Thread by @FPLHAL

1/ Okay, this is going to be controversial, but in #FPL:

*There is no such thing as form*

Hopefully I can convince you with data, this is going to be a (very) long thread

2/ Throughout the #FPLCommunity the key discussion from week-to-week is which players to transfer in and out. To do that we usually look at both player fixture and form.

Whilst fixture difficulty is not too hard to estimate, trying to quantify form is a lot harder.

3/ The problem comes from Apophenia. The tendency to perceive connections and meaning between unrelated things. This is also commonly thought of as the "Gambler's fallacy" (en.wikipedia.org/wiki/Gambler%2…)

4/ The other issue, especially when considering form, is that people are very poor at thinking about 'random' patterns. Ask a group of people to spread themselves out randomly in a room and they inevitably place themselves evenly.

5/ In practice, randomness comes in 'clumps'. These clumps are then mistakenly thought to be non-random (en.wikipedia.org/wiki/Clusterin…)

6/ Take for example a player which scores 15 goals in a season. A random shuffle of these goals throughout the GWs may look like this: 1,1,1,1,0,1,0,0,1,0,0,1,0,1,0,0,0,0,0,0,0,1,1,0,1,1,0,0,1,0,0,1,0,0,0,1,0,0 (this random sequence was generated in Python). Where '1's are goals

7/ If this was a striker in #FPL would everyone be jumping on the train/party in those first 6 weeks citing fantastic form thanks to a good rest over the summer and bemoaning the run of 7 games without a goal in the middle of the season? Perhaps during the busy festive periods?

8/ Perfectly reasonable, but we're just finding patterns in randomness.

9/ Now, lets try to quantify the above with data. First, lets make our dataset. Consider every player over the last 2 seasons and for their points we will follow the 'adjusted' points [AP] model of @mathsafe_fpl and @WGTA_FPL (points with the appearance points removed)

10/ Further to this I have also removed any bonus points scored and tried to account for fixture difficulty. To do this I have calculated the average AP scored and conceded by a team respectively. A (reasonable?) way to account for the difficulty is then removing these 2 numbers.

11/ For example, consider Man City v Watford. City players score on average 1.12 AP points per game and Watford concede 0.61. So in that match City players have 1.73 points removed. However Watford score 0.26 and Man City concede -0.02 so Watford players have 0.24 points removed.

12/ Then for each player, for each season, the average (mean) of their scores are removed.

That leaves us with this kind of thing (this is the time series of average removed adjusted points for Sterling for the 2018/19 season, the gaps are games where he played no part).

13/ Looking across the whole dataset we have 20,875 data points. Of those 7,180 are positive (above the mean) and 13,767 are negative (below the mean). The distribution of the points is shown below.

14/ So, overall, there is a 34% chance that a player will exceed their average points in a given GW.

Note this doesn't vary hugely depending on the player. For 2 examples: in 2018/19 Sterling exceeded his average 38% of the time, in 2017/18 Maguire exceeded his 37% of the time.

15/ Using this value as a basis, we can work out how this compares with the number of "above mean" weeks in weeks following one above the mean. E.g. the impact of form.

If form exists we would expect there to be groupings of "above mean" weeks rather than them being random.

16/ To draw this conclusion we need to know what the statistically expected values would be. To calculate those we assume that the scores are distributed by a binomial distribution (en.wikipedia.org/wiki/Binomial_…).

17/ This is a distribution for calculating the number of successes (in this case above mean weeks) in a sequence of trials.

Using a base rate of 34% (if we randomly pick a player and a GW there is a 34% chance that it is above the mean) we can compare stats with the actual data

18/ First we look at "3 week form" that is, if a player scores above their mean in a given GW, what is the chance that they score above the mean *at least once* in the next 2 weeks?

19/ Well, from the statistical distribution, we would expect there to be a 57% chance of this happening, and if we look at the cold hard numbers from the last 2 seasons then it happened 52% of the time. Pretty close!

20/ If we do the same for "4 wk form" (following an above mean score the player posts another above mean score at least once in the next 3 wks). Stats says 71% chance, the #FPL data? 61%.

"5 wk form" stats says: 81%, our data: 72% and finally "6 wk form", stats: 88%, data: 78%.

21/ So what conclusions can we draw from that?

If a player scores above average (for that player) in a given GW, it has no impact on their expected returns in any of the following 5 weeks.

22/ Similarly if you look at something like the chance of "2 above mean weeks in 4", or "3 above in 5" or any such combo, the same thing happens. The empirically calculated results are (very much) in line with the statistically expected values.

23/ Dare I conclude:

There is no such thing as form.

24/ Clearly there are other short term factors which can impact a player, which we as keen #FPL players should focus on.

For example a short term injury to a key player could have a large impact on another (e.g. an injury to Kane might be a good time to bring in Moura/Son).

25/ TL;DR:

There is no such thing as form. When picking players, worry about fixtures and other factors which may impact *future* performance, not how they have performed "in the last couple of weeks".

/END

@threadreaderapp unroll please