A machine learning and explainability-driven

methodology for identifying winning strategies in Rugby

Union
Arnaud Odet, Thomas Bechard, Pierre Moretto, Sebastien Dejean, Cristian
Pasquaretta

To cite this version:

Arnaud Odet, Thomas Bechard, Pierre Moretto, Sebastien Dejean, Cristian Pasquaretta. A machine
learning and explainability-driven methodology for identifying winning strategies in Rugby Union.
Decision Analytics Journal, 2025, 15, pp.100568. �10.1016/[Link].2025.100568�. �hal-05021650�

HAL Id: hal-05021650

[Link]
Submitted on 4 Apr 2025

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.

Distributed under a Creative Commons CC BY-NC-ND 4.0 - Attribution - Non-commercial use - No

Derivative Works - International License
 Decision Analytics Journal 15 (2025) 100568

Contents lists available at ScienceDirect

Decision Analytics Journal

journal homepage: [Link]/locate/dajour

A machine learning and explainability-driven methodology for identifying

winning strategies in Rugby Union
Arnaud Odet a,b , Thomas Bechard b , Pierre Moretto b , Sebastien Dejean a , Cristian Pasquaretta b ,∗
a Institut de Mathématiques de Toulouse, France
b
Centre de Recherches sur la Cognition Animale, Centre de Biologie Intégrative, France

ARTICLE INFO ABSTRACT

Keywords: Interest in predicting sports match outcomes has grown significantly, driven by advancements in machine
Machine learning learning techniques and widespread adoption. However, the utilization of these predictive models in enhancing
Explainable artificial intelligence tactical team performance remains relatively limited. We propose a methodology that combines machine
Performance analytics
learning and algorithm explainability techniques, which were demonstrated through a case study on Rugby
Match prediction
Union. Our study unfolds in two phases: first, we identify the most suitable modeling approach for our data
Team evaluation
by establishing a prediction model based on performance indicators observed during games. Subsequently,
we applied an analysis based on SHapley Additive exPlanations (SHAP) values to interpret the predictions of
this model. Our findings serve three primary purposes: (i) from a global standpoint, identifying performance
indicators that primarily determine match outcomes; (ii) from an aggregated point of view highlighting
strengths and weaknesses of any given team; and (iii) from a local perspective, offering technical staff
diagnostic analyses of past games.

1. Introduction pre-game reports and statistics to fit an ANN to predict ice-hockey

games, Prasetio and Harlili [6] used a logistic regression to predict
Sports games are inherently uncertain events, with outcomes shaped soccer games, and O’Donoghue et al. [7] used a linear regression to
by a multitude of factors that interact in complex ways. These fac- predict the 2015 Rugby World Cup games results, with the observation
tors include, but are not limited to, team composition, player per- that prediction is better when the data violates the assumptions of the
formance, weather conditions, refereeing decisions, tactical strategies, linear regression. However, they are rarely compared. Researchers tend
injuries, and even psychological states of athletes. Such elements can to choose one of these algorithms and compare the prediction of their
dynamically shift during the course of a game, contributing to the model with the predictions of one or several baseline models.
unpredictability of the final result. Due to this complexity, prediction Model predictions are mostly addressed as a binary classification
of the results of sports games has attracted increasing interest from problem. As noted in [8] for soccer, in some cases researchers forecast
researchers [1,2]. With the recent popularization of machine learning scores instead of outcomes. The author found this approach less perti-
and the increasing access to computing power, an established body of nent when trying to predict the outcomes of the game. The predictors
literature is dedicated to modeling and predicting sport game outcomes. used to address this classification problem can be of two types: derived
Although the effectiveness of machine learning (ML) techniques to
from past game results (for example, the number of points scored, rank-
predict game outcomes is now widely recognized [1], the scientific
ing before the game, etc.), see [9,10]) or performance indicators within
literature remains divided on the choice of the most suitable predictive
games (for example, number of passes, number of shots, etc.), see [11]).
models [2]. The methods employed may include black-box models
The fundamental difference between these two types of predictor is
(neural networks, support vector machines, ensemble methods, etc.)
that the results of previous matches provide no information about a
or white-box models (linear methods, decision trees, etc.). Loeffelholz
team’s strategy, unlike performance indicators which can quantify the
et al. [3] used an Artificial Neural Network (ANN) to predict National
relevance of tactical and technical choices.
Basketball Association (NBA) games, Delen et al. [4] found a Decision
Moreover, despite significant advances in forecasting the results
Tree outperforms ANN and Support Vector Machines (SVM) predicting
American Football College games, Weissbock and Inkpen [5] combined of sports games, the use of such analyses to guide technical staff

∗ Correspondence to: CRCA-CBI, Rue Marianne Grunberg-Manago, 31062 Toulouse, France.

E-mail addresses: [Link]@[Link] (A. Odet), [Link]@[Link] (T. Bechard), [Link]@[Link] (P. Moretto),
[Link]@[Link] (S. Dejean), [Link]@[Link] (C. Pasquaretta).

