If you’ve heard the term moneyball, then you’ll know that in 2002 the Oakland ‘A’s Major League Baseball team began to use statistical analysis to identify and sign undervalued players, in order to compete against their richer competitors. The approach is credited with getting them to the playoffs in both 2002 and 2003 and has since been adopted more widely.
In the football world, Brentford FC are reportedly embarking on a similar journey using the data-driven approach proven at their Danish sister club.
So could an AI algorithm win fantasy football?
The short answer is probably yes (and maybe already has!), but here is how a less successful attempt went this season!
Fantasy Football 101
The English Premier League’s free fantasy football competition lets you pick a squad of 15 players from the various Premier League teams, with the aim of seeing how your squad performs compared to other peoples. There are restrictions on the make up of your squad, including a maximum cost that prevents you loading up with too many top players.
Each week you select a team of 11 players from your squad to “play”, and then score points depending on how well those players perform in actual matches that week. If a player is not performing then you can transfer them out and another player in. There are limits on the number of free transfers that you can make and player costs change depending on demand (making popular players less affordable). There are two opportunities to make unlimited transfers during the season by playing a transfer wildcard.
My goal was to have a fully automated team entered into the competition. The algorithm would select a squad at the start of the season, collect data each week, make any transfers and select the best team to field based on that data. I didn’t achieve full automation this season, and so manually applied aspects of the algorithm instead.
Firstly, this algorithm isn’t really AI. It’s more of a first attempt to work with some of the data involved!
Initial squad selection: Step 1 was to select the best combination of players that met the cost and team structure rules. The query I developed attempted to ”brute force” valid combinations of players and then order those by overall score last season. There were some obvious downsides to this approach!
- The query ignored players that were not involved last season (new signings and those from promoted clubs)
- The highest scoring combination of players wouldn’t necessarily have won last season (transfers and captain choices play a large part in overall score).
- The number of combinations considered had to be limited in order for the query to run in a reasonable time
- The best scoring combination I could identify was some way off both the actual top scores from last season and the theoretical maximum for a squad with no cost restrictions (the best 2 keepers, plus the best 5 defenders, etc.)
On a weekly basis: Gather data, make transfers and pick team.
- Get latest player performance data
- Get performance data for my squad
- Get betting odds for forthcoming games (which teams are most likely to win)
- Transfer out the player with the worst form and/or least likely to play
- Transfer in a player with better form from a team with good odds (and within cost rules, etc.)
- Select a team based on home advantage and odds for the forthcoming games (e.g. favor Chelsea at home)
An initial issue with the transfer approach was that budget bench players would be transferred out until the point where the algorithm could no longer make a useful transfer and expensive players were sat on the bench tying up funds. Consequently I adapted the rule after the first half of the season to generally leave budget bench players alone and only transfer the regular players.
Periodic squad reselection: At two points during the season the squad was reset using the transfer wildcards. I tried different approaches to the selection on each occasion. The variations tried involved:
- Pre-selecting best value budget bench players (playing regularly and picking up some points)
- Consider form and value as well as overall points when selecting main players
The squad was entered from week two.
It finished third in a small private league of eight (second if you discard the points from the first week).
It had an overall rank of 455,222 out of 3,502,998 (I.e. in the top 10-15%)
And was ranked 5,412 out of those entering in game week two.
What would I do differently next time?
The fantasy premier league form calculation wasn’t wholly reliable. Their form calculation is an average over a thirty day period, consequently a player who scored very well 3 or 4 games ago but has since been poor, can still have a relatively high form. Instead it might be better to recalculate form, factoring in how consistently the player is scoring points and/or whether their performance is trending up or down.
Too many selection factors were retrospective – just because a player has done well doesn’t mean that they will continue to do so. Instead it would be interesting to gather more predictive data (e.g. fan tips) and try to identify the factors that are most linked to future performance.
It would be nice to more fully automate the processing and team management. The tool-set used should allow this to be achieved.
Data was gathered using Python and HTQL and stored in a Microsoft Azure SQL Server database.
Analysis was performed using SQL queries.