Quantifying Passing Decision Quality Under Pressure
Introduction
Soccer is a game of decisions under pressure. On every touch, a player must quickly identify the available passing options and choose the best one while defenders close space and block passing lanes.
Most existing metrics evaluate the outcome of a pass, such as whether it was completed or how much threat it created. This creates outcome bias: we cannot tell whether the player chose the right option, whether that option was a hard one, or whether the pass was completed through skill rather than luck (Kang & Narasimhan, 2026).
Recent studies consider the full set of passing options available to a player. Goes et al. (2022) estimate the risk and reward of the selected pass and realistic alternative options. However, less is known about which defensive, spatial, and contextual factors explain why the highest-threat-available option is not selected,and whether some players choose better or worse than their situations would suggest.
This situation shows a counterattack. The ball carrier drives through the centre toward goal with two passing options available: one to his left and one to his right. The animation stops at the moment of the passing decision.
The main question is: Which passing option is more dangerous, and will the player choose it?
The second animation shows how the play ended. The pass was completed and the possession finished with a shot, so traditional metrics would consider it successful. However, the option on the left was more dangerous. A better opportunity was available but not chosen, and outcome based metrics do not capture that.
Research Questions:
Which contextual, defensive, and spatial factors influence whether an attacker selects the highest-threat passing option?
Which players consistently choose more threatening or safer passes than expected?
Data
Source: The dataset used in this analysis was provided by SkillCorner. It contains data from 520 matches played in Major League Soccer (MLS) during the 2023 season. For each match, we had three sets of data: Event Data, Tracking Data, and Match Information Data. Event Data consists of detailed information on both on-ball and off-ball activities, including ball carrier coordinates, shots, goals, fouls, game state, interceptions, possession outcomes, and many other variables for four main event types: Player Possessions, On-Ball Engagements, Passing Options, and Off-Ball Runs. Each event has a start frame, end frame, and timestamp, which can be linked to the tracking data. Tracking Data provides the ball coordinates and the coordinates of all 22 players on the field at 10 frames per second (fps). Match Information contains the home and away teams, stadium, date, competition details, and information for all players, including substitutes. 18 matches did not have high-quality event data; therefore, 502 matches were included in this study.
Preprocessing: Since we were interested in analyzing passing decisions, we joined the Player Possessions data with the Passing Options and On-Ball Engagements data for each event. We also linked the event data with the tracking data to extract important defensive features. The event data was standardized so that the attacking team always attacked from left to right. To ensure consistency, we also standardized the attacking direction in the tracking data.
Methods
To answer our research question, we first needed a way to quantify passing decisions. To do this, we used the outputs of the SkillCorner models available in our dataset:
Expected Receiver Model: This model identifies realistic passing options based on the passing option score. The passing option score is the probability that a player will be the target of a pass. It is calculated using a Graph Neural Network (GNN) model that considers features such as the ball carrier’s coordinates, speed, movement angle, inter-player distance, and teammates’ positions. For an option to be considered a passing option, it must have a passing option score > 0.6.
Xthreat Model: This model estimates the expected threat (xThreat), defined as the probability of a goal occurring within 10 seconds if a player receives the pass. Using a similar GNN model trained on the same features as the Expected Receiver Model, it estimates the xThreat value for all passing options.
Passing Threat Reduction (PTR) Metric: For each passing situation, we calculated the Passing Threat Reduction (PTR), defined as the difference between the maximum threat available and the xThreat of the pass actually chosen by the ball carrier. We also applied a realistic filter by considering only passing options with a completion probability (xPass) > 0.68. This allows us to evaluate whether players select the highest-threat option without penalising them for passes that were unlikely to succeed. The threshold follows SkillCorner’s criterion for a valid line-breaking option, which requires a high probability of completion. A PTR value of 0 indicates that the ball carrier selected the highest-threat passing option available in that situation. A PTR value > 0 indicates that the ball carrier chose a safer pass despite a higher-threat passing option being available.
PTR = max(Xthreat_{realistic}) - Xthreat_{chosen}
Response Variable: We formulated our response variable as a binary outcome to determine whether the attempted pass was optimal or sub-optimal. We defined a pass as optimal if the ball carrier chose the best (highest-threat) passing option and sub-optimal if the ball carrier did not choose the highest-threat passing option and instead selected a safer option.