[Link]
Received 5 December 2024; Received in revised form 20 March 2025; Accepted 25 March 2025
Available online 27 March 2025
2772-6622/© 2025 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC license
([Link]
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

decisions remains marginal. Here, we aim to leverage recent advances This illustrates how SHapley Additive exPlanation (SHAP) values can
in sport game outcomes forecast to provide technical staff with valuable elucidate the contribution of specific features to model predictions,
information for tactical use. We suggest doing so by modeling the game offering actionable insights for coaches and analysts. Lalwani et al. [18]
results from performance indicators collected during the game applying explored xAI methods to predict outcomes in the Brazilian volleyball
an algorithm explainability framework to assess the importance of league. They used directly interpretable models like logistic regression
features over the predicted result. and boolean rule column generation, as well as post-hoc interpretation
In this regard, we first performed an empirical study comparing methods like SHAP and ProtoDash (a method for selecting prototypical
existing machine learning algorithms for the problem of predicting examples that capture the distribution of dataset), to provide both
game outcomes. Second, we propose a model-agnostic explainability global and local interpretability. Their findings demonstrated the effec-
approach to the prediction of the model to produce a tactical evalu- tiveness of these explanations in understanding the models’ predictions
and ensuring the explanations are sensible. Similarly, Odong and Bou-
ation at multiple levels. At the ‘‘Global’’ (sport) level : our aim is to
quet [19] used SHAP values to explain winter sports performance
identify the key features that determine the outcome of games. At the
prediction, introducing a multi-scale analysis to analyze both local
‘‘Aggregate’’ (team) level: we apply this analysis to identify strengths
(individual prediction) and global (feature importance across the entire
and weaknesses of a given team. Finally, at the ‘‘Local’’ (game) level,
dataset). Cavus and Biecek [20] introduced explainable expected goal
we propose a diagnosis of the game, highlighting features that were
models for performance analysis in football analytics, using aggregated
decisive in determining a specific game’s outcome. This model-agnostic profiles based on ceteris-paribus profiles to evaluate team and player
approach is here applied to Rugby Union. Rugby Union is a full-contact performance. By explaining black-box models at both local and global
team sport that originated in the United Kingdom, characterized by its levels, this approach provides a more comprehensive understanding
physicality, strategy, and camaraderie. It is played between two teams of the factors influencing expected goals values and offensive perfor-
of 15 players each, with the objective of scoring points by carrying or mance. Finally, Procopiou and Piki [21] discussed the role of xAI in
kicking the ball into the opponent’s goal area. With its rich history, football, addressing its conceptualization, applications, challenges, and
global reach, and intense competitions, Rugby Union has become one future directions.
of the most popular and beloved sports in the world. These studies collectively highlight the importance of combining
predictive accuracy with interpretability in sports analytics, enabling
2. Related work stakeholders to make informed decisions based on a clear understand-
ing of the factors driving model predictions.
The application of machine learning to predict sports results has In this paper, we bring novelty by (i) performing an extensive
become a prominent area of research, with various studies aiming to empirical study of existing machine learning models, (ii) proposing a
improve prediction accuracy and provide actionable insights. model-agnostic, 3-levels explainability approach and (iii) applying our
Wong et al. [12] developed machine learning models for English research to Rugby Union.
Premier League soccer matches, incorporating novel features such as
3. The dataset
momentum and fatigue alongside ensemble techniques to achieve pre-
diction results comparable to those of bookmakers. This demonstrates
3.1. Description
the potential of machine learning to rival expert predictions.
Although some models inherently offer a degree of interpretability We created our dataset using data of 2058 regular season games
through their structure, others require additional techniques to reveal collected from the website [Link] for the 2021/2022 to 2023/2024
their decision-making processes. Calderón-Díaz et al. [13] employed seasons of the following championships: Premiership, Top14, ProD2,
various machine learning algorithms to identify biomarkers of mus- and United Championship.
cle injuries in professional soccer players. They found that hamstring Our dataset comprises 79 features per team, related to possessions,
muscle strength and stiffness were key predictors, achieving up to lineouts, scrums, kicks, rucks, passes, linebreaks, tackles, cards. Tries
78% precision with eXtreme Gradient Boosting (XGBoost). The intrinsic and penalties were excluded as they directly determine the result of
explainability of models such as decision trees and logistic regression the game. Most of these features have already been shown to have
allows medical teams to interpret these models and gain trust in their an impact on the game result: Watson et al. [22] emphasized the
results. Gifford and Bayrak [14] also utilized machine learning to importance of ball possession and highlighted the fact that winning
predict outcomes in the National Football League (NFL). Their study teams tend to have fewer passes and rucks, Vaz et al. [23] stressed the
focused on quantifying the influence of team statistics on regular season importance of winning lineouts and Hughes et al. [24] the importance
wins using logistic regression, identifying offensive turnovers as having of ‘‘stealing’’ lineouts introduced by the opposing team, while Ortega
a significant negative impact and defensive turnovers as having a et al. [25] studied the importance of linebreaks and percentage of
strong positive impact on game outcomes. In these papers, the models’ tackles completed. Moreover, Coughlan et al. [26] found that actions
coefficients provide information and allow for a direct interpretation of following lineouts, scrums, and kicks receipts tend to precede tries.
the influence of different factors, thus enhancing the reliability of the However, a joint analysis of the relative importance of these features
is missing to the best of our knowledge.
results.
Possessions and linebreaks are divided according to the position
To address the limitations of ‘‘black-box’’ models, researchers have
in the field. The field (of approximately 100 m in length) is divided
increasingly turned to explainability techniques that provide insights
into 4 zones: when the team is in its 22 m, between its 22 meters
into model behavior. Explainable Artificial Intelligence (xAI) is a flour-
and the median line, between the median line and the opponent’s
ishing topic, with application in various fields such as healthcare,
22 m, and in the opponent’s 22 m. Scrums, lineouts, and tackles are
manufacturing, transportation, and finance, see [15]. In sports, recent divided according to quality indicator : they are manually labeled in 3
papers implemented xAI methods to highlight important features. Ye- subgroups (positive, neutral, negative) by aiasports’ analysts. It should
ung et al. [16] proposed a framework for the prediction of interpretable be noted that there is no further quality indication. As a consequence,
match results in football, emphasizing the importance of providing the dataset does not distinguish between good and bad passes and
coaches and players with actionable information. They rely on feature between kicks that would make the team considerably advance in
importance of a XGBoost model to determine which factor are the most the field and kicks that would not. The duration of possessions are
important. Plakias et al. [17] used SHAP values to identify key factors expressed in seconds and are classified into categories: from 0 to 30 s,
influencing team standings in French Ligue 1. Their analysis revealed from 30 to 60 s, and more than 60 s. The duration of rucks is also
that short passes and through balls positively impact team success, expressed in seconds and classified into categories: between 0 and 3 s,
while long balls and attempted tackles negatively affect performance. between 3 and 6 s, and more than 6 s.

