Expected possession value of control and duel actions for soccer player’s skills estimation

Abstract

The paper presents enhancements to the Expected Possession Value (EPV) model for soccer, addressing challenges such as selection bias. The improved model places greater weight on events leading up to a shot (decay effect), accurately incorporates possession risk, and evaluates individual players’ ability to win duels. Using these improvements, the model predicts players’ performance for the upcoming season, considering opponents’ strength.

Introduction

Evaluating soccer players’ skills is crucial for managers, scouts, and analysts, but traditional methods like video analysis are impractical for a large pool of players. Hence, metrics correlated with player performance are more effective for selection. Soccer analytics has advanced in estimating players’ ball possession skills, moving from basic statistics (goals, passes) to advanced metrics. These metrics, however, rely on historical data, making future performance predictions challenging, especially in diverse leagues where player adaptation is key.

Related Work

The paper references J. Hollinger’s Player Efficiency Rating (PER) for the NBA, which uses basic statistics to evaluate players, introducing the concept of possession value to reward or punish player actions.

Proposed Approach for EPV

Background

The model defines possession as a series of uninterrupted control actions by a team. Effective playing time (active gameplay time) is used instead of dirty playing time (total game time) for more accurate metrics.

Possession Value of Control Actions

The EPV model’s target variable is the possession value, predicted as the expected possession value (EPV). Traditionally, EPV is defined by assigning 1 for a goal and 0 otherwise, but this approach is avoided due to overfitting risks. Instead, EPV is defined by the cumulative sum of goal probabilities (xG) from future shots within the same possession.

Decay Effect

To address the drawback of equal importance for all possession events, a decay effect is introduced. Actions leading to shots are weighted more heavily, modifying the EPV formula to reflect this.

Possession Risk

Possession risk considers the probability of losing possession during actions. The model assigns penalties for turnovers, especially when leading to goal opportunities for the opponent.

Possession Value of Symmetrical Duels

The model calculates EPV for symmetrical duels (aerial and ground duels). The possession value of the first control action after a duel is assigned to that duel, recursively applied for a series of duels. The reward for duel outcomes considers the probability of winning, adjusting rewards based on the mismatch between players.

Reward Metrics

Rewards for player actions are categorized into five outcomes: control by the same team, control by the opponent, a goal, the end of the half, and symmetrical duels. Rewards and penalties are assigned based on the difference in EPV before and after actions.

EPV Implementation

Expected Goals (xG)

The xG model predicts the probability of a shot resulting in a goal, factoring into the EPV calculation.

Symmetrical Duels

EPV for symmetrical duels incorporates the probability of winning and the difference in outcomes between the actual and average player to avoid inflating EPV for superior skills.

Metrics Prediction for the Upcoming Season

The model predicts metrics like season pass carry reward, using training sets and features such as club and league ratings. It addresses challenges like presence-only data and the probability of players staying in the dataset.

Results

The results section, though not detailed in the provided text, presumably presents the findings of the EPV model’s predictions and their accuracy compared to actual player performances.

Conclusion

The paper concludes by summarizing the enhancements to the EPV model, emphasizing its improved accuracy in predicting player performance by incorporating the decay effect, possession risk, and duel outcomes. These improvements make the model a valuable tool for soccer analysts in evaluating player skills and making informed decisions.

Acknowledgments

Acknowledgments are typically given to contributors, funding sources, or institutions that supported the research, though specific details are not provided in the text.

Case Study: Gianluigi Donnarumma

A case study on Gianluigi Donnarumma’s passes to duels over two seasons demonstrates the application of the EPV model. It highlights how the model evaluates and predicts player performance based on historical data and the newly introduced metrics.


Source:

DOI: https://doi.org/10.48550/arXiv.2406.00814