be have scored on average for the season so far based on simulations of the xG value of their chances.)

• Total shot difference per game

(The difference of shots attempted and shots conceded)

• Total shot difference per game (game-state adjusted)

(Teams that lead have a shot difference per 90 of at least -5 (so teams that trail have a shot difference of at least +5).  This makes leading/trailing teams appear worse/better than their real level.  I compare total minutes leading to minutes trailing for each team to adjust for this.  For example, a team that has 5.5 more 90s leading vs. trailing in the first half of the season will on average have had their shot difference reduced by ~28 shots owing to game state.  As this is due to game state and not their ability I add 28 shots on to their total shot difference.  I have written a few articles so far on game state before on my blog so check them out if you want to read more:)

https://syzygyanalytics.co.uk/2025/08/01/soccer-team-rating-iii-game-states-i/  (part 2 linked at the bottom)

• Total shot difference (game-state & big chance adjusted)

(A shot that Opta has tagged as a big chance counts double)

• Total shot difference (game-state, big chance & goals adjusted)

(A goal is worth a bonus shot)

• Market implied team ability

(The betting odds at kick off are potentially an excellent measure of every team’s ability as a lot of smart money has shaped these markets.  This makes the team ratings we can derive from betting odds an interesting metric for comparison)

• My own team rating metric

I have developed this over about 10 years and included it for the sake of comparison.

***Thanks to Joseph Buchdale at footballdata.co.uk for the betting market odds, whoscored for shot data and understat.com for xG data.***

I will be measuring the performance of the above 8 metrics in 2 different ways. 

1. How well does performance in the 1^st half of the season predict goal difference in the 2^nd half of the season and
2. How well does performance in the 1^st half of the season predict betting odds at kick off for the 2^nd half of the season.

Results

Figure 1

Figure 2

Interpreting the results

Figures 1 and 2 shows R² values when each metric for the 1^st half of the season is correlated to each team’s goal difference in the 2^nd half of the season. 

• xPts and npxGD hold similar value.  Personally, I slightly prefer npxGD as every chance is credited with its xG value.  xPts counts similar chances different because a chance created in a close game is worth more than a chance created in a game where one team is dominant (think about how many xPts a 0.5xG chance is worth in a game with an xG total of 3.0xG – 0.2xG vs. a game that’s 0.9xG – 0.9xG).  This is potentially a complex point and not the topic of today’s article.
• It is up to you whether you consider figure 1 (goal difference) or figure 2 (market derived ratings) as a better indicator of team strengths.  I think goal difference is more susceptible to randomness but it’s also more authentic.
• As the 1^st half metrics are based off 17-19 games (quite a big sample), we see goal difference alone predict to a respectable level.
• Each adjustment we make to the raw shot difference is valuable. Each extra adjustment may have diminishing returns.  This is because playing in leading game states makes it easier to create a higher proportion of big chances.  Similarly, more big chances mean more goals.  I tried making the adjustments in a different order and found the game adjustment was the best single adjustment followed by big chances then goals.
• All 3 adjustments combined outperform the understat xG metric.
• My metric shows what extra is still possible (although calculation complexity is on a par with xG, for example it includes shot location). 

Conclusion

The conclusion here is fairly clear – you can make a very solid metric with simple quantities like shots, goals, big chances, minutes leading/trailing.  This test was based off a sample of 17-19 games, if we were to use a smaller sample, xG would fare even worse against these shot count metrics.  I think xG is a fairly disappointing performer in terms of complexity vs. performance and arguably even outright performance.

   

Please send feedback if you found this interesting today, have any questions or any requests on what content you’d like to see!

This week I am writing down my thoughts on week 25 of the Premier League.  My goal is to beat the closing lines, so let’s try and nail down what I think the true price for every game should be!  I am working on the assumption that the closing price = the true price.

Leeds vs. Nottingham Forest

Market price (home – draw – away) (3pm GMT Friday 6^th February): 2.29 – 3.425 – 3.675

A tricky one because there is some disagreement between previous market prices for Leeds and my ratings.  My spreadsheet is spitting out 2.07 Leeds (an advanced metric using team performance from this season and predicted lineup strength compared to averages for each team).  That doesn’t consider whether this game has a draw bias – both teams may feel like a point is a reasonable result?  Leeds have much better form in the last half of the season so if you have a read that’s signal rather than noise then this could mean Leeds should be shorter.  Leeds’ rating this season might be slightly hot and Forest’s might be slightly cold.  Leeds underperformance this season relative to shooting metrics is included a bit in my rating but possibly it deserves a little more weight.

Quite torn but I’m going to go with a true price of the same as the market

Predicted Closing price: 2.29 – 3.425 – 3.675

Manchester United vs. Tottenham

Market price (home – draw – away) (3pm GMT Friday 6^th February): 1.615 – 4.95 – 5.65

Tottenham’s injuries are a little shaky but if we can be confident that Romero/Van de Ven and Solanke are all good to start I’m predicting a small move back towards Tottenham.  Tottenham are still quite injured but have been all season so we can potentially call Spurs’ line-up ‘average’ while the added freshness of not having CL (not a factor for United who never had CL) is also key. 

