Which Machine Learning Models Perform Best for Football Match Prediction?
A practical overview of machine learning models, data sources, and evaluation methods for soccer match result forecasting.
The following summary critically reviews the research paper titled "Machine Learning for Soccer Match Result Prediction" by Rory Bunker, Calvin Yeung and Keisuke Fujii. 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
Predicting professional football match outcomes remains inherently complex due to "draws being a common outcome in the sport, as well as its low-scoring nature and often highly competitive leagues". The field attracts diverse stakeholders (bettors, bookmakers, fans, analysts, and coaching staff) each requiring different model attributes. While high accuracy is vital for betting markets, coaches and performance analysts prioritize model interpretability to uncover influential match features ("performance indicators" [63]) that can be targeted for strategic refinement.
Historically, statistical models dominated this domain. Early work by Stefani [105] introduced least-squares-based team ratings, while Maher [81] and Dixon & Coles [37] advanced Poisson-based goal modeling, addressing temporal variance and data irregularities. These evolved into more flexible distributions such as dependent Poisson, negative binomial, and extreme value forms [51, 9, 10, 91]. The Bivariate Poisson model, despite its simplicity, still "provides strong performance", as demonstrated by studies using the engsoccerdata R package [78, 34].
Constantinou [28] proposed a taxonomy classifying prediction models into three groups:
Statistical models, encompassing Bivariate Poisson and related goal-distribution models [81, 37, 51, 9, 10, 91], along with ordered logistic regression [4, 100].
Machine learning and probabilistic graphical models, such as fuzzy logic, genetic algorithms [112, 87, 54, 95], and Bayesian networks [68, 31].
Rating systems, including Elo ratings [41, 65], pi-ratings [30], and more recent innovations like Berrar ratings [8], GAP ratings [118], and PageRank adaptations [15, 59].
However, this taxonomy has proven insufficient given the increasing use of hybrid models. Top entries in the 2017 Soccer Prediction Challenge [7] exemplified this trend, combining machine learning with football-specific rating inputs [8, 59, 28]. Examples include Random Forest models enriched with Poisson-based team rankings [53], or the use of logistic regression which straddles both statistical and ML categories.
A paradigmatic case is the Bradley-Terry model [11], employed in the 2017 Challenge [116] and widely used for team ranking [78]. Its dual identity, as a statistical foundation and the basis for Elo-type ratings [33], illustrates the fuzzy boundaries among the three groups. Consequently, hybrid methodologies have flourished, blending model types and rating features to enhance predictive power and practical insight.
The research promises a comprehensive examination of this landscape, detailing datasets (notably the Open International Database and the 2017 Challenge), model classes (traditional, ensemble, and deep learning), interpretability concerns, feature types (including ratings and team/player data), and evaluation techniques. By doing so, it aims to support both academic inquiry and practical application in the football analytics community.
2 Data
The growing interest in football match result prediction has prompted the development and release of various datasets, though these differ substantially in coverage, structure, and feature richness. Notably, many sources omit key match-level features such as in-play event data or betting odds, making feature engineering and dataset merging essential for comprehensive modeling efforts.
2.1 Available Datasets
Publicly accessible datasets for football include the European Soccer Database on Kaggle, which comprises data for 25,000 matches and 10,000 players across 11 leagues, and includes betting odds and limited event features. However, datasets often differ in league and season coverage, necessitating manual consolidation. For instance, the Open International Soccer Database, used in the 2017 Soccer Prediction Challenge, contains over 216,000 matches from 52 leagues but includes only goals as match events. The engsoccerdata R package [34], StatsBomb open data, and Wyscout event data [90] also provide partial access to structured football data. As noted, "spatiotemporal data is often only available to professional teams themselves", which limits academic replication and benchmarking.