2
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

Table 1 This analysis is performed once and is not changed in the further
Models accuracy and F1-scores (best in bold). development of this work.
Model Accuracy F1-score Cohen-Kappa As part of the feature selection process, we conducted a Variance In-
Support Vector Classifier 0.854 0.903 0.608 flation Factor analysis to mitigate the multicollinearity issue present in
Logistic Regression 0.851 0.901 0.601 the dataset. Indeed, some feature are by definition correlated (e.g. the
PLS-DA 0.835 0.890 0.566
total number of scrums is the sum of positive, neutral and negative
ANN 0.832 0.886 0.568
Ada Boost 0.823 0.885 0.508
scrums), and some others are expected to be correlated (e.g. number
XGB Classifier 0.817 0.881 0.491 of possessions and possession time).
Random Forest 0.771 0.862 0.246 We finally retained 40 features per team, those features have al-
k - Nearest Neighbors Classifier 0.768 0.858 0.272 ready been studied as having an impact on game results, as mentioned
Baseline ELO 0.756 0.848 0.260 in 3.1, plus the ELO difference, summing up to 81 features per game,
Baseline home 0.726 0.841 0.000
and we validated with Rugby Union professional coaches the relevance
of our feature selection.

3.2. Feature engineering 4. Methodology

Machine learning algorithms. We used different types of models among

In order to account for the intrinsic level difference between two
the most commonly used (including in works related to sports, see [2]):
teams, we created a variable ‘‘ELO points’’ [27]. An ELO rating system
is a rating system in which each participant (here, each team) has its • Linear models : logistic regression and partial least squares —
score and after each confrontation between two teams, the winner’s discriminant analysis (PLS-DA),
ELO score increases by 𝑚 points, and the loser’s ELO score decreases • Ensemble methods : random forest, adaptive boost classifier, eX-
by the same 𝑚 points, where 𝑚 is proportional to the difference in treme Gradient Boosting (XGBoost) classifier,
ELO score prior to the game. High values of 𝑚 denote unexpected • Support Vector Machine with linear, polynomial and Radial Basis
results, whereas low values of 𝑚 correspond to foreseeable results. We Function (RBF) kernel,
initialized the score of all teams at 0 at the start of the 2021/2022 • k-Nearest Neighbors (kNN) classifier,
season and then used the games of the first half of the 2021/2022 • Artificial Neural Networks.
season for each team to have an initial ELO score representative of its
level. Those games were put aside and not used in the rest of the study. We also used two baseline models to evaluate the accuracy gain
The ELO scores kept evolving after each game, so the ELO score of a brought by our analysis: the Baseline Home model, which systemati-
given team before a game reflects its performance considering all games cally predicts a victory for the home team (69.7% of the matches in
previous to the considered game. our dataset), and the Baseline ELO model, which consists of a logistic
We set the fixed factor 𝐾 set at 20, applying the values used by regression with the only predictor being the difference in ELO points
the International Chess Federation (FIDE, [Link] between the two teams before the start of the game.
chapter/B022024). This factor determines the magnitude of the number Machine learning algorithms were implemented using the scikit-
of points that teams exchange: the higher 𝐾, the more points are learn [31] Python library, and TensorFlow [32] was used for ANN
exchanged, and the less stable the ranking is. According to Lampis optimization. To optimize the performance of these algorithms, a ran-
et al. [9], the determination of this variable 𝐾 can be the subject of dom search was performed in a predefined range. The predefined
ranges for the hyperparameters optimization are standard ranges in
further research. The ELO rating system has been used as a relevant
machine learning. The code is available at [Link]
predictor of game outcomes in literature [28,29]. In this work, we
odet/explain_winning_strategies.
tested different values for 𝐾, from 10 to 40, and found it made no
As specified in 3.2, we set aside the first half of the 2021/2022
significative difference (less than 1% in accuracy).
season for the initialization of ELO. We used games from the second
Finally, we defined a ELO difference variable as the difference of
half of 2021/2022 season as well as games from the 2022/2023 season
the ELO score of the two teams : 𝐸𝐿𝑂𝑑𝑖𝑓 𝑓 𝑒𝑟𝑒𝑛𝑐𝑒 = 𝐸𝐿𝑂ℎ𝑜𝑚𝑒 − 𝐸𝐿𝑂𝑎𝑤𝑎𝑦 .
as a training set (𝑛 = 1042), games from the first half of the 2023/2024
This variable takes a positive value when the home team is expected season as validation set for the search for hyperparameters (𝑛 = 328)
to be stronger than the away team and low values when the away and finally games from the second half of the 2023/2024 season as test
team is expected to be stronger than the home team. The higher set for evaluation (𝑛 = 328).
the absolute value of 𝐸𝐿𝑂𝑑𝑖𝑓 𝑓 𝑒𝑟𝑒𝑛𝑐𝑒 , the greater the expected level
difference between the two teams. Comparison metrics. We chose to compare the best performing models
Furthermore, we classified the games according to whether they in terms of accuracy. We also displayed their respective F1-score to
were won by the home team (1) or not (0). This binary feature will account for their sensitivity to false positives and false negatives and
be our target variable for the remaining of this paper. One limitation their respective Cohen’s Kappa score, as our dataset may be considered
of this choice is the consideration of 56 drawn matches (2.7%), which as unbalanced (69.7% of games are won by the home team).
are thus classified in the same category as a defeat for the home team.
5. Results and analysis
This approach is justifiable by the existence of a home advantage in
sports [30]. Moreover, considering the draw would lead to a three-
5.1. Model comparison
class problem with imbalance issue, since draws represent a very small
proportion of games, and as highlighted in [2], in soccer results predic- The results of the model comparison are presented in Table 1. Each
tion, the accuracy drops considerably when considering a three-class line corresponds to be the score of the best set of hyperparameters for
problem compared to a two-class problem. the corresponding model. We observe that the Support Vector Classifier
with RBF model achieves the best scores in terms of accuracy, F1-score
3.3. Data preparation and Cohen’s Kappa. The following of the paper is based on this model,
with the best hyperparameters found by randomized search.
The dataset does not contain missing values : all games are provided It should be noted that while the linear models (logistic regression
with all features available. and PLS-DA) do not perform as well as the Support Vector Classi-
Features were scaled using either standard scaler, robust scaler fier with RBF model, they have the advantage of being more easily
or mix-max scaler, based on visual inspection of their distribution. explicable.

