Data revolution in soccer as well?

[Borut]

I am reading Football Hackers by Christoph Biermann at the moment. The use of data analytics has been very pronounced in basketball and baseball, but not yet so much in football, for example the Basketball forum has its own section for data analysis. Why would you guys think that is?

I believe football is a more complex sport so it is harder to operationalize the qualities. Basketball has a larger tradition of tracking useful data.

[Foye]

Honestly, there‘s probably more analytics in the background of football than there is for basketball.

Teams do everything from scouting, to analyzing all measurable data, tracking player health.

You can be sure that each team buying player xy for millions has seen every game from the player in the last few years and put focus on statistics such as distance covered, passing percentage, goals scored per shot, assists, tacklings, headers won, body strength etc etc.

The difference being these statistics are not so interesting for the common viewer compared to basketball.

[coldfish]

There is an article floating around the internet about how the owners of Liverpool are numbers guys from baseball and they have taken the same approach. They hired Klopp based on metrics and several of their buys were based on numerical analytics, like Salah and Keita.

The big difference right now is that the methodology is top secret. I suspect a lot of teams are doing it but they aren't advertising what exactly they are doing.

Here it is:
https://www.nytimes.com/2019/05/22/magazine/soccer-data-liverpool.html

[coldfish]

The methodology is really different from basketball or other sports from what I have seen. On/off numbers don't mean the same thing in a sport where you only get 3 subs. If you read the article I posted, its almost like golf analytics where you assess each touch and quantify the value created. There is also a layer for off ball movement and I have no idea how they put numbers to that as far as its value but I do know its not just distance covered. They are looking for players who move into space in a threatening position, not just jogging around aimlessly.

[madmaxpeake]

I think there's room for a similar data revolution but maybe not at the same level as in other sports. I've always felt that stats say a lot less than "the eye test" in this sport.

[HIF]

Football would benefit greatly from better use of digital and data analysis. I think Sousa at Bordeaux is really looking at improving this area of the club and I've thought for many many years that prem clubs could be more efficient if they hired some of the basketball video guys to adapt their skills for football.

[freethedevil]

I think there's room for a similar data revolution but maybe not at the same level as in other sports. I've always felt that stats say a lot less than "the eye test" in this sport.

Nonsense. Analytics are very much present in soccer here, people just aren't knowledgable enough (likely because they're hard to find publicly)to make common use of them:

Watch on YouTube

[SgtPepper]

It's true motion tracking data is walled much more and that some unique complexity is added from the sport being low-scoring, but the whole analytics ecosystem is more immature than other sports such as basketball or baseball. In those sports the leagues or national organizations themselves provide predictive metrics and the interchange between bleeding edge sports theory and the data science is much smoother. Just look at how many major analytics folks in soccer roll into major appointments in the industry vs other sports. In basketball we have guys like Morey, Goldsburry, Partnow who have made major contributions both within the league and to public research in the sport. In soccer most of the guys within industry are silo'ed in clubs and the few who are consultants whether in firms or as independents spend their time pecking at each other on twitter or posting tactical marketing pieces rather than doing a real bake-off.

[andyhop]

Football started using player tracking technology a long time before basketball did.

https://www.wired.co.uk/article/premier-league-stats-football-analytics-prozone-gegenpressing-tiki-taka

A small indication of how analytics work in the premier league and its history

[SgtPepper]

^The research exposed to the public at conferences show that bball and American Football are well ahead of the curve comparatively. Also Luke Borrn is the leading name in soccer motion tracking analysis and most of his stuff came adapted from basketball.

[Borut]

Interesting to go back to this thread. I just wrote a blog about my own application of data science in soccer: https://medium.com/@borutflis1/application-of-data-science-in-soccer-eadfc9bd6da

[SgtPepper]

Belated, but thanks for sharing. I think the post had 3 issues:
1) The main unique challenge with football compared to other sports is how low scoring it is. That means there's far less room to work backwards from the outcomes in mapping the process of gameplay, and more innovation and speculation is needed upfront to cover some underquantified things like zonal efficiency especially in defense.
2) As a result of #1, PCA backed out from outcomes usually doesn't cluster players in a useful way since the initial parity within the dataset is too small and requires convolution. A more correct technique is DBSCAN.
3) The 'clusters' you drew in that team scatterplot graph is invalid from a mathematical or data science definition of that word. You can see that just via the eye test because there is no separation between the colored clusters. You can also formally demonstrate the invalidity by seeing that the clusters do not separate by any of the classic cluster separation conditions (whether centroids vertex/knn, error, etc).

[HIF]

Belated, but thanks for sharing. I think the post had 3 issues:
1) The main unique challenge with football compared to other sports is how low scoring it is. That means there's far less room to work backwards from the outcomes in mapping the process of gameplay, and more innovation and speculation is needed upfront to cover some underquantified things like zonal efficiency especially in defense.
2) As a result of #1, PCA backed out from outcomes usually doesn't cluster players in a useful way since the initial parity within the dataset is too small and requires convolution. A more correct technique is DBSCAN.
3) The 'clusters' you drew in that team scatterplot graph is invalid from a mathematical or data science definition of that word. You can see that just via the eye test because there is no separation between the colored clusters. You can also formally demonstrate the invalidity by seeing that the clusters do not separate by any of the classic cluster separation conditions (whether centroids vertex/knn, error, etc).

Sounds good