No (to SK) - his formula assumes a linear relationship between the variables, which it is not.

Bale is better value than Milner - As shown by % ownership, but formula assumes price is correct

My post was a lil test of the reality of what you are discussing. Surely it merits a reply, or is the thread totally abstract?

Whilst I do love a good multiple regression, and without wanting to piss on anyone's parade as some value can be gained from this type of thing, but my memory from statistics (which was gained sat at the back of the lecture theatre whilst half asleep) is that a regression with an adjusted R Square of 0.17 is a not a very good model for predicting future out comes. Is my memory correct?

A players score in their next game is just far to random to be able to build a model to predict it if you ask me. However it could be more useful to predict the score over say 6 games, taking all opposition over that period into account, as then players will tend to perform to average over a longer period.

OK, based on my research I would say your squad is Pants.

Oh

I'd suggest that if anyone (myself included) had any real faith in our predictions that we likely wouldn't share them. We'd be looking instead for cash leagues.

Although I have no idea at how Wyld got to his formula, the one thing I did notice is that he uses form as the player's last score. This is a player's form only if they are playing under similar conditions (home/away, strength of opponent & position). Is Bale's prior week's points away to Chelsea as a left back his form for an upcoming match at home against Wigan & playing outside mid? (Anyone else notice how Etuhu's shots went from 2/game to 0 when moved from left to right mid?).

As it turns out, my predicted points for Bale this week are higher than Milner (significantly higher actually). And although based on a small sample, include a goal, 2 assits & 9 bonus points in three away games to teams "similar" to West Ham. Milner has only one goal and no assists over the past two years at home to Top-5 clubs. But that was with Villa of course, and Man City *should* have more scoring potential. But against Chelsea right now? Not likely (and so cue up an assist hat-trick).

When a player changes clubs, or changes his FPL position, obviously his potential is affected. So maybe last season's stats aren't worth much.

But we do have this season's stats (albeit for 6 weeks only)...

Bale (7.1m) 30 points from 540 minutes
Milner (9.1m) 30 points from 534 minutes

My formula is fairly easy to understand...

predicted score next match (FPL points) = 0.34*price + 41*historicalYield + 0.053*opposition + 1.01*venue + 0.04*form - 1.63

where:
price = the player's price we have
historicalYield = points per minute played
opposition = the rank of the opposing team...let's take my 1-20 ranking above: Bale v Villa (rank 6), Milner v Newcastle (rank 15)
venue = home advantage: Bale at home = 1, Milner at home = 1
form = points scored in last game: Bale 2, Milner 3
-1.63 = the regression intercept (don't worry about it, just subtract it)

Thus...
Bale, predicted score next match (FPL points) = 0.34*7.1 + 41*(30/540) + 0.053*6 + 1.01*1 + 0.04*2 - 1.63 = 4.5
Milner, predicted score next match (FPL points) = 0.34*9.1 + 41*(30/534) + 0.053*15 + 1.01*1 + 0.04*3 - 5.7

Thus my formula predicts that Milner will score more points this weekend. Note that, for this particular comparison, the factors historicalYield, venue, and form are nearly the same or else trivial.


My formula says that the main difference between these two players, in this instance, is their price. Simply by being a more expensive player, Milner is expected to score more points.

Along with this, Milner also has a slight advantage this week in facing weaker competition.


Ownership: no I didn't consider that. That is an interesting suggestion. Where is the data?
Ranking: the 1-20 ranking is, I think, good enough. Of course the order in which you rank the teams is important. I have toyed with the idea of using bookies tables for this.
Home advantage: please show the data which support your contention.




The test of my formula should simply be this: does it predict the score of a player (over a reasonable number of games) significantly better than chance? If it does, it has some worth. (Wasn't anyone impressed, by the way, that it spat out some reasonable looking values for the points each player is likely to score?)

In fact I will go further and say that I believe that my formula is the most accurate formula for predicting a player's FPL score next gameweek ever posted on FISO. I challenge everyone and anyone to come up with a more accurate predictive formula. I'll go even further than that and say that I believe my formula will do better even than a human being who just says what he thinks the player's score will be based on his knowledge of the game.

(Again, of course, based several players over a reasonable number of games.)

Least-squares multiple regression is quite a robust statistical method, and works well even when the data is a liitle skewed. I took the trouble to make sure that my variables were independent.