3
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

Fig. 1. Illustration of the global-scale explanation obtained on the test dataset from the Support Vector Classifier with RBF model using the train dataset as background. The
𝑥-axis represents the SHAP values for each predictor at each game. Color gradient corresponds to the value of the corresponding feature for each point: high (red) or low (blue).
SHAP values larger than 0 increase probability of the home team’s victory. For readability purposes, only the most impacting features are displayed. Feature impacts are ordered
by 𝜎 ∗ = (𝜎ℎ + 𝜎𝑎 )∕2, where 𝜎ℎ and 𝜎ℎ are the standard deviation of SHAP values for home and away team respectively. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

5.2. Diagnostic analysis feature takes a high value are located to the right of the central axis,
indicating that this feature pushes the prediction towards 1 (victory for
5.2.1. SHAP values the home team). In contrast, field, clearance, and pressure kicks (lines 2
In order to explain a given game and provide technical staffs with to 5 in Fig. 1) appear to increase the probability of victory for the team
the potential reasons that may have led to the final result, we suggest that makes the most. This observation is illustrated by the fact that
analyzing the model predictions using the methodology developed when the predictor is evaluated at the home team (respectively away
by Lundberg and Lee [33], introducing the SHapley Additive exPlana- team) level, the probability of winning for the home team increases
tions (SHAP) values. This methodology allows to measure the impact of (resp. decreases) for large values of the predictor (red dots). One last
different features on the prediction. It is based on Shapley values [34], observation that can be made is about the possession: possession high
developed in game theory to assess the contribution of each player of in the field (in the opponent’s 22-meter area) is positively correlated
a coalition in a cooperative game, defined as follows: with chances of success, and inversely, short possessions (shorter than
∑ |𝑆|!(|𝑁| − |𝑆| − 1)! 30 s) are negatively correlated with the probability of victory.
𝜙𝑖 = [𝑣(𝑆 ∪ {𝑖}) − 𝑣(𝑆)]
𝑆⊆𝑁⧵{𝑖}
|𝑁|! Here, the difference in the ELO points that precede the match seems
where 𝜙𝑖 is the Shapley value associated to player 𝑖, 𝑆 is a sub- to be the variable that most influences the result of the match indicating
coalition of 𝑁 that does not contain player 𝑖 and 𝑣(𝑆) is the value a difference in team performances previous to the game is likely to
obtained by the coalition. be reproduced in the game. High values of ELO difference indicate
The analogy between machine learning and game theory can be a strong difference in ELO score in favor of the home team, which
presented as follows : a feature (predictor) is considered as a player, appears to be positively correlated with the probability of victory for
the analyzed coalition is the set of all features used by a model, and the home team, and low values of ELO difference indicate a strong
finally, the value generated by the coalition is the value predicted by difference in ELO score in favor of the away team, which appears to be
the model. Thus, the objective is to determine which features contribute
negatively correlated with the probability of victory for the home team.
the most to the observed prediction.
This observation is rather intuitive, as a stronger team is expected to
prevail over a weaker team.
5.2.2. Global diagnosis — analysis over the test dataset
In Fig. 2, we (i) grouped the features by type (e.g. Kicks includes
Our approach allows highlighting the features that have the most
influence on the outcome of a game in our test set, as illustrated in Fig. various types of kicks, Scrums include positive, neutral and negative
1. The features with the lowest concentration of points near the central scrums, etc.) and (ii) separate them according to their impact when
axis are the feature that most influence the model’s predictions. Indeed, playing at home or away. It appears that kicks and tackles have a sim-
the lower |𝜙𝑖 | for a given prediction, the less important is feature 𝑖 for ilar importance when playing at home and away, while rucks features
this prediction. from the home team have a higher weight on the game outcome than
Moreover, considering the signs of the SHAP values and the feature those of the away team. On the other hand, the passes and scrums
value, we can infer if it is desirable to have high values for the feature performed by the away team weigh more on the outcome of the game
of interest. For example, the number of simple passes by the away team than those of the home team. Rucks are related to contacts in possession
(6th line of Fig. 1) appears to be a factor that is negatively correlated phases, when a team carries the ball and tries to push the defending
with the probability of victory for that same team as it increases the team towards its try line. This result is not surprising, as players at
probability of victory for the home team. Indeed, the points where this home tend to gain more meters, see [35].

