Monday 15 July 2013

FSCORE (more analysis)

Following suggestions from a couple of wise readers this post is going to deconstruct F.SCORE some more, to see how it's accuracy varies when modelling FPL points based only on a player's statistical performances, by

  • GW range through the season - do stats paint a better picture of FPL points midseason? (thanks to Lucien for the suggestion)
  • FPL Points scored - how do stats explain points scored between high vs. low scoring players? (thanks to 2ndMan)
  • FPL Points vs FSCORE vs Team
There's also the small matter of the OMGFPLPRICELIST is released!  If you hadn't figured by now, I am blogging at InsideFPL which is a place for much more direct fantasy football content, the answers so to speak. This site will remain where the questions get asked ;) So please do check out InsideFPL for more from me. 

I'll have more of what the future holds soon enough but for now back to the hardcore stats!

F.SCORE vs Actual Points Through the Season
I've plotted Points Scored / 6GW and FSCORE / 6GW through the season to see how the model varies at different stages.


6 week rolling avg. 

At the start of the season FSCORE and Points hold pretty well but then actual points scored goes up quite suddenly around GW13-15 until GW27ish when it drops right down below.  In other words teams/players score goals at an expected rate per shot (etc.) at the season starts, mid-season they score more goals per shot, and then start missing loads at the end of the season.

Needs a fair bit more investigation this one!


F.SCORE vs Actual Points by Overall Points Scored
Points 0-75  R=0.90

Points 75-210   R=0.83

It's quite easy to see the correlation gets worse the higher the points scored, but thankfully not horrifically so.  There's definitely an unacceptable spread as pointed out in my last post at the 150pts mark.  I suggested last time that this could be due to Game State and Shot Zones but assists from set pieces or penalties will also factor heavily I think. 

F.SCORE vs FPL Points by Team
I observed last time that this spread featured players like Mata, Hazard and Tevez over-achieving their stats (above the line) and the likes of Jelavic and Cisse under the line, and hastily assumed that this was because the above-the-line lot were from freer-scoring teams.

In retrospect, this assumption needs testing and in truth I just went head-over-heels to take a closer look at Game State following 11Tegen11's own analysis. Back-tracking some then, let's take a look at how total F.SCORE for team's compare with their total FPL points..

For FWD & MIDS - FPL TAEM Points vs. FSCORE

This is pleasing and backs up the assumption that "truly better" clubs tend to turn their     chances (a.k.a. stats/FSCORE) into goals and FPL points.  Sunderland's little leap from the line needs to be checked out!

CONCLUSIONS
Firstly, quick-posts-before-bed seem to turn into epics! In general the correlations are still very strong between underlying player or now team stats and FPL points scored. There are significant variances from club to club (or elite club's players vs. non-elite)  and at different stages of the season and the aim of further analysis is to determine the data-driven factors that cause these variances and build that into FSCORE and Points Projections.
  • The FSCORE model works better for low-scoring players than high scoring ones.
  • At season start teams score goals and convert shots to point at closest to expected rate
  • Mid-Season they score more
  • End Season they score less
  • Players in teams who 'score for fun' (Utd) generally convert chances at a better rate than those who can be consider high scorers but easily "frustrated" (e.g. Liverpool)






6 comments:

  1. Hello SoT, nice to see you posting here again :)

    I'd like to add to your conclusions and see what you think of them:

    1. Your FSCORE model's accuracy is probably good enough. After all, FPL players only really want to know whether a player will "Do good" a particular week, Its virtually impossible imo to pinpoint exactly how well a player would do.

    2. For your second conclusion, I'd like to ask a confirmation question then: Does that mean that we should pick up your model's recommendations OFFENSIVELY early season?

    3. With your last 2 conclusions, I am reading that regardless of how freely a team scores, players will convert more shots into goals/points?!?!? Typo probably, but it would be nice to get a clarification from you anyway.

    Thanks for the good work man, and do drop your opinion on what you think of the new Bonus Points System!!

    Catch you around.

    Lucien

    ReplyDelete
  2. Thanks for the follow up post, really interesting stuff! Pleased to see the correlation for high scoring players is still strong (although the report on the graph and caption are different?).

    The F.Score vs Actual Points over the season is interesting too. You could certainly come up with an explanation of players settling into systems and hitting full fitness mid season, so they perform at their highest, but by late season become more tired, injuries are more common (thus disrupting the team and rotations) and top teams have to rotate around later stages of knock-out competitions. Or, it could all just be random variation about a mean. Wonder is the pattern this season is the same.

    Thanks again for the good article

    ReplyDelete
  3. Cheers for the feedback, Gents.

    @Luc - 1) you're right, and I'm moving on soon from this level of precision.
    2) Unknown. the conversion rate seems to vary but we don't know if this favours parricular teams or players.
    3) typo! (corrected)

    I'll have a post on the new BPS soon at insideFPL

    @2ndMan

    graph has R-squared, caption is R.

    I like your explanation. I think Lucien's initial interest in variance thru' the season was founded on a curiosity whether the relatively "crazy" start and end to the season are any different to the midseason. Seems they may be, but as mentioned to Luc above, this variance has to be considered the same for all teams atm.

    ReplyDelete
  4. so i've built my team like this using both rvp and bale , for the first 3 weeks :

    http://madjacquedisease.blogspot.in/2013/07/playing-rvp-bale.html

    let me know what you think

    ReplyDelete
  5. Hi Ste,

    Hope you will be keeping the good work on this site up too? i don't seem able to leave comments on the inside fpl, but signed up anyhow.

    now back to your fscore, interesting times...

    i) As your fscore is distinct to the goals and assists, it seems it may be better at predicting the new BPS system, have you put pen to paper in this space, think this will become even more correlated now... ?

    ii) For your past captain of the week predictions, if you were to use the fscore to select your captain each week (highest fscore, 1 player) how many points would you have ended up with from your skipper for the season? given RVP on his own scored 524 this is a good comparison.

    iii) I pulled together some research here on captains. (Not trying to plug.)
    http://www.fpl-dugout.co.uk/features-and-opinion/fpl-201213-in-stats-and-lessons-learnt
    Would be interested to know if the 'leader' beat your fscore, although I understand you have the added advantage of selecting any player without worrying about transfers. Would you say his average of 18 points a week was obsene? I'm stuggling to see where I went wrong, or any patterns to it & I've spent many hours looking at this. Would appreciate your input!

    Keep up the good work.
    Cheer Mate.
    Matt

    ReplyDelete
    Replies
    1. Hey Matt, thanks for your comment. I linked your article in the InsideFPL roundup. I do like FPL Dugout. In response to your questions, I will be posting here too. I took a quick look at BPS and will post that analysis at InsideFPL. I will definitely look at comparing my captain rankings vs RVP's 524. This is an excellent test. It will be interesting, and will be my next post here, this week hopefully. Good idea.

      Delete