For this kind of model, an adjusted R Square of 0.17 is neither high nor low. It means that my formula accounts for 17% of the variation in player scores, or to put it another way it doesn't account for 83% of the variation. An important question to ask might be, what proportion do we actually expect luck to play? Some weeks a player scores 1 point and other weeks he scores 13.


That's a good point about the weak measure for form that I used. Do you have any data to back up the contention that a player's form over, say, the last two games or four correlate with his actual score?

By the way, I built the Packing Solver.

Please see attached jpg showing positions at end of season, for last 3 seasons and then a league showing % of home points gathered. Seems like the bottom team last season didn't take use of home advantage but overall, seems like hancockjr is correct.



It is definitely the only formula that is "spewing" out numbers. How accurate is yet to be seen.



So I think it will be interesting to set this challenge. Anyone can enter - I'll be happy to judicate. We should spread this across 8-10 GWs so we give each "formula" a chance to balance itself out. I'm happy to randomly choose 25 fpl players and they will stay constant during this period.

People can sign up now saying they are interested in joining the challenge - whether they have their own formula; go with gut feeling; speculate or read tea leaves. Every week, by Friday night at midnight, each player either posts or PMs me the points that they think each one of the 25 players will get that GW (we can decide whether it should be PM or open post if we think it can influence another player.

At the end of the 8-10 GWs we can see who has the most accurate way of predicting points.

Who is interested??

No, and I don't really use form, but it did occur to me that since you were, a more precise measure of form might be their last game at home when home or away when away.

Cool. What tool (ap, programming language, stat's package, etc.) did you use?

I'd be up for that, but if we're starting in GW7 I'd need to know asap so I can fire up Deep Blue to run the calculations

Seriously, I want in - I have a formula of sorts that eliminates most (but not quite all) judgement calls, based on some of what I have read in this thread and my own thoughts, but I'm not confident enough of it to base my real FPL decisions on it (or to publish it just yet), so this would be a good chance to test it out

I wrote it in an interpreted language called REBOL (which I had to learn). The results were very interesting and not at all what I expected.

What do you mean by "% of home points gathered"? Is your data for 1) FPL points or 2) league points (3 for a win, 1 for a draw, 0 for a loss)?

They would be very different things I think.

The home advantage factor in my equation is for FPL points. It assumes that players in the top teams and in the bottom teams gain approximately the same advantage playing at home. This advantage is around 1 FPL point per player, regardless of the team that they are playing in.

Hancockjr seems to be saying that he thinks this is not true, that bottom teams will have a better home advantage, and hence that Blackpool players will score proportionately more points at home -- although it not clear to me if he is talking about FPL points or league points.

I think the data needed to test this would be by comparing the ratio:

FPLSeasonPointsHome/FPLSeasonPointsAway

for the top and bottom teams (or for all the teams).

If the ratios are approximately the same, I'm right. If they are significantly different (statistical meaning) in favour of bottom teams, Hancockjr is right.

And if he is right, then the ranking of the player's own team would be a useful addition to the predictive formula.

The % of home points gathered is badly explained on my part - apologies. Yes it relates to point 2 above

However, does it not still have relevance to your point 1 above too? eg. Burnley got 30 pts last season & finished 18th getting relegated. 26 of those points were collected at home, meaning their home points equated to 86.67% of their overall point tally. Chelsea in 07/08 came 2nd in the league - 50% of their points came from home games and 50% from away games.

I could look further into goals scored, cleansheets etc but when considering fpl points - a player from Burnley is most likely to possibly get 86.67% of their points at home (more chances to score, get assists, get cleansheets etc), whilst a Chelsea player's fpl points are going to be more spread home and away.

Or am I missing something here?

I don't think you are missing anything at all. I've edited my post above to make it clearer what data needs to be compared to prove Hancockjr's hypothesis. (And might I add, I quite like the look of his hypothesis.)

Wyld's new formula for predicting a player's score in his next match.

