Can Data Tell If Two Players Will Click on the Pitch?
Measuring and Predicting Football Chemistry from Match Event Data
The following summary critically reviews the research paper titled “Player Chemistry: Striving for a Perfectly Balanced Soccer Team” by Lotte Bransen and Jan Van Haaren. All data, figures, and analysis presented here are drawn from their original work; I do not claim any authorship or ownership of the content. This summary has been written to provide a concise and technically informed synthesis of the paper’s findings, methodologies, and implications, while maintaining fidelity to the authors’ intellectual contributions.
1. Introduction
The paper opens with a concrete and well-known transfer example to motivate the concept of player chemistry. The contrasting Liverpool signings of Andy Carroll and Luis Suárez during the 2010/2011 winter transfer window illustrate how individual quality alone does not guarantee success. Carroll, described as “a tall target man thriving from crosses,” struggled in a team whose playing style did not suit him, while Suárez “gelled much better with the players surrounding him on the pitch” and delivered consistently strong performances. This contrast frames the central argument: recruitment decisions typically ignore team balance and chemistry, despite their obvious influence on outcomes.
The authors argue that both traditional scouting and most existing analytics tools focus almost exclusively on individual performance. Assessing how well players fit together is difficult through limited visual observation, and most large-scale analytics frameworks are not designed to capture interactions. However, recent work in soccer and basketball has begun to address interactions and cooperation in team sports. Building on this emerging literature, the paper asks a precise question: “How well does a team of soccer players gel together?”
Two complementary settings are introduced. In the observational setting, chemistry is measured for players who have already played together, which is relevant for line-up selection. In the predictive setting, chemistry is estimated for players who have never played together, which directly supports recruitment decisions. The key assumption underpinning the paper is explicit: “two players who have high mutual chemistry perform better than players who have low mutual chemistry, all other things being equal.”
The introduction concludes by outlining four main contributions: the definition of offensive and defensive chemistry metrics, machine-learned models to predict these metrics for unseen player pairs, a Team Builder that assembles line-ups maximizing chemistry, and a set of empirical observations and applied use cases.
2. Dataset
The dataset consists of Wyscout match event data from 106 domestic and international competitions starting in the 2015/2016 season. The authors justify the use of event data by emphasizing its global availability and balance between coverage and granularity, making it “ideally suited for developing performance and style metrics that are aimed at player recruitment.”
All events are converted into the SPADL representation [6], which standardizes actions such as passes, dribbles, shots, and take-ons while preserving temporal and spatial context. Each action includes identifiers for the match, team, and player, timing information, start and end locations, outcomes, and body parts used.
The dataset is further enriched in two ways. First, player-level attributes are added using proprietary SciSports indicators, including current ability (SciSkill), potential, Player Roles, and physical performance metrics such as speed, duel strength, and work rate [7,8]. Second, relational and contextual features are computed for each pair of players, including shared nationality, language, region, previous competition experience, and the number of matches played together at club and international level.
In total, the dataset spans 361 seasons, over 106,000 matches, more than 38,000 players, and nearly 100,000 unique player-season combinations, as summarized in Table 1. This scale underpins the statistical robustness of the proposed metrics.
3. Measuring the mutual chemistry for a pair of players
The authors begin by restating the core methodological challenge: chemistry is not directly observable. Instead, it must be inferred from joint performance. Since no established metric exists for pairwise performance, the paper introduces two new measures built on top of the VAEP framework [6], which assigns values to individual on-the-ball actions based on their impact on scoring and conceding probabilities.
Two distinct metrics are proposed: Joint Offensive Impact (JOI) and Joint Defensive Impact (JDI). Together, they aim to capture how pairs of players contribute to attacking success and defensive stability.
3.1. Joint Offensive Impact (JOI)
JOI is designed to quantify how two players jointly increase their team’s likelihood of scoring. The intuition is straightforward: when players interact effectively, their combined actions produce more offensive value than isolated contributions.
3.1.1. Computing the offensive impact for interactions
A match is modeled as an ordered sequence of actions, each performed by a player and belonging to one of five action types: pass, cross, dribble, take-on, or shot. An interaction is defined as two consecutive actions performed by two different players. For example, a pass followed immediately by a take-on by the receiving player constitutes a single interaction.
