MAny of those will be impossible to recerate on the build dataset - you may know them now but not hostorically to establish the relationship.

Price is pretty correlated with historic ppg.

I agree price is poor - I suggested use % ownership, maybe in combination, as this sort of shows whether the price is still considered accurate, but no-one shoudl be surprised it shows as significant in the model, as it is simple, available and intuitively sensible.

This is pedantic I know (sorry ), but it gets to causation. A player doesn't score more points because he's expensive, he's expensive because he scores more points (or is expected to).

I think you are on to something here. But maybe not % ownership, but change in % ownership (which you might have meant anyway). Price says Rooney will score, but if his ownership percentage has dropped (and I'm guessing it has), the change in ownership might be more significant than the actual ownership (this might also have a direct correlation to change in price, rather than actual price).

Along these same lines, I wonder if volatility tells us anything. Very active buying & selling, while not significantly changing either % ownership or cost, may mean something (but I don't have any ideas on what it might mean, or whether it actually happens or not - I only thought of it because of it's effect on stocks).

It doesn't need to be a causal relatinship, just a relationship.

Change in % may work but if Drog is scoring week after week after week, eventually people will stop buying but that should not lower his prediction.

Think the only way we can be sure is if we set up as many formulas with different variables as possible and see which on average is the best!


price shouldn't affect a persons game week score should it?

Imagine you know nothng about rugby, but you have to pick a fantasy team. Which players do you think will score the most in the first (or any) week? The expensive ones or the cheap ones? Who generlaly score the most points - Gerrard, Drogba, Fab and Torres or <insert own list of journeyman pros at £5m each>?

I think you are both right. My analysis says that both a) price and b) historical yield are useful measures. It says nothing about causality, only that there is a relationship between them.

Price and historical yield are actually more useful predictors when treated separately (i.e. account for more variance) than they are if combined (e.g. as a measure of "value" like historicalYield/price). I interpret that to mean that there is valuable predictive information in both measures that is not contained in the other.

That's what I'm hoping anyway

The problem is, the higher price does not mean the player will score more points, not in the short term anyway. i.e. Kuyt vs Arteta, I don't think there's any question Arteta is the better player (by far IMO) and should/will score more ppg despite the price. Kuyt is living off one good season and maybe playing for Liverpool ( ). I don't see the correlation in enough (a high % is enough).

You can't just quote isolated examples. I have looked at two pages from the stats sheet - highest priced mids and 6m to 5.4m. Price and season points so far are 41% correlated. It really is that simple:

Lampard CH 12.8 15 42%
Fabregas AR 11.9 22
Gerrard LI 11.5 29
Malouda CH 10.3 51
Arshavin AR 9.9 33
Milner MC 9.1 30
Kuyt LI 8.7 16
Cole LI 8.7 2
Silva MC 8.5 12
Cahill EV 8.5 23
Young A AV 8.4 31
Nani MU 8.4 45
Arteta EV 8.3 24
Lennon TH 7.9 18
Valencia MU 7.9 4
Van der Vaart TH 7.5 13
Diaby AR 7.4 19
Kalou CH 7.4 34
Modric TH 7.3 11
Aquilani LI 7.3 0
Wright-Phillips MC 7.2 5
Nasri AR 7.2 20
Essien CH 7.1 30
Bale TH 7.1 30
Benayoun CH 7.1 8
Ramsey AR 6 0
Bilyaletdinov EV 6 5
Diamanti WH 5.9 1
Huddlestone TH 5.9 13
Gamst Pedersen BR 5.9 16
Murphy FU 5.9 15
Osman EV 5.8 10
Etherington SC 5.8 26
Bentley TH 5.8 0
Nolan NU 5.7 29
Petrov BW 5.7 23
Taylor BW 5.6 5
Larsson BI 5.6 24
Song AR 5.5 17
Petrov AV 5.5 17
Anderson MU 5.5 1
Cattermole SU 5.5 1
N'Zogbia WA 5.5 9
Henderson SU 5.5 22
Hargreaves MU 5.5 0
Collison WH 5.5 0
Walters SC 5.5 10
Noble WH 5.4 18
Parker WH 5.4 26
Kompany MC 5.4 1

However I don't understand how basic situations like Kuyt vs Arteta arise, completely wrong prices, seems so obvious. If it is so correlated that solves most of our problem then, just pick your most expensive xi every week, NO exceptions (problem is there are lots)?? I mean who actually knows how much FPL takes into account with the prices.
But Wyld has stated already "Simply by being a more expensive player, Milner is expected to score more points", though 'everyone' was astoundingly (wrongly) on Bale's side last weekend. So gut went against formula bigtime.