PTR_{binary} = \begin{cases} 1, & \text{PTR } > 0 & \text{sub-optimal pass = high threat pass} \\ 0, & \text{PTR } = 0 & \text{optimal pass = safe pass} \end{cases}
Visualising PTR:
The solid arrow shows the pass that was played, while the dashed arrow shows the highest threat realistic option available.
In the left example, the ball carrier selected the best available option. The played pass had an xThreat of 0.052 and an xPass of 92%, so the situation was classified as PTR = 0.
In the right example, the ball carrier chose a pass with an xThreat of 0.044 and an xPass of 42%. However, another teammate was available with a much higher xThreat of 0.435 and an xPass of 88%. Because this alternative was both more dangerous and more likely to be completed, the ball carrier should have selected that option. This situation was therefore classified as PTR = 1, meaning the highest threat realistic option was not chosen.
Spatial Features: In addition to the variables provided in the SkillCorner event and tracking data, we added some spatial features to better explain the defensive context during each passing situation. These features were designed to capture defensive pressure, defensive organization, and the spatial relationship between defenders, the ball carrier, and the best available passing option.
- Defensive Midfield Spread:To evaluate the compactness of the defending team, we calculated the spread of the defenders and midfielders at the passing moment. We chose defenders and midfielders because they are usually responsible for maintaining the overall defensive structure. Using the similar idea used by Forcher et al., the spread was calculated as the sum of the squared Euclidean distances between each defender or midfielder and and the centroid of the defending unit:
S = \sum_{i=1}^{n} \left[ (x_i - \bar{x})^2 + (y_i - \bar{y})^2 \right] where (x_i,y_i) is the position of defender or midfielder i, (\bar{x}, \bar{y}) is the centroid of the defense, and n is the number of players included in the calculation. The smaller value of S indicates that defense was more compact whereas larger value indicates more spreaded defense structure.
- Passing Lane: For quantifying how open the best available passing option from the ball carrier, we defined the idea of passing lane. Based on the idea of the line corridor proposed by Cascioli et al., we modeled the passing lane as a fixed-width rectangular corridor centered on the line segment that connects the ball carrier p_a and the best available passing option r_i. The corridor is defined as:
R_w(p_a,r_i) = \left\{ x \in \mathbb{R}^2 \;\middle|\; \operatorname{perp\_dist}\!\left(x,L(p_a,r_i)\right)<w/2 \right\}. where L(p_a,r_i) denotes the line segment connecting the ball carrier and the best available passing option. We used the lane width (w) of 4 m in our analysis considering the variation in player and ball positions.
Unlike the original line corridor approach, which classified a passing lane as blocked or not, we used the corridor to derive two continuous measures of defensive coverage:
Nearest defender distance to the passing lane: minimum perpendicular distance between any defender and the passing lane.
Number of defenders in the passing lane: Total number of defenders inside the lane.
Predictor Variables: After data preprocessing, correlation analysis, and feature engineering, 20 variables were selected for modeling:
- Match context: game state, possession duration
- Ball carrier context: location, playing position, Nearest defender distance
- Defensive pressure: Pressing (group pressure), pressure, defensive organization
- Passing options: Dangerous options, line-breaking opportunities
- Best passing option: Distance, defensive pressure, passing lane features
Analysis
Because our response variable was binary, we began by fitting a baseline Logistic Regression Model. The baseline logistic model is given by:
\log\left(\frac{E(Y \mid X=x)}{1 - E(Y \mid X=x)}\right) = \beta_0 + \sum_{j=1}^{p} \beta_j x_j. where E(Y | X=x) denotes the probability of sub-optimal pass given the predictor variables, x_j represents the j-th predictor, and \beta_j are corresponding regression coefficients.
Generalized Additive Model (GAM)
The logistic regression assumes linear relationship between all of its predictor variables and response. But we wanted to include the ball carrier’s location, which is inherently nonlinear, so we extended the baseline logistic regression using a Generalized Additive Model (GAM). This allowed us to add the carrier’s x- and y-coordinates as a smooth two-dimensional function,
\log\left(\frac{E(Y \mid X=x)}{1 - E(Y \mid X=x)}\right) = \beta_0 + s(x_{\text{carrier}}, y_{\text{carrier}}) + \sum_{j=1}^{p} \beta_j x_j. where s(⋅) is a smooth spatial function. Now, the effect of the ball carrier’s position could vary nonlinearly across different regions of the pitch while the linear effects for the remaining predictors remains the same.
