Evaluating Playmakers in 9 Leagues

---

We build a data-driven model of the majority which estimates the number of chances created (CCs) that we should expect of the typical player, based on the opportunities given him, such as minutes on the pitch, crosses he gets to make, opportunities given him for passing in the final third, etc. This creates a sliding scale against which to benchmark all players as playmakers. Anyone who makes better use of these opportunities (by making more incisive passes, crosses) will outperform the typical by making more CCs than expected. This excess in opportunity-adjusted chance creation is our measure of playmaking talent, and the very best playmakers will register as outliers in excess CC.

We use public domain data from Opta Analyst to replicate the analyses over 8 different European leagues, 2025-26, plus a single legacy dataset from the previous season for one of the leagues. We build the model of the majority using robust regression, which down-weights the few observations that do not follow trend. This gives a cleaner model of the majority than OLS regression. But the main benefit is that robust regression clearly identifies outliers that do not follow trend. Usually, outliers are simply rejected from analyses as aberrant data due to coding error, unidentified exceptions of equipment failure, and so on. But in judging playmakers, (positive) outliers are to be valued as a central focus of interest.

One of the subplots of this year’s EFL was Bruno Fernandes’ 21 assists for the season, breaking the previous bests of de Bruyne (20), Henry (20), and Özil (19). To achieve such a feat you have to be an outstanding playmaker, but fortune has to be on your side too, because BF’s expected assists for the season are recorded as only 12.32, suggesting that luck, excellent finishing, and / or indifferent goalkeeping helped give his numbers an after-pass boost. Then there has to be opportunity: BF played 90% of the season’s minutes, which is high for players in forward positions; and he took most free kicks and 90 corners. His xA reduces to just 9.12 in open play when set pieces are subtracted. Finally, team-mates needed to help by putting BF in positions where he could create those chances, which Manchester United certainly did. For instance, BF made 871 passes in the final third, which was 10% more than the next highest in the league (Dominik Szoboszlai).

In view of these boosts, how remarkable was BF’s underlying performance when adjusted for opportunity? And which other playmakers are worthy of note (who may not have been given the same opportunities)? 

Set pieces are undoubtedly important. Arsenal won this year’s EPL on the back of them, and being a dead ball specialist, such as James Ward-Prowse (marked in the right scattergram with an “X”), can certainly improve an otherwise nondescript CC tally: 5 in open play, augmented by 18 from set pieces. But most players don’t get a chance to take free kicks or corners, which obviously skews the statistics in favour of those who do. Set pieces also require a different skill set to open play creation. Consequently, we restrict our attention to CC in open play, and leave set pieces for another stage.

The left scatterplot did the rounds in social media and TV. But reworking the data as in the right scatterplot is more instructive. BF (●) had more CCs both in open play. Next nearest was Enzo Fernandez (●), who had his own boost of even more minutes than BF. BF also gained more CCs from set pieces than anyone else – next nearest was Dominic Szoboszlai (●). 

This graph shows the smoothed distributions for all 9 leagues. What Tukey called far outliers are marked as filled black circles; they are defined as lying greater than 3 x IQR above Q3 (≈ 3.5 x IQR above the median) – similarly for far outliers below the median. These are much more extreme than the usual outliers found in Tukey boxplots, defined as lying 1.5 x IRQ above (below) Q3 (Q1). Because leagues have different numbers of games (max = 46, min = 34), the raw, expected, and excess CCs will naturally differ in magnitude. Measuring Excess CC in league-specific IQRs units, as in the graph, allows us to compare across leagues. Evidently Bruno Fernandes (in red) is the most extreme outlier in all leagues. Although his team certainly gives him more opportunities than most to create chances, once this boost is adjusted for, BF is still revealed to be exceptionally efficient at turning those opportunities into chances. In the next section are the top 10 players in each league, measured by Excess CC. Far outliers are highlighted in green.

The top half of the table (green) are the regression coefficients, and the bottom half (blue) are the corresponding t-values. The coefficients have been scaled for easier viewing. They are to be read as follows: if a player competing in the EPL (English1) plays for 1000 minutes, makes 100 passes in the final third, makes 100 crosses, and makes 100 crosses NOT in the final third, then we expect him to create:

3.601 + 5.259 + 4.500 – 1.251 + 0.807 = 12.915 chances

If instead he makes 200 passes in the final third, we expect him to create an additional 5.259 chances, and so on. Note that minutes per se do not have much bearing on chance creation, and that the more passes not in the final third, the fewer the chances we expect to be created. R2 goodness of fit for the 9 leagues models were: 0.816, 0.859, 0.799, 0.785, 0.822, 0.859, 0.826, 0.824, 0.837, in turn.

The coefficients (highlighted in green) across the leagues are remarkably similar. To appreciate quite how similar, suppose we mistakenly applied F1’s regression coefficients to E2 players, for instance. How much difference would it make to CC estimates for each player? The table below shows all the Pearson correlation coefficients of expected CCs when a ‘wrong’ set is applied, correlated with expected CCs using the correct set. Specifically, row 2 column 6 answers our question. Wrongly apply the F1 coefficients to E2 players and the resulting estimates have an extremely high correlation of 0.9903 with the correct estimates (E2 weight applied to E2 players). The other boxed correlation in the table is the vice versa: wrongly applying E2 weight to F1 players. Unlike in a standard correlation matrix, these figures differ slightly. 

The conclusion is that model coefficients are highly interchangeable between leagues, with the possible exception of F1 with E4 (highlighted in yellow). Highest coherence is between E3 and E3a (pink highlights), as we might expect, E3 and E3a being the same league in consecutive seasons; but also between I1 and G1 (blue highlights). However, these are mere nuances. The very high correlation throughout the table suggests we could construct a single ‘super model’ of how chances are created that would serve almost as well as any of the 9 league-specific sub-models. A quick and dirty way would be to average, column-wise, the coefficients in the table above. The very great similarity of E3 with E3a coefficients suggests that a one-size-fits-all set of coefficients might also work well, if not in perpetuity, then at least for a few seasons.