MACHINE LEARNING + Fantasy Premier league (@fpl__ai)

Can machine learning models be used to predict FPL points of players?

I decided to build a machine learning project to check if this is possible, and to check the quality of the predictions. A twitter account created to monitor the results of the model can be found here .

Source code can be found here . This repo will be updated weekly

Data used?

Finding FPL data was pretty hard for me at first , I checked data from Opta stats and most popular football stats platform but I couldn't get something reasonable, after searching deeply I came across a GitHub repository for FPL data and thankfully the data had FPL data from 2016/17. I also discovered that FPL has an API that gives access to fantasy football data for the current season.

Data from the github repository ( 2022/23, 2021/2022 and 2020/21 seasons) were used to train the model, the FPL API will be used to get data for every gameweek (Gameweek 1 and 2 data already collected) in the current season.

Getting and cleaning Train Data

28/10/23: Some of the python files re-written into colab notebooks, to make running faster, since colab has more memory than my computer.

• merged_previous_seasons.py
  The data for each player in each gameweek from the 2022/23, 2021/2022 and 2020/21 seasons were downloaded and merged together, new features were also added like the total stats (points, bonus points, goals,saves and co) the player amassed for the previous seasons, also a new feature that represents the position on the final EPL table the player's team and every opponent the player played against had in the season before. Another interesting feature added is a column that shows the percentage a player's value contributed to the teams total value for every gameweek ,and a column that shows how many players are more valuable than a player in his position for his team.

Cleaning previous season data

• clean_previous_seasons.ipynb
  Some unused columns were dropped ["xP","opponent_team","expected_assists","expected_goal_involvements","expected_goals","expected_goals_conceded","element_type_ex","team_h_score","team_a_score","element" , "round" ,"fixture","starts" ] mainly because they had many missing values or don't have much relevance. Some set of historical features of type list for some columns (history_stats in the config.py file) were created, this list stored the last 3 values of a stat the player had in the last 3 gameweeks, for example, last 3 assists, last 3 goals,last 3 bps, last 3 saves. I then calculated the mean and std of some of these historical features (mean_features and std_features in the config.py file). Finally the last 3 stats features were dropped.
  
  Another type of features created were the team Goal scored, team Goal conceded, match_result, and the opp team Goal scored, opp team Goal conceded, opp match_result: this features were calculated (their average) for the last three games and start of the season, the features were created to show the form of the team and the opponent they are facing.

Get next gameweek fixtures

• weekly_fixtures.ipynb
  This file basically scrapes test data for the next gameeweek (players,teams , costs,home team ,away team, kick off time and co) from the FPL api.

Get gameweek results

• weekly_results.ipynb
  This file basically scrapes results and actual player performance on the previous/ just concluded gameweek from the FPL api.

Clean next gameweek fixtures

• clean_fixtures.ipynb
  This file cleans the scraped the data for the next gameweek and prepares it for modelling, new features were added too (as was done in the merged_previous_seasons.py file). The overall values the player had for some features in the last season (2022/23 season) were added.

Modelling

• train_model.ipynb
  

I built 2 different models

The first model is a classification model (f1_score and accuracy used as metric) that predicts whether a player will start a game or not.

The second model is a regression model (root mean squared error used as evaluation metric) that predicts the total points of players that played. The reason for this approach is a lot of players don’t play games at all and just predicting the points of all the players directly means our test dataset will have many 0’s which will strongly affect the quality of our regression model.

PLOTS

I then created bar plots for the top 10 players (with the highest predicted points) in each position on a colab notebook using matplotlib. Goalkeepers Defenders Midefielders Forwards

Feature importance plot

Feature importance

I intend on making predictions for every game week and I will be documenting the models prediction and progress here @fpl__ai.

Folders

• datasets folder
  The folder contains the raw uncleaned data webscraped directly from the FPL API and the FPL data repository. The fixtures and results subfolder contain the fixtures and results from the fixtures for each gameweek in the current season (2023/24).

• cleaned_datasets folder
  This folder contains the cleaned CSV files from the datasets folder, new features have been created and some unuseful features have been removed.

• predicted_dataset folder
  This folder contains the predictions made on the starting players and points by the model for every gameweek.