Predicting the Top 10,000 In Blank Gameweek 31

This article follows on from my earlier look at how the Top 10,000 Fantasy Premier League managers are geared up for Blank Gameweek 31.

This time around I will go one step further and attempt to predict what final moves they will make to set themselves up for this weekend’s limited fixture list, when only four matches take place.

For this latest analysis I will detail the results of a complex predictive statistical model that I created over the weekend. This takes into account various factors about the history of each manager in the Top 10k and predicts the chance of them using their Free Hit chip (FH) for Gameweek 31 and how many hits they will take. The goal is to provide a simulation of what’s going to be happening next week.

Gameweek 30’s Moves

Some of the assumptions I made last week were very simple in order to just get an initial feel for Gameweek 31, but hopefully the complex model I built will hugely improve on that. Analysing the transfers made for Gameweek 30, we could test some of the assumptions:

Assumption: Every team who already used their FH will use all free transfers to bring in Gameweek 31 players.

Reality: In Gameweek 30, only 0.6% of players who already used FH used free transfers to get non-Gameweek 31 players. There focus was Gameweek 31 players, so that was proved correct.

Assumption: Every team who won’t use FH in Gameweek 31 will use all free transfers to bring in Gameweek 31 players.

Reality: In Gameweek 30, 50.09% of players who didn’t use FH used FTs to get non- Gameweek 31 players. So around 50% of managers with their FH intact ignored Gameweek 31 completely in their Gameweek 30 transfers. If we say the 50% ignored it because they plan to use FH anyway in 31, then this gives an estimate of 3018 managers who will use FH in Gameweek 31. This is still close to the predicted number from last week which was 2780. But the article should get a much more infromed estimate, so keep reading.

The assumptions made in that previous article about hits were not very concrete and that’s what made me build the new statistical model (that’s my speciality in real life).

I also included weights for different players in Gamweek 31 so that players that are likely to be benched, such as Everton’s Oumar Niasse and Huddersfield’s Colin Quaner, are not equated with nailed on players.

Explaining the predictive model

The algorithms I developed are very hard/not useful to explain in an article but I will explain the general idea behind them.

For every Top 10,000 manager we extract their transfer history to know how lenient he generally is with taking hits. Using that and the general understanding that in Blank Gameweeks it is usually easier for people to take hits, we build a distribution of probabilities for each manager to take 0, 1, 2,.. hits this week.

This distribution also takes into account the situation of the player with regards to how many players he already owns, how many free transfers he has, so managers with a small number of potential starters are more likely to take hits.

Example:

As an example, let’s take the #1 player in FPL. He currently owns 6.5 players (with Niasse as the 0.5) but he has two free transfers. So, he could field 8.5 players with no hits. The algorithm puts that into account + the fact that over the course of the season, he took two -4 hits and one -8, and predicts that his likely moves this week are: 39.2% no hits, 40.5% 1 hit, 20.3% 2 hits.

Which gives him an average of 9.31 players with average deducted points of -3.24. Note that although he usually doesn’t take many hits the algorithm predicts he’s probably going to take one hit because he might prefer fielding 9.5 players with a hit rather than only 8.5.

On the other hand, for the current #2, he only owns three players now with one free transfer. If he were to keep hold of his FH,  the algorithm predicts that he will be able to field only 5.38 players with -5.52 points deducted. So, he will probably use his FH, but determining the probability of that is the task of the second part of the algorithm.

The Model for FH Probability:

The first part of the algorithm runs for all the Top 10,000 to get their expected hits and number of starters. Based on that, the second part of the algorithm tries to predict the chance each manager will use their FH, if available.

A predicted small number of starters and a predicted big number of hits will result in higher chance of using FH. Other factors included in the algorithm include how much the manager has been planning for Gameweek 31 by looking at the percentage of recent transfers that were used to get those players in. Another factor included (with a smaller weight) is whether they have two free transfers. If this is the case they may be tempted to hold their FH as deploying it would mean they lose one of their free transfers.

Combining those factors in a proper statistical way yields the probability of whether each manager will use their FH. For example, if a player who usually hates taking hits is in a position where they have to take multiple hits to field a reasonable number of players, they will be more likely to FH than a player in an identical situation but who is historically fine with hits.

Examples 

For the #1 player, the algorithm predicts a chance of 24.5% that he will use FH. While for #2, the probability is 78.8%. This reflects their situations with number of players and FTs they have, as well as their transfer history plus the unpredictability of human beings.

Here are some examples for the predictions of the model:

Again, note that the algorithm just tries to predict the behavior, and is not in any way saying what a player should or shouldn’t do. It’s like a chess computer stalking us and predicting our next move.

Also, the model doesn’t just go with the most probable scenario for each manager, but keeps the whole spectrum of choices a FPL manager has so that averages can be calculated correctly.

Results

Now, we can run the algorithm on all Top 10,000 managers to get the probability of FH use for each in Gameweek 31. From that we get the following simulation results:

Expected number of FHs in the Top 10,000: 3268.9 

This is significantly higher than previously thought (2780), but still comparable to the crude estimation at the beginning of the article (3018).

Expected number of Gameweek 31 players in teams with no FH: 7.37 players

Also lower than previously calculated.

Expected number of point deductions in teams not using their FH: -5.24

Now, we will use this simulation to calculate the scores teams will get depending on how an average Gamweek 31 player will score:

Note that this gain is comparing the average non FH team to the FH team, but your team could be better than the average so a FH team is not beating you by the same points as in the table.

The average team will field 7.37 players with around one point hit so you might be better.

Note also that even if you field seven to eight players (so average), you might choose the wrong/right players and be well below/above the non FH average.

If you want to calculate your predicted points to compare, use this formula and compare that to the averages in the table above to judge your potential gainloss.:

(points_per_player)*(num_of_fielded_players) – 4*hits_taken

Final Thoughts

I will not attempt to say whether the use of the FH chip in Gameweek 31 is the right decision, because as you can see, it is very subjective and we don’t know much yet about Gameweek35. But from this predicted simulation, it seems that players who will use the FH will get a decent boost over active non FH teams (maybe 16-22 points), and a very significant boost over dead teams, which I didn’t even consider (since I have assumed most in the Top 10,000 are active).

I will also be glad to tell you what the algorithm predicts about your next move, but again this is not an advice but just an observation of your patterns. If interested, let me know your FPL id in the comment section below. Good luck.