You can only play as well as your opponent allows?

This somewhat unusual data structure requires the use of cross-classified multilevel models, in which a single Level-1 observation (team-match) can simultaneously belong to several Level-2 units (teams). The analyses must also take into account the specific nature of the dependent variables: number of shots, number of shots on target, number of goals, and xG. The first three are count variables, with right-skewed distributions and values corresponding to non-negative integers. The model I estimated accounted for this by using a Poisson distribution. xG, by contrast, can take non-integer values but cannot fall below zero, while high values are relatively rare. For this reason, in addition to a standard Gaussian model, I also estimated a model in which xG was log-transformed. The results of the two analyses were very similar.

To decompose the variance in these variables into three sources (i.e. the focal team, the opposing team, and the match / random factor), a null model with no explanatory variables was sufficient. The data came from SportMonks and FotMob, and the transformation history from raw data to the gold layer can be traced here. The .csv file and notebook needed to replicate the results presented below can be found here.

FIRST THE ATTACK

Let us start with attacking performance: what did the league hierarchy look like in terms of shots, shots on target, expected goals, and actual goals per match?

AND WHAT ABOUT THE DEFENCE?

We get a fuller picture of the teams' performances when we look at statistics calculated with respect to their opponents.

With the opening of the 2026/2027 English Women's Super League season fast approaching (4 September, with the heavily invested-in London City Lionesses facing an underinvested Manchester United), I decided to take a closer look at data from the previous campaign. To begin with, I wanted to establish the extent to which WSL teams' attacking performance depends on their own characteristics, the characteristics of their opponents, and match-specific factors / randomness. In other words, I examined how true, in the context of the best women's football league in the world, Kazimierz Górski's famous saying is: you play as well as your opponent allows you to. (The shared surname is purely coincidental.)

The analyses used data from 132 matches and 264 team-match observations (12 teams × 22 rounds), meaning that each match was considered twice. Why? Take, for example, a match between LCL and Manchester United. In a single analysis, we obtain two sets of statistics from that match (e.g. shots on target, goals): one for the London team and one for the Manchester team. In addition, each such team-match observation is nested within two teams — the focal team and the opponent. The data structure therefore looks roughly like this:

It turns out that, regardless of the indicator considered, the team hierarchy was fairly stable. I would divide the twelve league teams into as many as five groups (not on the basis of any algorithm, but simply at face value). The first would consist of the big three — Manchester City, Arsenal, and Chelsea — with Manchester City in a dominant position and Chelsea performing somewhat below their usual level. All three teams take clearly more shots, in which they outperform the rest of the league. The second group, the challengers, would include Tottenham and Manchester United, who remain behind the big three but nevertheless create and convert more chances than the mid-table teams: Aston Villa, London City Lionesses, and Brighton. The fourth group would consist of three relegation candidates — Everton, Liverpool, and West Ham United — with West Ham appearing the most attack-minded of the three. Leicester would form a category of their own.

The figure below, showing average expected goals and actual goals scored per match, leads to some interesting conclusions. The grey line represents the situation in which reality matches expectations exactly, i.e. match-level xG equals the number of goals scored. Importantly, it separates overperformance (everything above the line) from underperformance (everything below it).

Arsenal, Manchester City, Manchester United, Everton, Tottenham, and Aston Villa scored more goals than would be expected from the quality of the chances they created. These "excess" goals relative to expectations may reflect the strong finishing ability of the players involved — although it is worth remembering that xG models have their limitations, and part of this difference may be a measurement artefact rather than a genuine advantage. Leicester, West Ham, Brighton, and Chelsea, by contrast, scored fewer goals than their chances would suggest. Perhaps they should invest in attacking players during the current transfer window (which Chelsea, it seems, have already done by signing Melvine Malard and Manaka Matsukubo).

While Manchester City dominated in attack, Arsenal emerged as the best defensive side. Opponents took the fewest shots against them, the chances they conceded were of low quality, and they allowed goals exceptionally rarely (0.64 goals per match). This time, Manchester United were not far behind the big three — in terms of shots and shots on target conceded, they actually looked better than Chelsea. Liverpool turned out to be a mid-table side defensively, suggesting that their penultimate position in the final table resulted from a weak attack rather than defensive shortcomings. Thus, signing Vivien Endemann looks like a very sensible move, at least on paper. Defensive weakness, by contrast, was a genuine problem for Tottenham. The arrivals of Alice Sombath from Lyon and Caitlin Dijkstra from Wolfsburg could meaningfully improve the team's position next season. It is also worth looking at two teams that finished next to each other in the table — Everton and Aston Villa — but followed very different paths to the middle of the pack. Everton combined an average defence with a weak attack. The Birmingham club, by contrast, combined an average attack with a weak defence.

SHOOTING IS TO A LARGE EXTENT A TEAM CHARACTERISTIC

But back to the main question: what determines the number of shots and goals to the greatest extent? The figure below answers this question by presenting the results of the multilevel analyses.

Nearly half (47%) of the variance in the number of shots taken was attributable to factors associated with the focal team, while the characteristics of the opposing team (26%) and match-specific factors / randomness (27%) played smaller roles. A similar pattern emerged for shots on target, although with a somewhat larger contribution from random or match-specific factors and smaller contributions from both the focal and opposing teams. This pattern suggests that the ability to generate shooting opportunities consistently is a relatively stable team characteristic. The opponent's characteristics, the course of the match, and randomness matter less in this respect.

BUT XG DEPENDS MUCH MORE ON THE PARTICULAR MATCH

The picture changes when, instead of looking at the number of attempts, we focus on their quality. For xG, only around 30% of the variance was associated with the focal team, and just 9% with the opposing team. Somewhat surprisingly, the quality of shots taken depends only to a very small extent on stable characteristics of the opposition. As much as 61% of the variation in this measure remains at the level of the specific match / random factor. One way of interpreting this is that teams differ fairly consistently in how often they shoot, but the quality of the chances they create depends much more strongly on the dynamics of a particular match.

GOALS? EVEN LESS DEPENDS ON TEAM CHARACTERISTICS

The same pattern is visible in the figure below. The closer we get to the final outcome, the less stable the team's "signature" becomes and the more variation is associated with the individual match. The number of shots reflects a team's more or less attacking identity. The number of goals, in contrast, tells us much more about what happened in that particular match.

The number of goals showed the weakest association with both the focal team (18%) and the opposing team (14%). As much as 68% of the variation in goals scored was attributable to the level of the individual performance. The results therefore point to an interesting continuum:

shots → shots on target → xG → goals

This is one of the reasons why assessing team quality solely on the basis of goals scored can be misleading. Goals are, of course, what determine the result, but statistically they are also much more variable than the process that produces them.

YOU PLAY THE WAY YOU WANT TO, BUT YOU WIN THE WAY THE MATCH UNFOLDS

Regardless of the indicator, the focal team's characteristics explained more variation than the characteristics of the opposing team. This difference was largest for shots and smallest for goals. That does not mean, however, that the opponent is unimportant. For the number of shots taken, opponent characteristics accounted for around one quarter of the variance — still a substantial share. But when it comes down to what ultimately matters — scoring goals — the course of the match / randomness becomes the most powerful factor.

Although these analyses demonstrate the importance of the three sources of variance considered here, they tell us nothing about which specific characteristics of the focal team, the opposition, or the match explain the observed results. That seems like a good topic for the next analyses, now that we know that shots and goals depend primarily on, respectively, the characteristics of the focal team and the course of the match.