Friday, 31 August 2012

F.SCORE - what the F. is it?

UPDATE:  03-Sep-2012, I've added appearance points to the F.SCORE.  Made sense.  

F.SCORE is my Fantasy Football ranking.  I imagine it is not too dissimilar to under the bonnet of the awesome Fantasy Football Scout's ICT index.

At the moment I have it set up to use the Fantasy Premier League's Scoring System which is detailed below, but it's easy enough to do for other games too.  I'm playing the Classic game with work so draft scoring system for this game too.

FPL Scoring
You know this already but here we go.

Clean Sheet - 4 pts Defender, 1pt for Midfielder
Goal - 4pts for a Forward, 5pts for Midfield, 6 for Defender (or Tim Howard!).
Assist - 3 pts

Appearance points and the rest are not factored in, yet, and may never be, 

Based on the above FPL scoring the F.SCORE for the key footballing actions is calculated as follows:

-Shot on Targets (F.SoTf)
Only shots on target are counted, shots off target are not.  The correlation between shots on target and goals is better than just all attempts so I am working with that for now.  Top level football players  score 1 goal per 3 on target (ratio 1:3) , on average, or so I believe.  Therefore, 1 shot on target is worth 1/3 of a goal, or 1.33 for a Fwd, 1.66 for a Mid & 2 for a Def.   

The awarding of bonus points (BP) in FPL is ridiculously correlated to goals scored, and at some point last year I worked out that on average a goal accrues you 1.2 bonus points, so I have thrown  an extra point in there.  I'm sure there's a better way of doing this, for example, if a defender scores and keeps a cleanie they are very likely to get more than 1 BP, but that's more future work, as is biasing the ration 1 goal:3 SoT for different team's or player qualities.

SoTf = 1.66 (FWD), 2 (MID), 2.33 (DEF/GK)

- Key Passes (F.KeyP)
Much simpler.   An assist is worth 3pts in FPL.  A Key Pass is pass that leads to an attempt to score.  This stat doesn't indicate if this attempt was on target or not, so I am using another established average at the top level of the game, 1 in 9 attempts on goal result in a goal.  Using this value, 9 key passes is good for 1 assist, so 1 key pass is worth 3/9 = 0.33 pts.

F.KeyP = 0.33

- Clean Sheet (F.SOTa)
Tough one.  I have tried various means of scoring this one, including complex calculations involving Poisson distribution but for now am taking the same principle for points scoring from goals and shots on target but reversing it.  The logic is that a team concede 0 shots on target (SOTa) they will concede 0 goals, and hence a clean sheet and 4pts.  1 shot on target conceded would give a 33% or chance of conceding a goal, or 67% chance of not conceding, a defenders F.SCORE for 1 SOTa would be 4*(67%)= 2.6.  2 F.SoTa is 4*(67%)^2 = 1.8pts, and so on.  

Still with me?  Good.  Works the same for midfielders but multiplied by 1 instead of 4 of course.  I'll admit this probably undervalues clean sheets potential.  Without using a calculator, a team conceding just 1 shot on target all game is very likely to get a clean sheet, and I think it will probably be quite easy to adjust this particular factor with some more analysis.  Further work ;)

1 F.SOTa = 2.64 (DEF/GK), 0.67 (MID)
2 F.SOTa = 1.80 (DEF/GK), 0.45 (MID)
3 F.SoTa = 1.20  (DEF/GK)  0.30 (MID)
4 F.SOTa = 0.81 (DEF/GK), 0.20 (MID)
you see where this is going.......

-Appearance Points
Simplest of them all.    Play > 60 minutes = 2pts.  Play less than that = 1pt.

-Rounding Up and Smoothing
I am not all that sure in the pure value in doing this, but it is aesthetically pleasing to me and I hope provides a more user-friendly figure.  One of my tiny criticisms of FFS's ICT index is the decimal point.  I guess it just appears too precise a number and feels a little inaccessible.  After adding up the above scores I multiply the figure by 10 and then round the result to it's nearest integer to give me the final F.SCORE.

(F.SoTf + F.KeyP + F.Sota) * 10 = F.SCORE (rounded to nearest integer)

For an example or two:
Rooney has 1 shot on target = 17 pts
Bale has a shot on target =  20 pts
Baines creates a chance = 3 pts
Huth / Stoke concede 2 shots on target = 16pts 

Essentially, it's the average or expected FPL points * 10.

An Arbitrary Evaluation
After running some real data through all this I arbitrarily and subjectively came up with the following scale to bucket an F.SCORE

F.SCORE = 0-29   : Not too special
F. SCORE= 30-49   Pretty Good.  Well played
F. SCORE = 50+     Special.

Does this work?
Let's see.  It will continue to be assessed and tweaked with lessons learnt through the season but it looks good to me as a starter for 10. 


  1. You have assumed that 1 goal on an average, gives 1 bonus point.

    So, 1 F.SoTf should give 1/3 bonus points. So your F.Sotf points allocation will be:

    SoTf = 1.66 (FWD), 2 (MID), 2.33 (DEF/GK)

  2. Yes, you are right. Thanks for such an acute observation. Will update the parameters of my model.

  3. Also, I'm bumping BPs up to 1.5 points on average/goal and 2 for DEFs.

  4. Actually the model's parameter were correct and did not need adjusting. Just this write up was wrong.

  5. Hi 'Shots on Target,'

    Have to say read this and thought nice!
    Interesting stuff, although lots of ratio presumptions, which are critical. but least you are a consistent being.

    1) Where did you get this info from, i.e.
    - 'Top level football players score 1 goal per 3 on target (ratio 1:3) , on average, or so I believe.'
    - 'Key passes - 1 in 9 attempts on goal result in a goal.'

