Leaving Goals on the Pitch: Evaluating Decision Making in Soccer

Maaike Van Roy, Pieter Robberechts, Wen-Chi Yang, Luc De Raedt, Jesse Davis

Abstract

Analysis of the popular expected goals (xG) metric in soccer has determined that a (slightly) smaller number of high-quality attempts will likely yield more goals than a slew of low-quality ones. This observation has driven a change in shooting behavior. Teams are passing up on shots from outside the penalty box, in the hopes of generating a better shot closer to goal later on. This paper evaluates whether this decrease in long-distance shots is warranted. Therefore, we propose a novel generic framework to reason about decision-making in soccer by combining techniques from machine learning and artificial intelligence (AI). First, we model how a team has behaved offensively over the course of two seasons by learning a Markov Decision Process (MDP) from event stream data. Second, we use reasoning techniques arising from the AI literature on verification to each team's MDP. This allows us to reason about the efficacy of certain potential decisions by posing counterfactual questions to the MDP. Our key conclusion is that teams would score more goals if they shot more often from outside the penalty box in a small number of team-specific locations. The proposed framework can easily be extended and applied to analyze other aspects of the game.

The Over-Correction of the xG Revolution

In the last decade, the "Expected Goals" (xG) revolution has fundamentally changed soccer. Because xG shows that long-distance shots have a low probability of scoring, teams have drastically reduced them—down nearly 20% in the English Premier League (EPL) since 2013.

However, researchers from KU Leuven argue this might be an over-correction. While a shot from 25 yards has a low xG, forgoing that shot to pass the ball carries its own risk: the team might lose possession before ever getting a better look. The paper tackles the "counterfactual" question: What would happen to a team's total season goal tally if they changed their shooting policy?

Methodology: Modeling Soccer as an MDP

The authors model soccer possessions as a Markov Decision Process (MDP):

  • State Space: The pitch is divided into a 22x17 grid in the offensive half, with a single state for the defensive half and three "absorbing" states (Goal, Missed Shot, Loss of Possession).
  • Action Space: At any location, a player can choose to "move" (pass or carry) or "shoot."
  • Transition Function: Learned from historical StatsBomb event stream data (EPL 2017/18 and 2018/19). A key innovation is using Gradient Boosted Trees to estimate the intended destination of failed passes, which standard event data doesn't record.
  • Probabilistic Model Checking: The researchers used AI verification tools (PRISM and STORM) to calculate the exact probability of a possession ending in a goal given a specific sequence of actions.

They simulated "what-if" scenarios by mathematically adjusting a team's probability of shooting vs. moving in specific zones, accounting for a "quantity-quality trade-off" (i.e., if you shoot more often, the average quality of those shots likely decreases).

Key Insights: The Passing Trap and Team-Specific Sweet Spots

The model reveals a "Passing Trap": relying on multiple passes to generate a "perfect" shot often increases the cumulative risk of losing the ball. The benefit of moving the ball starts to disappear rapidly as more actions are required before a shot.

Spatial shift in shooting behavior: more close shots, fewer long-distance attempts.

Furthermore, the "right" decision is not universal. For teams with elite long-range shooters (like Manchester City with Kevin De Bruyne or Chelsea with Eden Hazard), there are specific zones where shooting immediately is mathematically superior to passing. For struggling teams like Huddersfield Town, the probability of "ever" generating a better shot later in a possession was as low as 5% in many areas, meaning they almost always should have taken the shot they had.

Results and Impact

The study found that for most teams, a targeted 10-20% increase in shot frequency from specific "high-value" long-distance zones would lead to an increase of 0.5 to 1.5 goals per season. While a blanket "shoot more" policy helps some teams but hurts others (like Burnley or Newcastle), a targeted policy increased expected goals for every single team in the dataset.

This margin is significant; one extra goal per season roughly equates to one extra point in the league table. In the case of Bournemouth’s 2019/20 relegation, the model suggests a slightly more aggressive shooting policy could have provided the single goal needed to stay in the top flight.

Limitations and Future Work

The current model lacks "freeze-frame" context, such as defender positions or open teammates. The authors also acknowledge that their approximation of how shot quality drops as frequency increases is a simplified linear model. Future work could apply this framework to other tactical decisions, such as throw-in direction or the risks of "playing out from the back."

Summary updated

Abstract

Analysis of the popular expected goals (xG) metric in soccer has determined that a (slightly) smaller number of high-quality attempts will likely yield more goals than a slew of low-quality ones. This observation has driven a change in shooting behavior. Teams are passing up on shots from outside the penalty box, in the hopes of generating a better shot closer to goal later on. This paper evaluates whether this decrease in long-distance shots is warranted. Therefore, we propose a novel generic framework to reason about decision-making in soccer by combining techniques from machine learning and artificial intelligence (AI). First, we model how a team has behaved offensively over the course of two seasons by learning a Markov Decision Process (MDP) from event stream data. Second, we use reasoning techniques arising from the AI literature on verification to each team's MDP. This allows us to reason about the efficacy of certain potential decisions by posing counterfactual questions to the MDP. Our key conclusion is that teams would score more goals if they shot more often from outside the penalty box in a small number of team-specific locations. The proposed framework can easily be extended and applied to analyze other aspects of the game.
BibTeX
@misc{roy2021leaving,
  author = {Maaike Van Roy and Pieter Robberechts and Wen-Chi Yang and Luc De Raedt and Jesse Davis},
  title = {{Leaving Goals on the Pitch: Evaluating Decision Making in Soccer}},
  year = {2021},
  eprint = {2104.03252},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.03252},
}
Figure