Contest! Rate Player Attributes and WIN! (13/03/2014 03:36)

Why the heck are you blogging when you should be SLAVING OVER THE NUMBERS MACHINE??

- All the managers still waiting for their analyses

Exactly how important is each attribute to performance?

---

Introduction

OK, so we all know how important the attibutes are to Quality - for each position it is a simple weighted sum of attributes. Specifically:

 

  

 

You can find this in all the best guides. Very simply, the quality rating for each position relies mainly on the primary attribute, and to varying degrees on four others. So increasing a defender's tackling by +11 would increase his quality by +4, but increasing his perception by +11 wouldn't increase his quality at all. Nice and simple! Except, of course, that it isn't. As anyone who's played for a while knows, during a match players can end up using lots of different attributes that don't count towards quality. Your midfielders will perceive, your attackers will tackle, and every now and then your defenders will get a chance to shoot. Plus of course, Spinner has updated (i.e. complexified) the sim since initially coming up with those simple Q formulae, and has stressed before that actions are determined by realistic sets of attributes, rather than just a simple check against, for example, tackling. Plus if you actually sit down and read a match report - really read it - you'll notice that while there are a bunch of events that seem quite clearly tied to particular attributes ("He shoots!"), others like controlling the ball are more ambiguous, or refer to skills that don't seem to correspond neatly with a single attribute.

The upshot is that (for this and a whole bunch of other reasons) you shouldn't only think about a player's quality - that's referred to as "Q blindness". Instead, you should also take into account other apparently unimportant attributes, as well as factors like performance in previous matches. Anyone who's tried to sell a midfielder with low perception will know that it's highly valued, even if it doesn't contribute directly to quality.

So what is the real "value" of an attribute for a particular player? That's what we're all going to guess, and I'm going to try and measure. The person who guesses closest to what I measure wins 5 credits or, if they prefer, a Luxury analysis for Season 112. Each position is its own separate context so there are 4 credit-or-analysis prizes available. Finally, the person who gets closest overall without being top of any single position gets a 5-credit-or-analysis prize too, so that's 25 credits or sevaral hours of my life on the table. Needless to say, the kudos for demonstrating your deep knowledge of the sim will be the greatest prize of all. Some ground rules:

 

- We're going to define this value in terms of how well a player performs now. We don't care about his potential or resale value - basically if you had to pick players for one single crucial match, which attributes would you want to be high, and which wouldn't you care about? What do you think the match sim cares about?

- We're going to ignore hidden attributes.

- Of course tactics will affect how often different attributes are used, so we're talking about the 'typical' game. Specifically the average of about 10,000 matches in a bunch of Q80 departments that I picked out semi-randomly for this analysis.

- And when I say we're looking at performance, I literally mean the performance measure the sim gives us in match reports. For better or worse, we're going to take this as a reasonable measure of how useful a player is during a game.

 

OK, so here's the contest. I've taken average season performance ratings for ~5000 players in some moderately strong departments, in a couple of different leagues. Between them these players have made over 90,000 league appearances: Hopefully this is enough data to smooth out some of the individual variation in things like hidden attributes. I've picked departments that are all within 2.5Q of Q80, and which have few bot or zombie teams in them (and of course I excluded any players from bot/zombie teams too). I've put all that data, together with each player's 8 attrbute values, into a big linear model for each position - GK, Def, Mid, Att - to find out how strongly each attribute's value predicts performance. So now I have a matrix very similar to the one above, but derived entirely by looking at the actual performance of players in games. Can you guess what it looks like?

I'll give you some clues. In some ways, it looks quite similar to the formula for quality. In other ways, it looks very, very different. All of the values are positive or zero, i.e. higher attributes are never associated with worse performance (makes sense, right?). Also, I'll tell you for free that Tk, Sh and He contribute 0 to a goalkeeper's performance, and Kp contributes zero to all outfield positions. Actually, because I gave you so much for free, I'm going to make you guess something extra: I'm including players ages as a variable. Do older players perform better than younger ones with the same attributes? And if so, how important is age compared to all the other stats?

---

 

How To Enter

To enter, just leave a comment on this blog with the proportion you think each attribute - and age! - contributes to performance. So if you thought that age had no effect, and the Q formula above was the true answer, you might type something like:

(Pos) (Kp) (Tk) (Pa) (Sh) (He) (Sp) (St) (Pe) (Age)

GK 36 0 9 0 0 18 9 27 0

Def 0 36 18 0 9 18 18 0

Mid 0 18 36 18 0 18 9 0 0

Att 0 0 9 36 18 18 0 18 0

 

So long as your formatting is understandable, it'll be considered a valid entry. Don't worry if percentages don't add up to 100% either, I just need the relative proportions. I'll give you all until League game 9 has been played, so that's 15:00 CET next Friday, March 21st. At that point I'll post & discuss the results of the analysis, measure how close everybody got (Euclidean distance, since you ask) and dole out some prizes :-)

Good luck!

 

- Belizio

 

 

Bonus technical details if you're interested (you don't need to know these to enter the contest):

- The analysis was a simple multiple linear regression on the 8 attributes plus age, predicting performance. A separate regression was run for each position. Certain attributes were excluded for each regression - the 3 outfield regressions did not include Kp, and the Goalkeeper regression did not include Sh, Tk or He. 

- I used average season performance but weighted each player's contribution to the model by number of appearances, so the analysis is equivalent to one run at the match level.

- All coefficients were assumed to be >=0, in other words increasing in an attribute (or age) should not reduce performance. This was achieved by removing attributes with small negative coefficients from the regression, and re-running with the reduced set - continuing as necessary until all coefficients were positive. In total 24 of 36 possible attributes had a positive effect on performance (a bonus clue for reading this far!).

- Because age is on a different scale (17-40ish) than attributes (20ish-99) I z-transformed everything (attributes/age/performance) before running the regression. That just means everything is on the same, intuitive, scale: So for example, if age is more important than shooting for some position, then being in the top 10% for age but average for shooting should give you better performance than being average for age and in the top 10% for shooting.

 

 

  

Click Here to return to the Wonderstar Analytics home page