    2) Where do you actually get the data itself from. i.e.
    - The number of shots on target in a game per player
    - The number of key passes in a game per player

    3) I think you are definitely undervaluing clean sheets, by multiplying likelihoods for 'chance of not conceding' you are using defining distinct probabilities the chance of a clean sheet are in fact less. Have a look at the Newton Dice Probabilities should help.

    4) You are also applying same likelihoods for players at different clubs i.e. likelihood of top 4 team conceding vs. bottom 4 teams. Shouldn't this be taken into account, or are you suggesting there is there a need to have independence here and treat each player an individual, perhaps grouping the prem league in 4 or 5 levels/ group of teams hence giving rise to homogeneous sets?

    6) Would be interested to see an article on the mentioned Poisson Modelling, the key is obviously with the likelihood, lambda.

    FYI - might want to change '< 60' to '>= 60'

    Keep up the good work.


  6. skip of the tongue.
    *chance of a clean sheet are in fact more.

  7. Hi Matt - you wouldn't be Matt who did this awesome article on FPL Dugout are you?

    Thanks a bunch for some really insightful comments on this metric.

    I'll lot you know the above is version 1.0 of a very much more spohisticated model I am working and hope to have complete and refined for start of 13/14 season.

    The starting point values I've have got from opinions formed on a great website called Soccer By the Numbers which unfortuneatly seems to have ceased to be active. I've also analysed the stats myself. I've summed some of the important values in a post a few days ago called VEry Important Numbers.

    I have already built in a variable rate for conversion to my model, which is not reflected in the above description. I hope to update it soon. It essentially uses a players underlying stats to determine their expected shot:goal rate. It works quite well and I will publish the results once I have fleshed it out a bit more.

    The data itself is from a few places. has good data for the basics - S, SoT, KP. I am subscribed to EPL index too which has much much more data but obv. costs. For last season's data you must check out MCFC Analytics. An awesome initiative from the current Champions.

    It's hard to rate cleansheets due to their binary nature. I do not multiply probabilities together though, and my calcualtions is equivalent to a poisson distribution model. It actually works but you are right, I am undervaluing the resilience of the better defensive teams. I've found that some teams defend well by preventing shots and shots on target, others defend well by not allowing SoT to be converted into goals. I have some great data to demonstrate this but until I figure out the underlying drivers I am not going to incorporate this. I am not keen on saying just becasue a team is a big team then they are better defenders. The stats will tell me this one day I hope, and I do have a few promisng in-roads to this.

    I don't think you were thinking of a bsic poisson distribution article when you mentioned that but rather something more specific on what I do. However there are a few articles I've posted on the forum for reference

    Thanks for the comments and ideas gleamed. I invite you to get invovled on our fledging forum. You'll find a link a the top of the blog.

  8. Guilty! Yeah that is me, good spot & thanks, Glad you enjoyed the article. Seems it's a much smaller world than I first thought!

    Thanks for all those links, it's going to take me all week to read all the stuff on this site let alone anywhere else. (my missus will kill me!) I just applied for the MCFC Analytic to send me their data will be good to see what those guys are doing. Thanks for pointing that out!

    Re: Cleansheets, not sure if I have the wrong end of the stick, you mention...
    1 would be 4*(67%)= 2.6.
    2 is 4*(67%)^2 = 1.8pts,

    Then for 3 games are you implying this should be 4*(67%)^3.
    hence why I took it that you are multiplying the likelihoods.

    2/3*2/3*2/3. p(AnBnC) where A,B and C are separate games. That's why I suggest the Newton dice Prob example to solve.

    This doesn't seem equivalent to a Poisson distribution. Would be interested to see how you have included this, can see that from the above??

    Have you thought about goodness of fit for clean sheets, you'd need a lot of data but you seem to have that. Model this to see which distribution fits. The appropriate distribution will then produce better idea's on how to come up with likelihoods, than just the binary stats. (Rather than working the other way round, from likelihood and trying to fit the distribution.)
    Thinking it could fit to a negative distribution which would be interesting.

    Outside of this (but still on the cleansheets topic) I think I have an idea for drivers but again am lacking data i.e. final third possession stats (of the other team), final quarter forward passes. etc. Would it be possible to get info like this from somewhere?? guess this stuff won't be free. Plenty of other drivers I could think of that would be great to test.

    You are right, I'm wasn't talking about a basic Poisson distribution article, thanks for the link to that Poisson page, when I get some free time I'll have a look and get back to you on what I mean exactly.

    It's really cool to find someone that's done loads more work on this than me... Will be good to collaborate and pick your mind!

    Would like to say I'm not a massive blogger, but look at the size of this article... ha. Will have a look at the forum as well, going to be reading alot!!!

    Seems like a great site. Bravo.
    Kepp it up'Shots on Target.'

  9. Hi 'Shots on Target'

    Just read those article's on the Poisson models, other than that guys thesis, don't have a month to spare!

    They are pretty basic models, can you explain what you are doing on the poisson side with your model, I think that would make an interesting article.

    The pinnaclesports article adds the most value but doesn't ring true, I do not think this is right....

    Newcastle’s Goals = Newcastle’s Attack x Tottenham’s Defence x Average No. Goals

    it doesn't seem quite right. if you think in probabilty terms, what is this set?? (doesn't make sense) Will think it over and come back. The methodology though could perhaps be useful.

    I also added a few explanations to the article. As there seemed to be quite a few questions outstanding on that one.

    Cool, Cheers.

    1. Hi Matt, I will get back to you later today on these points. Apprecite your input.