It's only 42% though - so plenty of time to identify the other 58% - and of course some players may be completely mis-priced. Form etc also matter, that is the point of regression - find lots of explanatory variables and get a good combination of them all.

Historical statistics says that Arteta is the slightly better player (points/minute over 2 seasons). But he plays for the slightly poorer team. Therefore we should expect both players to have similar prices -- and they do. The fact that Arteta has a slightly lower price is just natural variation in the market that a sensible player takes advantage of.

I have a few ideas of some variables that may or may not improve the equation, but that I think may be worth checking (if you can be bothered of course ). They may be highly correlated with some that you already have, but I think they may have some added value. I don't have the stats to back my thoughts up, but I think they make sense, although possibly adding unnecessary complexity

Form I said earlier in the thread that I think this regression analysis could be more useful for predicting the next say 5 or so games as a players next score is a bit too random. Well I think that should also be true for how a players score in their last game, so I think a players ppg over the last 4 to 6 games (to be tested for which is best) should be more significant in predicting future scores than just the score in the last game.

Player Position Part of the usefulness of this may be in the Price variable, but I think strikers `tend` to have a higher ppg than midfielders, who themselves `tend` to have higher ppg than defenders. So I think it may be worth testing if a dummy variable for position is significant.

If not significant, then I think it may be when combined with some other variables.

Defensive Strength (Home and away) Some teams, such as Stoke, have a much better home defensive record than away, now this is covered in you Venue variable, but I think that the average goals conceded per game of the team that a defender plays for, depending on whether at home or away could add more value. So the variable would look something like (Def x Home x GC per game h)

Where Def = 1 or 0, Home = 1 or 0 and goals conceded per game at home of the team)

You would then also have (Def x Away x GC per game a)

You could then also have the strength of the opposition attack if the player is a defender by using their goals per game scored home and away, by having the variable along the lines of (Def x Op GS h or a).

The same could be said of attacking teams, so a variable on the attacking strength of the team home and away for strikers could also add value

Attacking strength (Home and Away) So along the same lines as above, the variable would look something like (Striker x Home x GS per game h). And then again the defensive strength of the opposition could also add value.

To make it simpler you could not bother with the home or away, but just have the teams goals scored/conceded per games ratios.

My gut feeling is that the next one may only be relevant for strikers, but I will call it Anti Form. If a player's recent form is above or below their long term average, then it might be likely that their form will return towards the norm. so the variable would be (CurrentYield - HistoricalYield) where CurrentYield = pointer per minute of the last 5 or so games, and historical yield is as in the equation above. So if I am right, this would be a negative relationship.

Anyway, those are my thoughts, hope they are clear. Discuss

But the ppm of Arteta/Kuyt are taking the 'poorer' team factor into account over the past 2 seasons so I don't think the prices should be about the same, Arteta should be about 9 and Kuyt about 8. You are right that there are always differentials that can be taken advantage of.

I'm bringing this topic back up to the front page, could do with a few more interesting discussions on this forum.

Interesting maths stuff going on here.

Wow. Tomorrow will be exactly one year since I bumped this topic and it's had only one reply to it since then. What's the matter with this place? It's nothing like it used to be back in the good ol' days.

perhaps you should add something to the debate if you're bumping it? you might (almost certainly will!) get responses that you'll find more interesting than mine that way

Interesting thread. Does anyone have a good dataset for last/this season they'd be able to share? If so please stick it up on Google Docs and post a link. It'd be fun to build a simple model and see next season how well a purely mathematical approach can do over a full season