4
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

Fig. 2. Feature importance when playing at home or away. The 𝑥-axis (respectively y-axis) represents the standard deviation of features while playing at home (resp. away).
The sizes indicates the number of features grouped under the corresponding label, and the dotted red line represents the 𝑦 = 𝑥 line to separate feature having more impact when
playing at home vs when playing away. Features close to origin are less decisive, whereas features located on the right (respectively on the top) of the graph are determining
feature when playing at home (resp. away). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 3. Example of a aggregate diagnostic graph for a given team. The graph is made of boxplots of SHAP values over the analyzed period. The 𝑥-axis represents the impact on
probability of winning predicted by the Support Vector Classifier with RBF. On the 𝑦-axis the corresponding features for the given team (i.e. Team A) and its opponents that play
a role on Team probability of winning. The red diamonds represent the mean of SHAP values over the season, and the red dashed line emphasizes the 0-impact line. Only the
most representative features of Team T’s games are represented (ie features with |𝑚𝑢𝑖 | > 0.01 where 𝜇𝑖 the average SHAP value for feature 𝑖).

5.2.3. Aggregate diagnosis — analysis of a given team model, the (low) number of kicks of Team T’s opponents (i.e. opponent
By analyzing SHAP values of features for a given team through a nb field kicks and opponent nb pressure kicks in Fig. 3) is the feature
certain number of games, we can identify the characteristics that led that most increased the probability of victory of the team during the
the team to outperform or underperform. An illustration can be found season, traducing an ability to prevent the opponent from developing a
in Fig. 3. It presents the aggregate analysis over 26 games of a Top kicking game. Team 𝑇 also seems to be a team performing well carrying
14 team (hereafter team T) in the 2023/24 season. According to our the ball as its number of offload passes, its number of breakthrough in

5
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

Fig. 4. Average SHAP values of the top 6 teams in ELO at the end of the 2023/24 season. Blue axes correspond to the analyzed team and orange axes to their opponents.
The red polygon materializes the 0-impact line. The scope of the analysis is the 2023/24 season. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

the opponent’s 22 m, and its number of beaten defenders correspond teams manage to influence the way their opponents play, preventing
to positive SHAP values. them from kicking and forcing them to carry the ball. Most of them
Similarly, we can observe that Team T’s kicking game decreases (respectively 5,4, and 5) have positive values for rucks, scrums, and
their winning probability, as they correspond to low SHAP values, tackles respectively, highlighting their ability to dominate physically.
despite being a winning strategy at the global level (see paragraph They also all show positive breakthrough SHAP values, demonstrating
5.2.2). an ability to carry the ball. This will to carry the ball is reflected in
In summary, during the period analyzed, the main strengths of their passes SHAP values, which are negative for 5 of them, and with
Team 𝑇 are its ability to carry the ball and prevent the opponent from very low values (average SHAP below -2%) for 3 of them. Considering
developing a kicking game, and its weakness is mainly its own kicking this finding in light of the results of Section 5.2.2, we can deduce that
game. passing the ball is globally a losing strategy, but the performing team
The aggregate level allows us to emphasize the characteristics of manages to win while passing the ball.
wining teams. In Fig. 4, we considered the top-6 teams in ELO at the
end of our analysis and represented their profiles in terms of average 5.2.4. Local diagnosis — analysis of a given game
SHAP values over the 2023/2024 season. We can highlight several Applied at the local level, this method shows which features have
similarities: they all have positive average SHAP values in terms of led to the elaboration of the prediction (see Fig. 5, result of this analysis
passes performed by their opponent, and 4 of them also display pos- for a game of Top 14 of the 2023/24 season). In this example, the model
itive values for the kicks performed by their opponent, meaning these predicts a victory for the home team (with a probability of 92.5%)

6
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

Fig. 5. Illustration of the explanation of a local prediction for a match in the Top 14 league during the 2023/24 season. The 𝑥-axis represents the probability that the home
team wins the match according to the model (the base probability calculated across the training set being 0.634), and the values shown on the 𝑦-axis are the raw values of the
corresponding features. Values written in arrows represent the estimated impact of each predictor in percentage on the probability of the home team’s victory. A predictor value
that pushes model predictions towards 1 (home team victory) is indicated in red otherwise in blue. (For interpretation of the references to color in this figure legend, the reader
is referred to the web version of this article.)

