Last month, I wrote an article exploring Fulham’s under performance of their Expected Goals (xG) stats. In that article, written after Fulham’s 18th game of the season (vs Manchester Utd), we noted that Fulham had only scored 75% of the goals expected by the Statsbomb xG model.
Many commentators, including Scott Parker himself, have noticed that when measured in Expected Goals, Fulham’s attack is performing quite well (about mid table for the premier league). However, this is not leading to actual goals being scored, and is a primary reason why Fulham find themselves in the relegation zone.
My previous exploration of this poor xG number indicated that this shortfall in goals scored was due, at the time, to two main factors i) Fulham’s poor performance from the penalty spot, and ii) a run games where opposition goalkeepers seemed to have consistently good performances.
The data at the time did not support the idea that Fulham’s shooting, at least from open play, was particularly poor.
I hypothesised that this implied a run of bad luck, and that Fulham could expect to see a ‘regression to the mean’ and start scoring at a higher rate going forward.
5 further matches in, I thought it worth updating the numbers and seeing what trends might be emerging (writing after the win against Everton).
Converting Expected Goals
Like my previous article linked above, this latest analysis looks at various measures of Expected Goals (xG), which, as a reminder, is the idea that we can estimate the probability of individual scoring chances being converted into goals, and therefore determine, by looking at the overall quantity and quality of chances created, how many goals we would expect a team to score.
Many are sceptical of these models, but the Statsbomb xG data that I use is highly accurate. In the premier league, so far this season, this model has estimated that teams should have scored 605.8 goals (excluding own goals), the actual number of goals scored (excluding own goals) is 617, which means the model has predicted the number of goals scored to within a 2% margin of error.
So the model is accurate in aggregate for the league, but individual teams differ from predicted performance, the question we need to ask is whether this is just due to bad luck (driven by small sample sizes) or is something else going on with those teams?
At the time of my last article, Fulham had an actual goals to xG conversion rate of 75% which is bad and, as such, they were ranked 18th in the league for this measure. Only Sheffield Utd and Burnley were doing worse.
At the same time, I pointed out Southampton were at the other end of the spectrum, over-performing with an xG conversion factor of 150%. I had hypothesised that, as more data was collected (i.e. more chances observed), most teams would tend towards the league average conversion rate.
I have now updated this data which is shown in the chart below, with Fulham’s position highlighted.
Most of the previous outliers (Southampton, Sheffield Utd, Arsenal, Burnley) have indeed reverted towards the league mean, as predicted, but Fulham, worryingly, have gone in the other direction. Their xG conversion is actually getting worse and they are now the tied worst for this measure in the league with 69% conversion of expected goals into actual goals.
So, that’s not good news!
Why are Fulham Failing to Convert Expected Goals into Goals?
When I last investigated this, I compared models of chance quality (standard xG) with those looking at shot placement (post-shot xG). I noticed at the time that Fulham were, roughly, taking shots of about the same quality that the models would expect of a typical premier league team.
In other words the shooting (excluding penalties) was fine. The problem was, at the time, that the shots were being saved at a higher than expected rate given their quality and this, I hoped, was just bad luck.
I have updated the chart below which looks at whether the number of goals being scored is more or less than the number expected, using a post-shot xG model.
Green bars represent games where the opposition goalkeeper has saved more shots than expected by the model and red bars where they have saved fewer than expected. The orange bar shows the cumulative over/under performance of opposing keepers.
I have overlaid a red arrow showing the phase of the season where goalkeepers were overperforming against Fulham, this phase of the season accounted for most of Fulham’s missing goals this season when I last investigated this.
My prediction (hope) was that this was a temporary run of bad luck with goalkeepers and would stop. And indeed this prediction was correct, the subsequent blue line shows that over the most recent 7 games, opposition goalkeeper performance has been in line with modelled expectations.
So this is good news (isn’t it?)…but why then has Fulham’s xG conversion continued to deteriorate as outlined earlier?
Shot Quality
The previous article (written after the Manchester United game) found that models looking at the quality of shots taken by Fulham were not highlighting any particular weakness. More precisely, the cumulative post-shot xG generated by Fulham, up until the Manchester United match, was only 2 goals less than predicted by the ‘normal’ expected goals model.
This difference can almost entirely by explained by Fulham missing the goal with 2 of their penalty shots, which means from open play at least, the post shot xG model was showing the same outcome as the normal xG model: and this means that Fulham’s shots were of expected quality.
This conclusion was contrary to the popular narrative that Fulham were poor at finishing.
The chart below updates this analysis, it shows the cumulative difference in expected goals between the normal and post-shot Statsbomb models as the season has progressed. The orange line marks the point in the season where I last wrote about this.
The table highlights Fulham’s new problem, the sharp drop in the blue line at the end of the chart shows that the quality of Fulham’s shooting has suddenly declined. In 5 games since playing Manchester United, Fulham are missing 4 goals entirely due to shooting being worse than expected, given the chances Fulham were creating.
This is what not being clinical looks like in data form!
So just as Fulham’s run of bad luck against opponent keepers comes to an end, it seems their shot quality has gone off of a cliff!
These different drivers of goalscoring underperformance can be brought together in the chart below. It shows the extent to which Fulham’s missing expected goals are explained by i) poor shooting (shown in blue) and ii) good opponent goalkeeping (shown in orange).
As you can see, I have split the season up into three phases:
- In the first phase of the season (first 11 games), Fulham were broadly scoring at the rate predicted by models. There is some shooting underperformance, and this can be explained by off-target penalties against Sheffield United and Everton.
- In the second phase of the season (next 7 games), Fulham have a run where their shot quality is in line with model expectations, but they have a run of facing overperforming goalkeepers (orange expands)
- In the latest phase (last 5 games), goalkeepers have not performed exceptionally against Fulham, but the quality of shooting has been very poor (blue expands).
Conclusions
Fulham continue to create good volumes of goal-scoring chances without scoring the goals these chances warrant. This is a problem and time to resolve it is rapidly running out.
Other teams who were previously under-performing their xG conversion metric (Burnley, Arsenal and Sheffield United) have recently demonstrated a reversion to the mean, with results improving (to some extent) as a result. Similarly, Southampton, who were over performing their xG conversion metric, have also reverted towards the mean with a corresponding decline in results.
Fulham’s deterioration is worrying because, as the sample of matches increases, the extent to which simple bad luck can explain the underperformance diminishes. This is further highlighted by the evidence from the post shot models which says that Fulham have started to shoot poorly.
I do still believe that a reversion to mean performance is quite likely, that xG conversion will improve and that the data is largely driven by bad luck. Even if Fulham are, collectively, bad at shooting, it seems implausible that as a unit the problem is so extreme that they need twice as many chances (of equal quality) to score the same number of goals as teams like Southampton, West Brom and Crystal Palace. Yet this is what the data says for the season so far.
If Fulham keep creating chances at a high rate, then in the long run, a good number of goals will follow. That is the principle behind xG modelling and I still hope we will see this principle lift Fulham’s season. If it does, we have to also hope it happens in time to drag Fulham out of the relegation zone.
Anyway I will continue to monitor xG conversion and the drivers behind it as the season goes on, it is clear to me now that this will be the critical factor in whether Fulham deliver on the clear potential that they have as a squad.
Cottage Analytica 