Predicted Closing price: 1.67 – 4.63 – 5.41

Arsenal vs. Sunderland

Market price (home – draw – away) (3pm GMT Friday 6^th February): 1.245 – 7.1 – 17.75

Arsenal minus Saka and plus Havertz while Sunderland still have no Xhaka are all relevant considerations but in the end, I have no disagreement with the current line.

Predicted Closing price: 1.245-7.1-17.75

Bournemouth vs. Villa

Market price (home – draw – away) (3pm GMT Friday 6^th February): 2.76 – 3.725 – 2.7

Another difficult one as the lineups for these teams have not been consistent recently.  Bournemouth have lost Semenyo to Manchester City and Tavernier to a hamstring injury, but their last few performances have remained competitive.  Villa have numerous injury issues but also some good news like Onana is back, and Luiz/Abraham signings.  In the end I am currently going for:

Predicted Closing price: 2.86 – 3.78 – 2.60

Burnley vs. West Ham

Market price (home – draw – away) (3pm GMT Friday 6^th February): 3.475 – 3.675 – 2.27

Burnley have been very poor in their last few games.  Josh Cullen appears to be a very important loss.  Surprisingly I estimate both teams to field starting lineups about level with their season averages (using transfermarkt.com player values).  This would show dramatic value on Burnley and even adjusting for Burnley’s recent poor performances I put enough stock in a few bad performances to justify the starting price we see now.

Predicted Closing price: 3.1 – 3.62 – 2.43 

Fulham vs. Everton

Market price (home – draw – away) (3pm GMT Friday 6^th February): 2.21 – 3.425 – 3.925

I rate Everton slightly better and additionally think their predicted lineup is better than their average this season.  This is because of valuable African players they missed during AFCON, Branthwaite’s return and an otherwise fit lineup.  The loss of Grealish is obviously disappointing but more than made up by these factors. 

Predicted Closing price: 2.34 – 3.37 – 3.57

Wolves vs. Chelsea

Market price (home – draw – away) (3pm GMT Friday 6^th February): 5.4 – 4.35 – 1.705

Some of Chelsea’s key players (Caicedo/Fernandez) have played a lot of minutes recently but Palmer should be good to start here.  Wolves race is nearly run so I have a slight preference to Chelsea

Predicted Closing price: 6.2 – 4.3 – 1.65

Newcastle vs. Brentford

Market price (home – draw – away) (7pm GMT Friday 6^th February): 2.15 – 3.725 – 3.775

Newcastle are fairly injured – Gordon is probably out, Guimaraes is 50/50, Joelinton is out plus various other defenders.  Brentford have a key suspension in Kevin Schade who was sent off against Aston Villa.  Brentford are surprisingly good, so I see some value siding with them at the current prices.

Predicted Closing price: 2.28 – 3.83 – 3.33

Brighton vs. Crystal Palace

Market price (home – draw – away) (7pm GMT Friday 6^th February): 2.09 – 3.725 – 3.975

Palace are recovering from their slump and face a Brighton lineup that is strong as it’s been all season. 

I’ve changed my mind a few times on how to rate Palace here, so I’ll just settle on no movement

Predicted closing price 2.09 – 3.725 – 3.975

Liverpool vs. Manchester City

Market price (home – draw – away) (9pm GMT Friday 6^th February): 2.43 – 3.875 – 3.025

Manchester City had a cup tie mid-week but were able to rest a few key players.  Liverpool’s line-up is in good shape with Salah poised to find some form again.  City are minus Doku and Gvardiol, both of whom are big misses.  I’m struggling to justify Liverpool being the price they are.  I have City provisionally priced at 2.5 which is quite far from the market price.  I hope Ruben Dias can return but it could be argued that coordinating with a new teammate (Guehi) for the first time away to Liverpool is not ideal.  I must still favour Manchester City here

Predicted closing price: 2.67 – 3.91 – 2.67  

I’ll review these predictions next week to see how many I got right.  I will be trading these positions on the Betfair Exchange.

I’m also working on an ultimate metric comparison to illustrate how metrics like raw xG or shot totals compare to my more complex team rating metric.  I hope to post more on this in the next few weeks.

You can subscribe using the white box on the home page to be notified on future posts

Thanks for reading!

In my last article I looked for evidence that manager changes had an effect on a team’s performance. I did not see any evidence that sacking a manager had a positive effect on the teams included in that sample (5 leagues from the 2024/25 season.) This week I have done the same analysis for a new sample, namely the top 5 European leagues in the 2023/24 season.

Results

The thicker red lines are the average of all the thin team performance lines.

Table 1

Table 2

Conclusion

The results here add little to what was seen last time out – namely a manager getting sacked had little influence on a team’s performance. The average on the team performance plot is less pretty than for the previous sample – it stays pretty flat. As teams had not suffered a recent performance decay before the managers were changed, their performance did not regress upwards. This is a little surprising but it could be a quirk of the sample. It’s possible if I looked at goal difference it might illustrate better why the managers were removed for the teams involved in the plot. I will include that metric next time I look at managers! Overall my stance has remained the same – I am sceptical that managers have a strong link to the performance of a team.