mainly due to the following features: the number of clearance kicks, may enhance their ability to advance the ball more effectively. This
field kicks and the possession in their own 22-meters area by the away finding is consistent with previous research [36], which pointed out
team contributed positively to the prediction of home team victory by that executing too many consecutive short passes is not an efficient
respectively +6%, +4%, and +5%, while the share of rucks of less than method for ball progression. More generally, understanding the optimal
3 s in the opponent’s 22-meters area performed by the home team and balance between passing and kicking strategies could assist teams in
the number of breakthrough by the away team between the median line optimizing their playing style. This approach would ensure that they
and the home team’s 22-meters area both decreased this probability by capitalize on the actions that are statistically more likely to contribute
-4%. This information can be valuable in the post-game analysis as it to their success, ultimately improving overall performance.
emphasizes the characteristics in which the decision was made. The aggregate-level analysis for a given team provides valuable
insights into its strengths and weaknesses, which can be used to adapt
6. Discussion one’s own game plan accordingly. For example, as shown in Fig. 3,
when facing the team under analysis, we could benefit from adjusting
In this study, we propose an explainability approach applied to our strategy by (i) preventing the opposing team from carrying the ball
machine learning models to predict the impact of technical and tactical (a weakness we can exploit) and (ii) creating more opportunities to kick
the ball, since the team under analysis typically limits their opponent’s
features collected during rugby games on the probability of winning.
kick opportunities.
Our approach achieves high accuracy, making it a reliable model for
This approach can be especially useful for video analysts, guiding
providing technical staff with (i) a global analysis of the most relevant
them to focus on specific situations to analyze, such as the key combi-
performance indicators, (ii) a team-oriented summary of strengths and
nations leading to efficient ball carrying, and thus better understanding
weaknesses, and (iii) a game-level diagnosis based on past matches.
the opposing team’s vulnerabilities.
This 3-level analysis differs from the 2-level analysis in the existing
This methodology provides teams with actionable insights that help
literature [18–20].
them adapt by targeting the weaknesses of their opponents while
The model selection approach we implemented in this paper suggest
minimizing the impact of their strengths, leading to more strategic and
that the Support Vector Classifier with RBF displays an accuracy of informed decisions.
85.4% in the test set and reveals good predictors of rugby performance. Finally, the local diagnosis of a given game (see Fig. 5) can help
Our global level analysis indicates that the probability of winning a identify which aspect of the game to put emphasis on during a post-
game increases with (i) a higher number of kicks and (ii) a lower game debriefing. For instance, the away team, which lost the analyzed
number of simple passes. These findings provide important informa- game, could draw the conclusion they did not kick the ball enough, and
tion on match preparation strategies, highlighting the importance of instead spent too much time in their 22-meters area. They could also
specific game actions in influencing outcomes. First, given that kick- see that they allowed their opponent too many successful rucks.
ing appears to positively impact the likelihood of winning, tactical Our approach could be improved with a higher volume of data. Sec-
staff could consider prioritizing kicking drills during training sessions. ondly, there is also room for improvement in the features being used:
This would help improve team performance in this critical area. It the lack of indicators of quality leads to considering all instances of a
has already been noted that kicks tend to precede tries, see [26]. given feature equally. More specifically, while our analysis highlights
Secondly, the observation that an excessive number of simple passes the importance of kicks, considering, for instance the number of meters
is negatively correlated with the probability of victory suggests that gained after a kick should improve the analysis.
coaches might focus on training sessions that emphasize fast, single Additionally, a difficulty lies in simulating the absence of a feature.
passes. By reducing the number of consecutive short passes, teams To quantify the importance 𝜙𝑖 , of feature 𝑖 over 𝑓 (𝑥) the predicted

7
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

Fig. 6. Illustration of the second order explanation of a local prediction for a match in the Top 14 league during the 2023/24 season. For readability purpose, only the most
impacting features highlighted in our global analysis are represented. Dots close to the feature name represent the SHAP values of the corresponding feature for the game considered,
and edges between two features represent the value of the Shapley interaction between the two linked features. Red (resp. blue) indicates values positively (resp. negatively)
correlated to the probability of a home win, and the size is proportional to the absolute value of the Shapley value (for dots) or Shapley interaction (for lines). (For interpretation
of the references to color in this figure legend, the reader is referred to the web version of this article.)

probability of winning for the home team in a given game, we use passes together with a high number of clearance kicks yielded a higher
SHAP values (presented in 5.2.1), which can be seen as a function of the increase of the probability of a home team victory than the sum of
difference between 𝑓 (𝑥), the prediction of the model with feature and Shapley values.
𝑓 (𝑥0 ) where 𝑥0 = {𝑥𝑗 ∖𝑗 ≠ 𝑖}. To compute 𝑥0 , 𝑥𝑖 is replaced by a value Furthermore, to complement the present work, it would be benefi-
extracted from a background (a subset of the dataset), and the choice of cial to consider a certain dependence between features and to model
this background influences the SHAP values (see [37]). In this work, we modal shift. Indeed, in the example developed in Fig. 5, if the away
used the whole training set as background, which would be equivalent team decided to use more clearance kicks, the team would necessarily
to comparing a given game to the ‘‘average game’’, but other alterna- have fewer passes and/or rucks and/or line breaks.
tives can be discussed : for instance, the background could be composed
by games played by teams with ELO score similar to an analyzed team, 7. Conclusion
or by games sorted by similarity to an analyzed game to better stress
the decision boundary. This approach could be more ‘‘pragmatic’’ as While our work does not claim to replace the experience and
it would compare teams with similar characteristics and provide more expertise of a technical staff, we believe that it can provide practitioners
‘‘achievable’’ recommendations. We leave this exploration for further with valuable insights about their team and their opponents. The main
research. advantage of using a machine learning approach is that it can analyze
Related to SHAP values, an additional limitation remains to be large datasets of many games that would be too complicated and costly
discussed : in this work, for simplification purposes, we did not consider for a technical staff to process. Applying machine learning algorithms
Shapley interactions (see [38,39]). These interactions extend Shapley combined with explainability techniques on statistics observed within
values to joint contribution of features. A straightforward illustration games, we can identify winning strategies. Moreover, we believe that
can be features such as latitude and longitude, from which only joint this approach, presented and illustrated here through a case study on
consideration can provide information about exact location. Fig. 6 Rugby Union, is transferable to other sports.
illustrates the second-order explanation of the game presented in Fig.
5. The second order relates to the importance of a pair of features. A
Declaration of competing interest
positive Shapley interaction value indicates that the joint value of the
feature is greater than the sum of the first-order (Shapley values) of
The authors declare that they have no known competing finan-
the corresponding features, while a negative SI indicates the opposite.
cial interests or personal relationships that could have appeared to
We can see in Fig. 6 that for the game considered in our local analysis,
influence the work reported in this paper.
the joint contribution of the ELO difference and the number of field
kicks by the home team, meaning the joint contribution of these two
Acknowledgments
features is lower than the sum of the Shapley values of these feature.
This indicates the features ‘‘cooperated badly’’ and their interaction
decreased the probability of a home team victory. As both features This research did not receive any specific grant from funding agen-
have a positive Shapley values we can deduce the observed number cies in the public, commercial, or not-for-profit sectors. We thank AIA
of field kicks performed by the home team is a desirable value (as it Sports for granting the data mentioned above.
has a positive Shapley values but given the ELO difference, it could
have been better. On the other hand, we observe a positive Shapley Data availability
interactions value between the number of simple passes and the number
of clearance kicks performed by the home team, meaning that their The authors shared the code used in the analysis, however, they do
joint contribution increased the probability of a home win. In the not have permission to share the data.
game analyzed, the fact that the home team performed few simple