Generalized Additive Mixed Model (GAMM)
Our possession level dataset had multiple passes by same players. To account for repeated observations and unseen differences in players’ passing behavior, we extended our GAM model by including a player-specific random intercept u_{player}, resulting in a Generalized Additive Mixed Model (GAMM):
\log\left(\frac{E(Y \mid X=x)}{1 - E(Y \mid X=x)}\right) = \beta_0 + s(x_{\text{carrier}}, y_{\text{carrier}}) + \sum_{j=1}^{p} \beta_j x_j + {u_{player}}. where u_{player} ~ N(0,\sigma_{u}^2) . In the model implementation, this random effect allows each player to have their own baseline probability of making a sub-optimal pass after accounting for the fixed effects. We also tested the random effect of the defending team, but its variance was negligible compared with that of the individual player. Therefore, we decided to exclude it to reduce the complexity of our model.
Extreme Gradient Boosting (XGBoost)
We finally compared all the statistical models with the XGBoost which is capable of capturing all the nonlinear interaction between the variables. We did this to assess predictive benchmark of the model and get idea whether allowing complex machine learning model would improve the prediction performance.
Model Evaluation
All models were then evaluated using five-fold cross-validation at the match level to prevent data leakage between the training and testing sets. The average performance across the five folds is summarized in Table 1. Model performance was assessed using Accuracy, Precision, and Area Under the ROC Curve (AUC).
As model complexity increased, predictive performance improved, with XGBoost achieving the highest accuracy and precision by capturing nonlinear relationships and interactions between predictors without specifying them in advance. However, the Generalized Additive Mixed Model (GAMM) achieved comparable performance and since it provides interpretable estimates of predictor effects, we selected the GAMM as the final model for inference.
Results
Above table summarizes the estimated odds ratios from the final Generalized Additive Mixed Model. Factors with OR > 1 increases the likelihood of selecting the highest-threat pass and OR < 1 decreases the likelihood of selecting the highest-threat pass. Overall, defensive positioning around the best passing option had the strongest influence on whether attackers selected the optimal pass. Measures of defensive pressure, passing lane coverage, and defensive organization were also significant predictors of passing decisions. We decided to use the odds ratio table from the GAAM instead of the variable importance scores from XGBoost because the odds ratio table provides a clearer interpretation of both the magnitude and the direction (positive or negative) of each variable’s effect on the probability of a high-threat pass.
Differences Between Midfield Roles:
We used XGBoost because it can capture complex relationships between defensive pressure, available passing options, and spatial context.The model estimates what a typical player in our dataset would be expected to choose in each situation based on its context. We then compare this expected rate with the player’s actual decisions, allowing for a fairer evaluation than raw passing percentages alone. We were interested in whether passing decisions differed across midfield roles, and our results show that they do.
The model expected both midfield roles to choose a lower threat pass in about 59% of possessions. Attacking and central midfielders did so only 56.7% of the time, meaning they selected the highest threat realistic option 2.6 percentage points more often than expected. Defensive midfielders selected lower threat passes 60.6% of the time, 0.9 percentage points more often than expected.
By taking into account the difficulty and context of each situation, the model reveals decision making patterns linked to player roles. Defensive midfielders often prioritize possession and control, while attacking and central midfielders are more likely to pursue higher threat opportunities.This suggests that passing choices may reflect a player’s role on the field.
Player Evaluation:
We focused on attacking midfielders because they are often their team’s main creative playmakers. They decide where to move the ball and are responsible for creating dangerous attacking situations.
Our model compares each player’s actual decisions with what was expected given the context of each situation. Players shown in green selected the highest threat realistic option more often than expected, which may reflect a more creative and aggressive passing style. For example, Emanuel Reynoso performed 14.4 percentage points above expectation.
Players below expectation selected lower threat options more often. This does not necessarily mean they were worse players. Their decisions may reflect their tactical role, position, or team strategy. Therefore, this measure describes passing decision style, not overall player quality.
The results also provide face validity for our model. Riqui Puig, an established playmaker, selected the best option 8.7 percentage points more often than expected and ranked among the MLS leaders in touches and successful passes into the attacking third. In contrast, Dániel Gazdag’s profile is more focused on finishing attacks inside the penalty area.
Goals and assists show what happened after the decision, while our model shows which players consistently identify the most threatening realistic available option.
Discussion
Our framework evaluates passing decisions relative to the options available at the passing moment and identifies the defensive, contextual, and spatial factors influencing selection of the highest-threat option. Comparing actual choices with model predictions among attacking midfielders provides a fairer assessment of individual passing styles within the same position.
