Valuing Players Actions in Football

This data science project aims to value all football player actions on the pitch. This provides a better view over how different actions contribute to goals, since most models only consider single actions or shots. This project builds up on the paper Actions Speak Louder than Goals by Decroos et al. 2019 and the corresponding Github repository. Some ideas were also inspired by Statsbombs On Ball Value metric.

The entire code for the project can also be found in this Github repository.

General Idea

Generally we consider a good action in football as an action that either increases scoring probability or decreases conceding probability or does both. So we can define the action value of an action with the following formula:

VAEP value(a_i) = ΔP_scores(a_i) − ΔP_concedes(a_i)
ΔP_scores(a_i) = P_scores(S_i) − P_scores(S_i−1)
ΔP_concedes(a_i) = P_concedes(S_i) − P_concedes(S_i−1)

Data

We use the statsbomb data from the mplsoccer python library, this data contains 200 games from the UEFA Euro 2020, FIFA World Cup 2022, 1. Bundesliga 23/24 and UEFA Euro 2024. For these compoetitions the data contains all game states at the time of an action. Each game state contains informations about the action, the acting player and the positions of the other players in the camera frame.

Features

We use a variety of features to estimate the probabilities in the model:

• Distance to Goal: the distance of the ball to the opponents / own goal
• Angle to Goal: the angle of the ball to the opponents / own goal
• Time Elapsed: the time elapsed in the game since the start of the game
• Duration: the duration of the action
• Score Difference: the score difference at the time of the action
• Body Part: the body part with which the action was performed
• Action Type: the type of the action
• Closest Defender Distance: the distance of the ball to the closest defending player
• Opponents in Front: the number of opponents in front of the ball with respect to the x coordinate
• Goalkeeper Distance to Ball: the distance of the opponents / own goalkeeper to the ball
• Goalkeeper Distance to Goal: the distance of the opponents / own goalkeeper to the goal
• Goalkeeper in Shooting Triangle: whether or not the opponents / own goalkeepeer is in the triangle between the ball and posts

Labels

For probability estimation we label the 10 actions previous to a goal. Actions from the team scoring the goal will receive a 1 for the scoring label and actions from the conceding team will receive a 1 for the conceding label. All other actions will be labelled with a 0.

Model

We use an XGBoost classifier to estimate the scoring and conceding probabilities given the features. The model outputs the following:

• predicted_goal_prob: Probability of scoring from the event.

• predicted_concede_prob: Probability the team concedes after the event.

• action_value: Estimated contribution of the action to the overall chance of scoring.

Project Structure and Setup

Project Structure

The files in the project are required to be run in the following order:

• 1_Loading_Data.ipynb
  Loads the data, constructs the features, assigns the labels, computes player minutes and stores everything as data_cleaned.pkl and player_minutes.pkl.

• 2_Training_Model.ipynb
  Trains the model, estimates the probabilities, computes action values and stores the data in data_cleaned_trained.pkl.

• 3_Result_Visualization.ipynb
  Visualizes the model results and shows various potential applications with examples.

Applications

1. Player Ratings

The model can rate all players in the dataset by their action value per 90 minutes.

2. Event Snapshot Plotting

Plots a single event on a pitch, including visible area and freeze-frame positions of players.

plot_event_with_360(event_row)

3. Action Chain Visualization

Generates a table and pitch plot for a sequence of actions (passes, carries, shots, etc.) leading up to a key moment.

plot_action_chain_by_id(df_model, SCENARIO)

4. Action Chain Animation

Creates an animated pitch plot of action sequences with:

• Player locations

• Arrows for passes/dribbles

• Dynamic predicted goal/concede probabilities

action_chain_animation.mp4

animate_action_chain(df_model, SCENARIO, save_path='action_chain_animation.mp4')

5. Team (or Player) Goal Probability Heatmap

Creates a pitch heatmap showing areas from which actions resulted in high predicted probabilities of scoring.

plot_team_goal_prob_heatmap(team_id, player_id=None)

Requirements

• pandas

• matplotlib

• mplsoccer

• numpy

• cmasher

• sklearn

• xgboost

• collections

Install dependencies with:

pip install pandas matplotlib mplsoccer numpy cmasher sklearn xgboost collections