8
A. Odet, T. Bechard, P. Moretto et al. Decision Analytics Journal 15 (2025) 100568

References [21] Andria Procopiou, Andriani Piki, The 12th player: Explainable artificial intelli-
gence (XAI) in football: Conceptualisation, applications, challenges and future
[1] T. Horvat, J. Job, The use of machine learning in sport outcome prediction: directions, in: Int. Congr. on Sport Sci. Res. and Technol. Support, 2023, http:
A review, WIREs Data Min. Knowl. Discov. 10 (5) (2020) [Link] //[Link]/10.5220/0012233800003587.
1002/widm.1380. [22] Neil Watson, Ian Durbach, Sharief Hendricks, Theodor Stewart, On the validity
[2] R. Bunker, T. Susnjak, The application of machine learning techniques for of team performance indicators in rugby union, Int. J. Perform. Anal. Sport. 17
predicting match results in team sport: A review, J. Artif. Intell. Res. 73 (2022) (4) (2017) 609–621, [Link]
1285–1322, [Link] [23] Luis Vaz, Michele Van Rooyen, Jaime Sampaio, Rugby game-related statistics
[3] Bernard Jacob Loeffelholz, Earl M. Bednar, Kenneth W. Bauer, Predicting NBA that discriminate between winning and losing teams in IRB and super twelve
games using neural networks, J. Quant. Anal. Sport. 5 (2009) [Link] close games, J. Sport. Sci. Med. 9 (1) (2010) 51.
10.2202/1559-0410.1156. [24] Angus Hughes, Andrew Barnes, Sarah M. Churchill, Joseph Antony Stone,
[4] Dursun Delen, Douglas Cogdell, Nihat Kasap, A comparative analysis of data
Performance indicators that discriminate winning and losing in elite men’s
mining methods in predicting NCAA bowl outcomes, Int. J. Forecast. 28 (2)
and women’s Rugby Union, Int. J. Perform. Anal. Sport. 17 (2017) 534–544,
(2012) 543–552, [Link]
[Link]
[5] Josh Weissbock, Diana Inkpen, Combining textual pre-game reports and statisti-
[25] Enrique Sodupe Ortega, Diego Villarejo, Josè M. Palao, Differences in game
cal data for predicting success in the national hockey league, in: Adv. in Artif.
Intell., Springer International Publishing, 2014, pp. 251–262, [Link] statistics between winning and losing rugby teams in the six nations tournament,
10.1007/978-3-319-06483-3_22. J. Sport. Sci. Med. 8 4 (2009) 523–527.
[6] Darwin Prasetio, Dra. Harlili, Predicting football match results with logistic [26] Molly Coughlan, Charles Mountifield, Stirling Sharpe, Jocelyn K. Mara, How they
regression, in: 2016 Int. Conf. On Adv. Inform.: Concepts, Theory And Appl., scored the tries: applying cluster analysis to identify playing patterns that lead
ICAICTA, 2016, pp. 1–5, [Link] to tries in super rugby, Int. J. Perform. Anal. Sport. 19 (3) (2019) 435–451,
[7] P. O’Donoghue, D. Ball, J. Eustace, B. McFarlan, M. Nisotaki, Predictive models [Link]
of the 2015 rugby world cup: accuracy and application, Int. J. Comput. Sci. [27] A. Elo, The rating of chessplayers, past and present, 1978.
Sport. 15 (1) (2016) 37–58, [Link] [28] Lars Magnus Hvattum, Halvard Arntzen, Using ELO ratings for match result
[8] John Goddard, Regression models for forecasting goals and match results in prediction in association football, Int. J. Forecast. 26 (3) (2010) 460–470,
association football, Int. J. Forecast. 21 (2005) 331–340, [Link] [Link]
1016/[Link].2004.08.002. [29] Anthony Costa Constantinou, Norman Elliott Fenton, Determining the level of
[9] Tzai Lampis, Ntzoufras Ioannis, Vassalos Vasilios, Dimitriou Stavrianna, Predic- ability of football teams by dynamic ratings based on the relative discrepancies
tions of European basketball match results with machine learning algorithms, J. in scores between adversaries, J. Quant. Anal. Sport. 9 (1) (2013) 37–50,
Sport. Anal. 9 (2) (2023) 171–190, [Link]
[Link]
[10] Maxime Settembre, Martin Buchheit, Karim Hader, Ray Hamill, Adrien Tarascon,
[30] Miguel-Ángel Gómez, Richard Pollard, Juan-Carlos Luis-Pascual, Comparison of
Raymond Verheijen, Derek McHugh, Factors associated with match outcomes in
the home advantage in nine different professional team sports in Spain, Percept.
elite European football – insights from machine learning models, J. Sport. Anal.
Mot. Skills 113 (2011) 150–156, [Link]
10 (1) (2024) 1–16, [Link]
[11] Dragan Miljkovic, Ljubisa Gajic, Aleksandar Kovacevic, Zora Konjovic, The use 156.
of data mining for basketball matches outcomes prediction, in: IEEE 8th Int. [31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.
Symp. on Intell. Syst. and Inform., IEEE, 2010, pp. 309–312, [Link] Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D.
