The unreasonable effectiveness of Soccermatics? (2017)
interaliamag.org
interaliamag.org
I'm the analyst mentioned in the article. Happy to answer any questions about our weird little industry.
I stumbled across some high resolution soccer data during college and began writing a blog that became popular in soccer analytics circles. The company that produced the data that I was scraping eventually hired me to their data science team. I spent a few years with them before I was recruited to my current team.
I have never worked in gambling but there are a few people in my sort of role that have that sort of background. It requires a pretty similar set of skills. I am not sure that I could beat the market by a large enough margin to make it worthwhile. It's quite efficient. But for the most part, I'm more interested in understanding the underlying mechanics of tactics than I am interested in predicting the result of games.
But, I would read "The Numbers Game: Why Everything You Know About Soccer Is Wrong" and read the backlogs of the StatsBomb blog. That will get you up to speed pretty quickly.
I work with a team's coaching staff and management to help them make more data-driven decisions. This ranges from topics around opposition analysis to player recruitment.
Anyways, what sort of things do you look at for defensive players? It seems that its when I look at things such as WhoScored's statistical team of the season, it has players such as Mustafi, who generally has a negative reputation for his play. I suspect he is so high because the rating metric used by whoscored overvalues offensive contributions of defenders, vs. pundits more likely look for a defender's defensive contributions. Are there any form of 'advanced stats' for defenders beyond the basic measured stats of challenges, interceptions, etc, that you and/or the industry looks for?
The first problem is the data. The soccer viewing public is largely familiar with event-level data, typically provided by (my previous employer) Opta. They've done a great job normalizing soccer statistics on the cultural level, but the information they've collected at scale isn't that useful for creating good defensive metrics.
Other companies have sensed an opportunity here and have started providing more detailed data around things such as defensive pressure. Suddenly, you can contextualize each offensive event with the level of defensive pressure applied to it. I think this will be a game changer, but we're in early days there.
Other companies provide player-tracking solutions that give you real-time position of all players on the field. This is great because you have a "complete" picture of the game, but it requires a lot of work to build more sophisticated spatial/geometric models.
There's also the "Howard Effect", coined largely as a basketball term, but it's similar to the Maldini example you provided. Some defenders are so good that they don't have to be "active" defenders. That's something which is really difficult to adequately control for.
Statsbomb is a little different. They've started collecting their own data and are offering some free data from various Womens leagues.
You can do the exact same thing in soccer, if you have the data: You can assign responsibilities to players, just through computer vision. If Messi gets 3 touches in an attacking position when Sergio Ramos is defending him, you can credit that to Ramos, and compare that to Messi's touches vs the average Barcelona opponent.
Packing is the measurement of how many opposition players are beaten by a pass (or dribble or other move).
Impect is the number of deep lying defenders beaten, which is obviously more valuable compared to high pressing strikers.
The key insight is that defenders beaten, particularly deeper-lying defenders are the measurement needed to identify who is expected to win (and therefore assessing the value of passing players, or defenders).
http://bundesligafanatic.com/20160610/impect-packing-the-fut...
> Impect disregards all passes that go backwards or don’t beat any defender. It’s a correct assessment, a pass that doesn’t travel forwards, doesn’t help you score. Can’t argue with that.
This is a really odd statement. Plenty of passes don't go forward but absolutely do increase the likelihood of scoring. For example, the final pass here [0] is backwards, might even allow an additional defender time to track back (-1 packing?) but opens up the space required to take a shot and score.
Useful backward passes fall into at least two categories. The ones that eventually result in an impect, and the ones that run down the clock at the end of the game.
Yes, that's basically what I'm saying is bad about the statement that backwards passes don't help you score goals.
> Useful backward passes fall into at least two categories. The ones that eventually result in an impect, and the ones that run down the clock at the end of the game.
I think this is almost correct, but I think there is a third category: Backwards passes that immediately result in a shot resulting in a goal. In the example video I posted above, the final pass is backwards and is immediately shot into the net. This doesn't increase impect (ie: Messi dribbles past 4 players, then shoots and scores past 3 more, is only impect +4, according to the article).
I'd guess that either "Impect disregards all passes that go backwards or don’t beat any defender" would be better phrased as "Impect disregards passes unless they go forward or beat defenders" or the statisticians would defend the original solely in terms of it being a heuristic that makes the problem more tractable.