The results show that defensive positioning is can influence the passing decisions and can reduce attacking threat even without resulting in a turnover or interception. We also found meaningful differences across midfield roles. Attacking and central midfielders selected higher threat options more often than expected, while defensive midfielders chose lower threat options more often. These patterns are consistent with their tactical responsibilities and suggest that passing choices reflect a player’s role, not only individual quality.
By combining tracking and event data with interpretable statistical models, this framework provides a quantitative way to evaluate passing decisions in context.This framework also provides a quantitative approach for evaluating passing decision-making while accounting for the context of each passing situation. By combining interpretable statistical models with tracking and event data, this approach offers a foundation for future analyses of both attacking decision-making and defensive effectiveness
For coaches, it shows which situations players struggle to read and how defenders can limit dangerous passing options. For scouts, it helps identify players who take more risks or prefer safer choices. For data scientists, it offers a new way to measure decision quality in context. It can also serve as a starting point for quantifying soccer intelligence by measuring whether a player consistently recognizes and selects the best available option.
Limitations:
- Only pass ending possessions were analyzed, excluding prior actions.
- Results depend on SkillCorner tracking, xPass, and xThreat estimates.
- Defensive influence is approximated using event and tracking features rather than a complete pitch control model.
- The study covers one MLS season, limiting season to season comparisons.
Future Work:
- Model continuous PTR for suboptimal passes using other statistical models like Zero Inflated Beta.
- Create team level analysis of attacking and defensive styles.
- Test how sensitive the results are to different xPass thresholds.
- Develop defender level evaluation to measure which players most effectively remove or reduce high threat passing options.
- Add pitch control to better quantify defender influence.
- Validate across additional MLS seasons, the NWSL, and other leagues.
Real World Applications of Our Approach:
- Sport: Models compare a player’s chosen pass with higher value alternatives to evaluate decision quality.
- Business: Systems rank every possible offer for a customer and pick the best one.
- LLMs: Compare possible responses and select the one expected to be best.
Acknowlegement
Special thanks to Daniel Wicker, Dr. Yurko, Erin Franke, Sara Colando, Quang Nguyen the CMSACamp TAs, and the CMU Statistics Department for their guidance and support, and for making this summer research experience possible, successful, and enjoyable
References
Cascioli, L., Wang, A., Stradiotti, L., Van Roy, M., Robberechts, P., Wouters, M., Jaspers, A., & Davis, J. (2025). Quantifying off-ball defensive impact through cover shadows. In Proceedings of the 2025 Hudl Performance Insights Conference. https://static.hudl.com/craft/performance-insights-research-stage/2025/Quantifying-Off-Ball-Defensive-Impact-through-Cover-Shadows-Lorenzo-Casciolio-Allen-Wang_amended.pdf
Forcher, L., Forcher, L., Altmann, S., Jekauc, D., & Kempe, M. (2024). Is a compact organization important for defensive success in elite soccer? Analysis based on player tracking data. International Journal of Sports Science & Coaching, 19(2), 757–768. https://doi.org/10.1177/17479541231172695
Goes, F., Schwarz, E., Elferink-Gemser, M., Lemmink, K., & Brink, M. (2022). A risk-reward assessment of passing decisions: Comparison between positional roles using tracking data from professional men’s soccer. Science and Medicine in Football, 6(3), 372–380. https://doi.org/10.1080/24733938.2021.1944660
Kang, A., & Narasimhan, P. (2026). Monte Carlo pass search: Using trajectory generation for 3D counterfactual pass evaluation in football [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2606.11120
Li, L., Chu, W., Langford, J., & Schapire, R. E. (2010). A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th International Conference on World Wide Web (pp. 661–670). Association for Computing Machinery. https://doi.org/10.1145/1772690.1772758
Masella, G., D’Inverno, G. A., Goldsmith, M., & Rozza, G. (2026). Evaluating passing decision-making in professional football: An enhanced MPNN approach to receiver selection [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2605.25696
Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M., Chess, B., & Schulman, J. (2021). WebGPT: Browser-assisted question-answering with human feedback [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2112.09332
Contact Information
Julia Biesiada, Indiana University Indianapolis, julia.biesiada.analytics@gmail.com
Swayam Chaulagain, University of Southern Mississippi, swayam2060@gmail.com
Code Availability
Code available on GitHub
Apendix
The following supplementary figures and information provide additional support for our analysis.
