There are many problems with reducing complex sports, events, etc a into simple numbers and models. A lot of the nuances might get lost in statistical methods if one does not carefully think about causation and correlation. Also our outputs are usually linear, which tells us A>B>C just like EPL Standings. The second issue has bothered me for a while; and last week I came across a real life example from clubelo to demonstrate it.
When I(or many others) come up with models and ratings to we tend to dismiss the fact that the sport is being played among two teams in any given game but focus more on what will happen after the end of the season. This makes sense;one game is difficult to predict and does not reveal much. In contrast a season is long enough for mean reversions, trends, etc to work through. Also we can always adjust for strength of schedule if we think fixtures were biased. But this still leaves something out. Take 3 teams with similar looking strength ratings in elo, oddsmakers, dingoR, etc. On paper they might look the same, but they might have different styles:
Let’s name our teams
Yes, we are going to play papers, scissors, rock. (Dingo Jr.’s stroller color was chosen by p/s/r. Mr.Dingo beat Mrs.Dingo 2-1, but Mrs. Dingo bought the one she liked anyways. For a more interesting p/s/r story read this.).
Going back to soccer, these teams might look like similar strength from their games with other(weaker teams) and when they play each other P always beats B, B beats S and S beats P. In ELO, the team which wins the last game would have the highest ranking but we know theoretically that they have same strength. Let’s move to a more complex situation.What if a league has 3 Scissors, 4 Rocks and 1 Paper In that case, league table would like this:
P 12 pts
S 5pts each (3 points from beating P, 1 each from other S games)
R 6pts each
Is P really that good? Well yes because somehow they are employing a style that dominate, but can we move P to Champions League or any other league that are full of S’s and P’sand expect same success? Not really. Of course real life is much more complex. I have no idea how many styles there are(or even if the style issue is relevant). Teams can play different players and tactics according to opponents. For example, managers can employ lesser but faster players just because they have a better chance of beating a better but slower team.
Assigning styles to teams is not easy. How those styles of play have developed can be more difficult.Was it a conscious choice? Did a manger or an executive decide that a league had too many “rocks” therefore chose to build a “paper” team? Is that what Pulis does? Do pundits have point about “Barca playing Stoke on a wet Tuesday night”? As I leave you with all these questions I want to move on to an example I came across last week while writing the Turkish History post.
Go to this link. It shows number of ELO points gained/lost by Turkish clubs versus specific countries. The list on the right is shorter; anything strikes you? Not one country on the Mediterranean; almost all to the north of Turkey. So I thought it has to be height; taller teams usually do better versus Turkish teams. I plotted height versus average ELO score gained. )You can find all kinds of demographic data here.) Yes,there was -41% correlation between avg. height of a league and ELO points gained for Turkish teams. Yet; as I was entering the data I realized there were many countries that were tall but did consistently bad against Turkey- especially from Central Europe/Balkans. So instead of height I used geography as input- used the following groupings(other than Turkey/Greece group I wanted the other 2 to be of same height.Also relatively speaking these leagues had rather similar strengths )
Paper: Turkey/Greece (Avg Height 181 cm)
Scissors: Czech-o-Slovakia (Avg Height 182.7)
Rock: Norway/ Denmark (Avg Height 182.6)
|All Time ELO|
|All Time ELO/Game|
Turkish and Greek teams do terrible against Scandinavians but do very well against Central Europeans. Yet Scandinavians tend to lose points consistently against Central Europeans.Considering avg point exchange in European Cups seems to be slightly higher than 5 (thanks @clubelo), these look significant (don’t shoot me if it fails some null hypothesis test- I am not going to throw out all my history books because conclusions they reach are not statistically significant;)). It is not height but how you use it that seems to matter.
I am not an expert in different countries’ tactics, but I guess variance in height(Scandinavians tend to employ really tall defenders, and therefore might dominate in set pieces) or other factors would be very interesting to look into.
As a next step I will try getting correlations among all countries and see if I can group them into certain categories. Then would come intra-league styles which will prove to be most difficult.
To conclude; I really like Simon Gleave’s (@SimonGleave) ISG coefficient which basically “compare(s) the results this season with those achieved in exactly the same fixtures as last season”. Read about it here. I think it captures both strength of schedule and style of play in one glance. Of course there is randomness and noise in all of this and maybe I am being fooled by those. But’ I think this is something real and if you have similar models/anecdotes/presumptions let me know @dingosports and we can try to quantify.
Good news for Galatasaray fans then! On average Turkish teams grabbed 1079 points in 62 games versus English teams.(English are the best oppopnent to play against for Turkish clubs; and Turks are the worst country to play for English).