Additionally, your linked example, with #25 picking up the ball in a danger position after a deflection, is the kind of "penalty box slop" that, naively at least, seems extremely difficult to handle analytically.
> A “shutdown” cornerback like Richard Sherman can be a star in the NFL thanks to interceptions, broken up plays and tackles. “Lock down NBA defenders” like Bruce Bowen and Dennis Rodman can prove their worth with steals, blocks and rebounds. Football has goals and assists, that’s it.
Just as in soccer, those "counting stats" are not great measures of defensive ability. The recently retired cornerback Darrelle Revis was regarded as the best defensive player in the NFL from about 2009-2011. Yet he did not rack up interceptions or pass breakups -- in fact, in 2010, he had 0 interceptions and only 9 pass breakups over 13 games. Why? Because the receivers he was covering were never open, so quarterbacks rarely attempted to pass in his direction.
Similarly, steals, blocks, and rebounds are only a vague indicator of defensive ability; it's never a bad thing to get a steal or a block, but if you routinely leave your man to try to poke the ball away from someone else's man, you're likely hurting your team overall. The NBA has been working on developing better stats, including deflections (you get your hand on a pass but don't necessarily come away with the steal) and shots defended (you're within a short distance of a shooter). But Darryl Morey, general manager of the Houston Rockets and a well-known stats nerd, has said in a Reddit AMA that no publicly available defensive statistic is useful.
Part of this is because all three sports in question are team sports, which means a great deal depends on the defensive scheme. You may not block shots, but is it your job to block shots or to stop the ballhandler from getting near the basket? You don't get tackles, but is it your job to get tackles or to funnel the ballcarrier right into the linebackers?
It's extremely difficult for any outsider to determine the defensive effectiveness of a player. The only thing we can offer is guesswork.
Sure, you have people who write a ton of code and implement a lot of features, and fix a lot of bugs, who clearly are contributing a lot. But you also have people who, though other means (code reviews, refactoring and other code-health work, etc) ensure that a project is maintainable and sustainable.
How do you measure the value of 100 bugs that never made it to production because of high quality code-reviews? Or those 5 high-value features which were a snap to implement because somebody took the time to clean up all the cruft from Mr Rockstar Bro who made a gigantic mess?
Of course there's also stuff like GoalImpact which just tries to apply plus minus to football, but all the old arguments about football being a low scoring game apply here.
This is what the Italians really do well with their "verticalizzazione" (I guess "verticalization" would be the closest English translation, even though I'm not sure that's a word yet), i.e. deep passing when you're (usually) 30 to 50 meters outside the opposing post. For an excellent example see this YT video (https://www.youtube.com/watch?v=wWr1tTFt7a4) of Sarri's Napoli doing it (Sarri is now Chelsea's coach, that should be interesting).
The key insight. It kind of reminds me of how physicists like Robbert Dijkgraaf (I think) have argued that physics and other fields are now contributing to mathematics through their own outlandish needs for weird maths.
For example, from your quoted section:
> Mathematics is not there to be discovered, it is part of the patterns of reasoning in all of our brains.
Isn't that the very definition of elegance? He later says, talking about A.J. Ayer's definition of "non-sense":
> Wigner freely admits that his idea about maths comes from a feeling that can’t be verified by known scientific methods
Elegance, while there is often a wide collective understanding of what it means in mathematics, is essentially a subjective, aesthetic property; an intuitive one. Is this not exactly what:
(a) Wigner means in that quote,
(b) A.J. Ayer means in his definition, AND
(c) the author refers to as "patterns of reasoning in all of our brains"?
We also like symmetry but symmetry is just a way to identify attack surfaces for compression.
[1] https://www.edge.org/conversation/david_deutsch-constructor-...
(DDG !libgen soccermatics)
Ah, looks very good!
I had the pleasure of attending the mentioned colloquium by Sid Redner back in 2015 and I was absolutely blown away by the effectiveness of a simple random-walk model.
This would seem to be consistent with the notion of "common sense" that it's possible even for an average person to be able to ask meaningful questions and cast doubt on subjects they're not specifically trained in.
from a blog post, shows passing networks and has some vector fields overlaid on the soccer pitch. https://www.fourfourtwo.com/features/soccermatics-how-mesut-...