1 of 22

Data-driven evaluation of soccer players performance

Paolo Cintia - University of Pisa

2 of 22

{'eventName': 8,� 'eventSec': 8.221464,� 'id': 217097515,� 'matchId': 2576132,� 'matchPeriod': '1H',� 'playerId': 8306,� 'positions': [{'x': 42, 'y': 14}, {'x': 74, 'y': 33}],� 'subEventName': 83,� 'tags': [{'id': 1801}],� 'teamId': 3158}

pass

high pass

accurate

1700 events

per match

(in average)

identifiers

Luca

But how this information is collected and stored. Well, using semi-automatic technologies, Wyscout collects all the events in a game which involve the ball, storing them in JSON format.

This is an example of soccer event data, consisting of a Python dictionary-like structure providing useful information about the event:

the type (a pass in this case)
the timestamp of the event (that is, when the event occurred)
the match, the team and the player identifiers associated with the event
some tags describing meta-information attached to the event.

This event describes a pass made at the 8th second of the first half of the game, in a specific position of the field. The tags tell us that the pass is accurate, that is it accurately reached the destination player.

In average there are 1700 (one thousands and seven hundred) of these events during a soccer game. Given the complex structure of the event object, the natural way to store it is to use a nosql (nosequel) database. Indeed, we choose to store the data in a MongoDB database.

3 of 22

Luca

Fortunately, the Internet of things and the data revolution help us answer this question. Indeed, nowadays data describing many aspects of the performance of players are available in massive quantity, collected from a variety of sensors, devices and technologies.

This is, for example, the visualization of all the events Lionel Messi produced during a game in the Spanish La Liga. In the figure, you see that Messi performed 47 passes (the red arrows), 5 shots, 15 tackles, 4 crosses and so on.

The aggregation of these events into numeric variables constitutes the “performance” of Messi in that game, that is a set of variables describing different aspects of Messi’s game. In this case, we have shown the simplest way to extract a performance: counting the number of events of each type generated by Messi. Of course, more complex variables can be extracted from the events, describing for example the dangerousity of a player, his contribution to the team and the accuracy of his ball possession.

Technically, this visualization comes out from a manipulation of raw soccer data collected by the company Wyscout, the leading company for soccer talent scouting.

4 of 22

Ranking soccer players

5 of 22

Multidimensional and flexible�existing metrics are based just on passes or shots (e.g., FS, PSV)

Characteristics of the method

PlayeRank

Dynamic and role-aware

compare apples to apples

Validated

existing metrics are validated just against goals or assists (!)

6 of 22

soccer

logs

Feature weighting

Role detector learning

Player Rating

Players

Rankings

Individual

performance extraction

feature

weights

Role detector

Learning

Rating

batch

online

𝜇

(a)

(b)

(c)

b1

b2

c1

(d)

Ranking

d1

d2

c2

7 of 22

Feature weighting

Paolo

With training data we can use the Weighter class. For those of you who knows sci-kit learn, this code sounds familiar. Weighter has the same structure of sci-kit methods. The fit method build an Support Vector Machine that, given the feature of a team, predict the final result of the game. Once fitted the SVM, we can analyze the class property “weights_”. It contains a vector indicating the weight for each feature on deciding the final result, as you see in the bottom plot. Note that we are not using goals, because it’s trivial to notice the goals are highly correlated with the result of a game. Actually, goals are the results of a game. Furthermore, we are not interested on goals scored now: even a commodore 64 is enough to rank players by goals scored. But the majority of the players are not scoring any goals.

From the picture we see the most important features, both positive and negative: so a red card foul is the worst thing you can during a game, while passes leading to scoring opportunities, have the most important positive impact on a game.

8 of 22

Role classification

9 of 22

Rating computation

performance rating

of u in game g

taking into account the number of goals

10 of 22

18 competitions

30 million events

20K matches

21K players

Experiments

11 of 22

Best players in the dataset

12 of 22

Evaluation of PlayeRank

We randomly extract pairs of players from the rankings
Ask professional scouts who’s the best for each pair
Ask PlayeRank who’s the best for each pair
Compute the agreement between the answers

Compare with state-of-the-art:

FC → Duch et al., Quantifying the Performance of Individual Players in a Team Activity, PLoS One
PSV → Brooks et al., Developing a Data-Driven Player Ranking in Soccer using Predictive Model Weights, SIGKDD

13 of 22

Evaluation of PlayeRank

14 of 22

Evaluation of PlayeRank

15 of 22

Evaluation of PlayeRank

16 of 22

Evolution of players

17 of 22

Evolution of players

18 of 22

Patterns of performance

19 of 22

Versatility of players

20 of 22

Versatility of players

21 of 22

https://arxiv.org/abs/1802.04987

22 of 22

Player ranking

The ranking of players (by role) can be computed by aggregating over all ratings of the players