The VAEP value of an interaction is defined as the sum of the VAEP values of its two constituent actions. Formally, for an interaction
its value is
This extension allows the framework to move from individual actions to dyadic cooperation.
3.1.2. Computing the joint offensive impact in a season
The JOI for a player pair in a match is computed by summing the VAEP values of all interactions involving the two players, regardless of who initiates the sequence. To ensure comparability across pairs with different shared playing time, the metric is normalized per 90 minutes, yielding JOI90. This normalization divides total joint impact by the total minutes the pair spent together on the pitch and scales the result to a standard match duration.
3.2. Joint Defensive Impact (JDI)
Defensive chemistry poses a greater challenge because event data records what happens, not what is prevented. The authors therefore adopt an indirect approach. They argue that if an opponent consistently underperforms relative to expectation, then the defenders responsible for that opponent must have contributed positively.
JDI captures how pairs of players jointly suppress the offensive output of their opponents, again leveraging VAEP as the underlying valuation framework.
3.2.1. Computing the offensive impact above or below expectation for each opponent
For each opponent in a match, actual offensive impact is computed as the sum of VAEP values for all attacking actions. Expected offensive impact is estimated from the player’s earlier matches in the same season, normalized per 90 minutes. To address small-sample issues, a Bayesian shrinkage approach is applied for players with fewer than 700 minutes, blending individual estimates with position-specific priors.
The difference between expected and actual offensive impact indicates whether a player overperformed or underperformed in the match.
3.2.2. Computing the responsibility share for each pair of players and opponent
Since multiple defenders can influence the same opponent, responsibility is distributed based on positional proximity. Without tracking data, the authors rely on default positions and map them onto a 5-by-5 grid of the pitch. Responsibility shares are inversely proportional to the Euclidean distance between positions. For a player pair, the responsibility share is defined as the average of the two individual shares.
3.2.3. Computing the joint defensive impact in a season
The JDI for a player pair in a match is computed by summing, across all opponents, the product of three terms: the opponent’s underperformance relative to expectation, the pair’s responsibility share for that opponent, and a scaling factor based on shared minutes. As with JOI, the final metric is normalized per 90 minutes to obtain JDI90.
4. Predicting the mutual chemistry for a pair of players
While JOI and JDI are informative for established partnerships, they cannot be computed reliably for players who have not played together. To address this, the authors introduce predictive models that estimate JOI90 and JDI90 for arbitrary player pairs.
4.1. Representing players as feature vectors
Each player is represented using a feature vector capturing demographic attributes, physical characteristics, positional information, nationality and language, physical performance indicators, and detailed Player Role scores. Player pairs are represented by combining individual features and adding relational features such as shared nationality, shared language, region, and prior matches played together.
4.2. Predicting the joint impact for a pair of players
Two regression models are trained using CatBoost [10], one for offensive chemistry and one for defensive chemistry. The dataset is split temporally into training, validation, and test sets. Only pairs with at least 700 minutes played together are used as training examples, resulting in hundreds of thousands of samples per split.
JOI90 and JDI90 serve as target variables. Both models outperform a simple baseline that predicts the average chemistry value, as shown in Table 3. The improvement is substantial for offensive chemistry and modest but consistent for defensive chemistry.
4.3. Analyzing the machine-learned models
Feature importance analysis reveals that Player Role scores are among the most influential predictors. Roles such as Mobile Striker, Deep-Lying Playmaker, and Ball-Playing Defender are important for offensive chemistry, while Holding Midfielder and Ball-Winning Defender roles matter more for defensive chemistry.
Interestingly, cultural features such as shared nationality or language have limited predictive power. The authors interpret this as evidence that “it is more important that two players communicate well with their feet and that they have complementary playing styles.”
5. Automatically assembling a maximum-chemistry team
Building on the ability to estimate chemistry for any player pair, the paper introduces a Team Builder that assembles an eleven-player line-up maximizing total chemistry. The problem is formulated as a mixed-integer programming task, where the objective function combines offensive and defensive chemistry with a user-defined weighting parameter.