Looks like Hancockjr was right. I ran the regression again, adding in the player's own team ranking. That gave the following adjustment to my formula:

predicted score next match (FPL points) = 0.27*price + 35*historicalYield + 0.04*opposition + 0.98*venue - 0.09*ownteam - 0.1

where:
price = the player's current price
historicalYield = points per minute played
opposition = the rank of the opposing team (Chelsea 1, ManUtd 2...)
venue = home advantage: home = 1, away = 0
ownteam = the rank of the player's own team (Chelsea 1, ManUtd 2...)
-0.1 = the regression intercept (don't worry about it, just subtract it)

form (as measured by points last game) was not significant and therefore dropped from the equation

Total variance accounted for (R square) = 18.8% (up 1.5%)

Updated predictions for Bale/Milner, next game:
Bale, predicted score next match (FPL points) = 0.27*7.1 + 35*(30/540) + 0.04*6 + 0.98*1 - 0.09*4 - 0.1 = 4.6
Milner, predicted score next match (FPL points) = 0.27*9.1 + 35*(30/534) + 0.04*15 + 0.98*1 - 0.09*6 - 0.1 = 5.4

(Remember, these predictions are based on only 6 gameweeks of data, so will not be so accurate yet.)

as good as your model looks, having no weight put on form seems crazy to me. how many points should Rooney have scored this season, for example?

Surley thats what historicalYield = points per minute played tries to factor in?

only if the 'historical' aspect of this is relevant only to this season rather than previous seasons

edit: actually reading back in this thread, historical yield is based on last season's stats. so by Wyld's calculations Rooney should be scoring big this season which he obviously isn't, as he isn't in form.

I completley get what you mean now! As long as historical Yield takes into account this year only then it is benificial

^yep. I reckon form needs a say in wyld's model alongside historical scoring. considering the scoring streaks and dry spells that strikers go through i'd say not only should it have some say in a prediction, but it should be the dominant factor. Especially for strikers.

I tend to agree my friend does anyone have an accurate Definition of what Historical yield is?

I dropped "form" from my equation because it wasn't statistically significant -- even though I wanted it to be.

The measure of "form" I took for my analysis was a little lazy -- it was just what the player scored the week before.

historicalYield has been defined several times on this thread already. You can derive it (in points per minute played) from last season's data (for the players who are still playing) or from the first 6 gameweeks of this season, or from several past seasons' data, or from all of these datasets with different weightings for each.

Let's say we were in gameweek 38. historicalYield could be calculated from the player's score on gameweeks 1-33. form could be calculated from gameweeks 34-37.

A simple test of your hypothesis that form is important may be derived thus:

1) Calculate each player's average score in gameweeks 1 and 2.
2) Calculate each player's average score in gameweeks 3 and 4.
3) Calculate each player's average score in gameweeks 5 and 6.
4) Correlate 1 and 2 with 3.
Is 2 a better predictor of 3 than 1 is? If yes, your argument is proved. If not, back to the drawing board.

Briilant thanks for the definition at the end of the day this is your formulae if it works for you then fair enough expect to be seeing some big point hauls though!

verry interesting



I bet there's a way of optimising how far you look back for form, too, depending on significance...

Is form an illusion?

Very brilliantly calculated. All the smart minds at work and I am sure it has intense logic behind it .
One thing i cannot understand though.
Why is PRICE a factor in determining a players score for the coming GW ?
I mean i just cannot, for the love of god, fathom that.
Two players - having had exactly a similar season - one priced a 9 and the other a 4.5 - both have scored 40 points in the last 6 GWs for example
Now your formula would predict higher points for the 9m player ?
Why ?
Whats the logic behind this ?
I am not saying it is wrong - i just cannot understand.
A Drogba at 5m will do worse than a drogba at 13.5 M ?
Why ?

I'm a bit curious to the answer myself, but would guess that it has something to do with the notion that FPL employs statistical experts to price a player according to their expected yield (do they?, I'm guessing here), and so in a way, a player's price takes into account the expertise employed by FPL.

Oh I think price should be in there but how to weight it...almost impossible as we don't know how many factors FPL take into account or don't take into account?? A gut prediction on a players ppg will not take price into account...or it shouldn't!

Expensive players score more points, in general, on average. It's as simple as that.

well spotted. abcd1234. price of a player doesn't come into it.

I think really before an accurate model can be created it needs to be determined which factors actually contribute to score and which are just byproducts. by all means a model can be made but at the end of the day it needs to be made using the correct variables. so price here is obviously a byproduct of other factors, in the same way that for example binge drinking has been connected with lung cancer... something that comes about simply because binge drinkers are more likely to smoke.

probably in the case of fpl point scoring the variables should be something like:
historic ppg/ppm record
form ppg/ppm record
historic quality of opposition attack/defence (also weighted for home/away)
current quality of service from teammates (approximated to league position and weighted home/away)
role in team (more/less attacking than fpl classification)

possibly also amount of bandaging apparent on player's head...

anyway if there was a way of using wyld's programming to establish the significance of each of these factors I think we could be onto a good thing but for me price shouldn't be a consideration.