10.1109/SISY.2010.5647440. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: Machine learning
[12] Albert Wong, Eugene Li, Huan Le, Gurbir Bhangu, Suveer Bhatia, A predictive in python, J. Mach. Learn. Res. 12 (2011) 2825–2830.
analytics framework for forecasting soccer match outcomes using machine [32] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen,
learning models, Decis. Anal. J. 14 (2025) 100537, [Link] Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin,
[Link].2024.100537. Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael
[13] Mailyn Calderón-Díaz, Rony Silvestre Aguirre, Juan P. Vásconez, Roberto Yáñez, Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh
Matías Roby, Marvin Querales, Rodrigo Salas, Explainable machine learning Levenberg, Dandelion Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris
techniques to predict muscle injuries in professional soccer players through Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Tal-
biomechanical analysis, Sens. 24 (1) (2023) 119, [Link] war, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol
s24010119. Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, Xiaoqiang
[14] Matt Gifford, Tuncay Bayrak, A predictive analytics model for forecasting
Zheng, Tensorflow: Large-scale machine learning on heterogeneous systems,
outcomes in the National Football League games using decision tree and logistic
2015, Software available from [Link], [Link]
regression, Decis. Anal. J. 8 (2023) 100296, [Link]
[33] S. Lundberg, S. Lee, A unified approach to interpreting model predictions, Adv.
2023.100296.
Neural Inf. Process. Syst. 30 (2017).
[15] Saranya A., Subhashini R., A systematic review of explainable artificial intelli-
gence models and applications: Recent developments and future trends, Decis. [34] L.S. Shapley, 17. A value for n-person games, in: Harold William Kuhn,
Anal. J. 7 (2023) 100230, [Link] Albert William Tucker (Eds.), Contributions To the Theory of Games, Volume
[16] Calvin C.K. Yeung, Rory Bunker, Keisuke Fujii, A framework of interpretable II, Princeton University Press, 1953, pp. 307–318.
match results prediction in football with FIFA ratings and team formation, PLOS [35] Teneale Alyce McGuckin, Wade Heath Sinclair, Rebecca Maree Sealey, Paul
ONE 18 (4) (2023) 1–15, [Link] Bowman, The effects of air travel on performance measures of elite Australian
[17] Spyridon Plakias, Christos Kokkotis, Michalis Mitrotasios, Vasileios Armatas, rugby league players, Eur. J. Sport. Sci. 14 (2014) S116–S122, [Link]
Themistoklis Tsatalas, Giannis Giakas, Identifying key factors for securing a 10.1080/17461391.2011.654270.
champions league position in French ligue 1 using explainable machine learn- [36] Mark G.L. Sayers, Jason Washington-King, Characteristics of effective ball carries
ing techniques, Appl. Sci. 14 (18) (2024) 8375, [Link] in super 12 rugby, Int. J. Perform. Anal. Sport. 5 (2005) 106–192, [Link]
app14188375. [Link]/10.1080/24748668.2005.11868341.
[18] Abhinav Lalwani, Aman Saraiya, Apoorv Singh, Aditya Jain, Tirtharaj Dash, [37] Emanuele Albini, Jason Long, Danial Dervovic, Daniele Magazzeni, Counter-
Machine learning in sports: A case study on using explainable models for factual Shapley additive explanations, in: Proc. of the 2022 ACM Conf. on
predicting outcomes of volleyball matches, in: 2nd Int. Conf. on Sports Eng., Fairness, Account., and Transpar., 2022, pp. 1054–1070, [Link]
ICSE 2021, [Link]
1145/3531146.3533168.
[19] Lawrence A. Odong, Paolo Bouquet, An introduction of explainable artificial
[38] Michel Grabisch, Marc Roubens, An axiomatic approach to the concept of
intelligence to winter sports performance analysis, in: 2023 IEEE Int. Workshop
interaction among players in cooperative games, Int. J. Game Theory 28 (1999)
on Sport, Technol. and Res., STAR, 2023, pp. 94–97, [Link]
547–565, [Link]
STAR58331.2023.10302671.
[39] Maximilian Muschalik, Hubert Baniecki, Fabian Fumagalli, Patrick Kolpaczki,
[20] Mustafa Cavus, Przemysław Biecek, Explainable expected goal models for per-
formance analysis in football analytics, in: 2022 IEEE 9th Int. Conf. on Data Sci. Barbara Hammer, Eyke Hüllermeier, shapiq: Shapley interactions for machine
and Adv. Anal., DSAA, 2022, pp. 1–9, [Link] learning, in: The Thirty-Eight Conf. on Neural Inf. Process. Syst. Datasets and
2022.10032440. Benchmarks Track, 2024.