Introduction

Football managers form a large part of football culture in Europe and beyond with hirings and sacking making up a decent chunk of the column inches written on football. Similarly, they are paid handsomely, with wages of the top managers comparing well with wages of the top players.  Today I am looking at data from about 40 manager sackings in the 2024/25 European season across 5 different leagues. I want to see what the data says about how this manager upheaval influenced the performance of these teams.

I am going to focus on two different metrics, one is team performance which considers shots, passes and goals and the other is derived by team’s betting odds at kick off.

The 5 leagues in this study are from the English Premier League, English Championship, English League 1, Serie A and La Liga.  On the following graphs there are 40 lines representing teams that changed manager last season.  There are some teams not included because the manager change was either too early or too late in the season to make a proper comparison of the team’s performance before and after. 

As you will see every line meets at (0,0) in the centre because the team rating is relative to the team rating at the time of the manager change and 0 refers to zero matches since the manager was fired. (The graphs are high resolution so you should be able to zoom in, if required.)

Results

Figure 1

The ratings here include as many adjustments as possible, such as fixtures, red cards, rebounds e.t.c.  This thick red line (the MEAN of all 40 lines) shows many teams suffered a decline in performances before seeing their manager sacked (not a surprise).  After another 5 games team performance starts to pick up again.

Figure 2 – (The image is high resolution, please zoom in)

The thick red line is the mean of all teams in the sample (40 teams).

This graph derives team ability from betting odds at kick off.  While we do see a slide in team ability in the run up to a manager sacking, this slide continues for 6 more games after a manager change.  It’s worth noting the values on the y-axis – this method of rating teams is less subject to fluctuation than the previous method that solely measured on pitch performance. 

Conclusion

I had a prior scepticism over the importance of most managers that has not been alleviated in this study.  While the first plot does imply a manager change has a positive influence this can’t be separated from natural variation.

To prove this, I calculated team performance in groups of 10 games to get the following plot:

Figure 3

The x-axis is comparing performance between games 1-10 and 11-20 of a season for the same 5 leagues I have used for the previous games.  This means data points/team to the right of the y-axis have experienced an increase in performance for the 2^nd 10 games of the season compared to the 1^st 10 games.  The y-axis then compares the team performance for games 21-30 compared to games 11-20.  The correlation and the line of best fit shows that a lot of data points to the right of the y-axis are also below the x-axis – i.e. all the teams that ‘improved’ in the 2^nd set of 10 games regressed by 50% in the next 10 games.

In fact, this is really intuitive!  Consider a team having performance of some number, x, for the first 10 games, and then a performance of 0.5x for the second ten games.  What performance do we expect for the 3^rd set of 10 games?  Assuming recent form is negligible we have to answer the average of x and 0.5x, so we get 0.75x. This is exactly what figure 3 demonstrates – performance change in the 2nd set of 10 games regresses in the 3rd set of 10 games. The fact the slope is slightly less than 0.5 likely relates to recent form playing a small role.

Now let’s go back and reconsider figure 1. We can not conclude there is any evidence that manager changes are affecting team performances.  Team performance for the 10 games after a manager change almost exactly represents the team performance for the 20 games before the manager change (the data point at x=10 is the average of the data point at x=0 and x=-10). The red line simply displays a natural regression.

The second graph also fails to show any benefit of firing a manager, in fact, it’s the opposite.  Betting odds at kick off are very efficient – I think they have summed up nearly all of the human knowledge of the strength of the 2 teams lining up on the day.  They demonstrate quite a dramatic contradiction of the perceived ‘new manager bounce’.  While this trope holds some weight in terms of results (although likely nothing to do with the manager change itself) the opinion of the smart money in the world often sees a team weaker than before.  In this specific sample the rating recovers after 14 games. It’s possible new managers experimenting with the lineup or caretaker managers with less experience result in the team experiencing a further drop in performance.

In my opinion it is quite hard to judge a manager in terms of results as it’s hard to say what level a team should be performing at.  So many times, has a manager been credited with a fair chunk of his team’s positive fortunes moved to a different team only to disappoint.  Similarly sacked managers have seen their reputation recover elsewhere. 

If had to choose a manger I would study his/her opinions and perspective on the game more closely than I would study his/her results. 

Thanks for reading I hope you found this interesting today!

Thanks to the following resources that helped me create my post today

Transfermarkt pages on managerial changes, e.g. https://www.transfermarkt.co.uk/ligue-1/trainerwechsel/wettbewerb/FR1/plus/?saison_id=2024

Whoscored data for advanced team data www.whoscored.com

And Joseph Buchdahl’s extensive betting market data at www.footballdata.co.uk

Please comment or like if you have any questions and subscribe to my blog using the white box on the home page to get notified when I post.