Constraints enforce realistic squad composition, including exactly eleven players and minimum and maximum numbers of defenders, midfielders, and forwards. This framework allows practitioners to explore trade-offs between offensive and defensive balance while explicitly accounting for pairwise interactions.
6. Observations
After introducing metrics and predictive models for player chemistry, the authors turn to large-scale empirical analysis. By computing Joint Offensive Impact and Joint Defensive Impact across all seasons in the dataset, they explore whether the proposed metrics produce interpretable, football-relevant patterns. The observations presented in this section serve two purposes. First, they validate that the chemistry metrics align with well-known partnerships and tactical structures. Second, they show how chemistry evolves over time and responds to contextual changes within teams.
6.1. Highest-chemistry player pairs
The authors begin by ranking player pairs according to their Joint Offensive Impact per 90 minutes (JOI90) across all seasons, restricting the analysis to pairs who have played at least 900 minutes together. The resulting list, shown in Table 4, is topped by Mohamed Salah and Roberto Firmino during Liverpool’s 2017/2018 Champions League campaign. This partnership is described as exhibiting “the highest mutual chemistry between two players” in the dataset. Other high-ranking pairs include Luis Suárez and Lionel Messi at Barcelona, as well as Yuki Nagasato and Sam Kerr in the NWSL, highlighting that the metric generalizes across leagues and genders.
The authors then narrow the scope to specific competitions and seasons. Table 5 presents the highest-chemistry offensive pairs for selected leagues during the 2018/2019 season. Examples such as Karim Bellarabi with Kevin Volland at Bayer Leverkusen, and Jozy Altidore with Alejandro Pozuelo in MLS, demonstrate that the metric captures productive attacking relationships that have also been recognized in media and tactical commentary.
A similar analysis is performed for defensive chemistry in Table 6, where high-ranking pairs often consist of defenders or defensive midfielders whose positional responsibilities naturally overlap.
Across these rankings, the authors emphasize that chemistry is not confined to globally famous duos. Lesser-known players and lower-tier competitions also feature prominently, suggesting that JOI90 and JDI90 capture structural interaction effects rather than reputation-driven signals.
6.2. Mutual chemistry for the 2018/2019 Liverpool side
To move beyond isolated pairs, the authors analyze chemistry within a full team context, focusing on Liverpool’s 2018/2019 Premier League season. This season is chosen due to Liverpool’s exceptional performance and tactical coherence.
Figure 1 visualizes mutual offensive chemistry between Liverpool’s regular starters, with green lines indicating strong links and dark red lines indicating weak ones. The visualization reveals “very strong offensive chemistry links among the three attackers” as well as strong connections between the full backs, Andrew Robertson and Trent Alexander-Arnold, and the attacking players. This reflects Liverpool’s tactical structure, where full backs play a key creative role in advancing the ball and creating chances.
Figure 2 presents the corresponding defensive chemistry network. Strong defensive links are observed in midfield and central defense, while weaker links appear between central defenders and full backs. The authors note that some links are omitted because certain player pairs were “not responsible for defending the same opponent for long enough,” highlighting how the defensive metric depends on shared defensive responsibility rather than mere co-presence on the pitch.
Overall, the Liverpool case study illustrates how the chemistry metrics recover intuitive tactical patterns: attacking cohesion among forwards, creative interaction between full backs and attackers, and defensive stability concentrated in central areas.
6.3. Mesut Özil’s chemistry at Arsenal over time
The final observation examines chemistry dynamics longitudinally, using Mesut Özil’s tenure at Arsenal as a case study. Özil’s performances were widely praised in his early seasons but increasingly criticized in later years, making him a natural subject for analyzing changes in player chemistry.
Figure 3 shows the evolution of Özil’s average JOI90, compared to Arsenal’s team average and the Premier League average, from the 2015/2016 season onward. The authors observe that both Özil’s and Arsenal’s offensive chemistry “have rapidly declined since Alexis Sánchez’ departure in 2018.” By the 2019/2020 season, Arsenal’s top chemistry pairs fail to reach even the league-average level.
Figure 4 further reinforces this finding by listing the top-ten Arsenal player pairs in terms of JOI90 since 2015/2016. Özil appears in six of these ten pairs, including the top four, with the Özil–Sánchez partnership appearing multiple times. This suggests that Özil’s effectiveness was closely tied to specific teammates, and that the removal of complementary players can materially affect a player’s joint impact.
7. Use cases
Having demonstrated that the chemistry metrics produce meaningful observational insights, the authors turn to applied decision-making scenarios. These use cases illustrate how predicted chemistry values can support recruitment, squad selection, and career planning.
7.1. Which center back should Manchester City acquire?
The first use case addresses Manchester City’s center-back shortage during the 2019/2020 season following Vincent Kompany’s departure and Aymeric Laporte’s injury. With Fernandinho deployed out of position, several elite defenders were linked to the club, including Toby Alderweireld, Kalidou Koulibaly, and Milan Škriniar.
Using the predictive chemistry models, the authors estimate both offensive and defensive chemistry between each target and key Manchester City players. From a defensive perspective, Koulibaly is shown to “gel best with center back Stones as well as defensive midfielders Fernandinho and Rodrigo.”
From an offensive perspective, however, Alderweireld exhibits stronger chemistry with City’s attacking players, partly due to his long-passing ability and prior experience with Kevin De Bruyne at international level.
The authors conclude that Koulibaly would be the preferred defensive option, while Alderweireld would be more attractive offensively, illustrating how chemistry-based evaluation can reveal trade-offs that are invisible in purely individual assessments.
7.2. Which right winger should Real Madrid field?
The second use case examines Real Madrid’s difficulties on the right wing following Cristiano Ronaldo’s departure. The authors compare Gareth Bale, Rodrygo, and Vinícius Júnior in terms of predicted offensive chemistry with the rest of the squad.
At the start of the 2019/2020 season, Bale shows the highest average predicted chemistry. However, when the number of matches played together is artificially increased, the models predict that Rodrygo and Vinícius Júnior’s chemistry will improve much faster. Vinícius is expected to overtake Bale after approximately 20 additional matches, with Rodrygo following later.
Further pairwise analysis shows that Bale maintains strong chemistry with most senior players, but Eden Hazard is predicted to “gel better with both Rodrygo and Vinícius Júnior.” The authors therefore recommend giving the younger players sufficient playing time to accelerate chemistry development, particularly in light of Hazard’s central role in the team’s future.
7.3. Which club should Hakim Ziyech move to?
The final use case adopts the perspective of the player rather than the club. For Hakim Ziyech, the authors first identify five plausible destination clubs based on predicted playing time, skill development, and financial feasibility: Internazionale, Roma, Chelsea, Bayern Munich, and Arsenal.
In a second step, they predict offensive chemistry for all player pairs within each club and use the Team Builder to assemble maximum-chemistry line-ups that are forced to include Ziyech. Comparing Ziyech’s average chemistry across these optimal teams, Bayern Munich emerges as the best destination, offering the highest predicted offensive chemistry with surrounding teammates.
This example demonstrates how chemistry models can support player-centric career decisions in addition to club-driven recruitment.
8. Conclusions
The paper concludes by summarizing its central contribution: taking “a first step towards providing insight into how well a team of soccer players gels together.” By introducing metrics that capture mutual offensive and defensive chemistry, the authors move beyond individual performance evaluation toward interaction-aware analysis.
They emphasize three main achievements. First, the definition of JOI and JDI provides a principled way to quantify joint impact using widely available event data. Second, the machine-learned models extend these insights to unseen player pairs, enabling predictive applications in recruitment and squad planning. Third, the Team Builder demonstrates how chemistry estimates can be operationalized in lineup construction and transfer evaluation.
Looking ahead, the authors outline two key directions for future work. They aim to extend chemistry analysis from pairs of players to larger groups, and to incorporate spatio-temporal tracking data in order to obtain “more accurate estimates” of defensive chemistry. These extensions are positioned as natural next steps toward richer, interaction-aware models of team performance.
Learn More
Here you have a direct link to get the free xG Mini-Guide, suited for those who want to fully understand the model in 5 minutes.
References
Bransen, L., & Van Haaren, J. (2020). Player chemistry: Striving for a perfectly balanced soccer team. arXiv preprint arXiv:2003.01712. https://arxiv.org/abs/2003.01712