2.2 2017 Soccer Prediction Challenge
The 2017 Challenge established a benchmark for evaluating predictive models using a shared dataset. Participants trained models on 216,743 matches and were evaluated on 206 unseen fixtures using the Ranked Probability Score (RPS) [42, 29]. The dataset's simplicity (limited to goals scored) was a deliberate design to maintain comparability across leagues and support generalizability.
2.2.1 Top Performers
Three top-performing studies exemplify different ML methodologies applied to rating-based feature sets. Hubáček, Sourek & Zelezný [59] used XGBoost on pi-ratings and achieved an RPS of 0.2063 and 52.43% accuracy, noting that "adding more model features, weighting aggregated data according to recency, and including expert guidance" could further improve performance. Constantinou [28] employed a Hybrid Bayesian Network on modified pi-ratings, emphasizing match outcomes over goal margins, reaching an RPS of 0.2083. Berrar et al. [8] introduced the Berrar ratings, applied XGBoost and k-NN to rating and recency features, and attained post-competition bests of 0.2054 RPS and 51.94% accuracy. They stressed that due to football's low scoring nature, feature engineering and domain expertise were "crucial" to performance.
2.2.2 Post-Challenge Studies
Subsequent studies have extended the benchmark. Robberechts & Davis [100] used Elo-based ordered logit models, while Elmiligi & Saad [40] proposed hybrid statistical-ML models, though with lower accuracy (46.60%). Razali et al. [96, 97] tested XGBoost, LightGBM, CatBoost, and TabNet on rating features, with CatBoost + pi-ratings achieving the best result of 0.1925 RPS and 55.82% accuracy, surpassing all 2017 Challenge entries. Conversely, Hubáček et al. [61, 60] showed that Berrar ratings slightly outperformed Poisson and Weibull models, although with modest absolute accuracy (48.54%).
Together, these works reinforce the value of football-specific ratings and boosting-based ensemble models in this domain, while highlighting that even simple baseline models like bookmaker odds remain hard to outperform.
3 Models
3.1 Model Objective
Defining the purpose of a match prediction model is crucial: it might aim for predictive accuracy (e.g., for contests or expert comparisons), inform betting strategies, or support performance analysis. In the latter case, interpretability becomes essential, as coaches and analysts seek models that "identify the most relevant features that are of importance to winning". Conversely, models used for betting also require strategic decision criteria, such as the Kelly Index [114].
3.2 Candidate Models
Model choice depends on the objective and user. High-performance goals may favor black-box methods, while interpretable models suit tactical or coaching needs. Classical classification models used in football prediction include Logistic Regression, Decision Trees, Random Forests, SVMs, Naïve Bayes, k-NN, and Bayesian Networks. More recent efforts have incorporated gradient boosting and deep learning. Comparative evaluations across leagues, such as those by Danisik et al. [35], Tax & Joustra [113], and Arntzen & Hvattum [4], highlight the need for generalizability across competitive contexts.
3.3 Model Interpretability
Interpretable models are increasingly prioritized, particularly by performance analysts and coaching staff. Studies employing SHAP [80], such as Moustakidis et al. [84] and Ren & Susnjak [99], have identified key performance indicators influencing match outcomes. Yeung et al. [124] proposed a framework analogous to GAP Ratings that emphasizes managerial and player-quality features, often derived from FIFA video game data. Random Forests, valued for their embedded feature importance, have also been used in hybrid frameworks (e.g., Groll et al. [53]), while models like Logistic Regression and Decision Trees retain their appeal due to inherent interpretability.
3.4 Ensemble Methods
Ensemble methods like boosting (e.g., XGBoost, CatBoost) and bagging (e.g., Random Forest) dominate recent high-performing models. The 2017 Challenge and follow-up studies affirmed that "gradient-boosted tree models... are currently able to achieve some of the highest performance". XGBoost [26] uses gradient descent to sequentially optimize weak learners, while CatBoost [93] introduces ordered target encoding to mitigate information leakage and enhance calibration of match outcome probabilities. Although less explored, Alternating Decision Trees (ADTree) [46] offer interpretable models using AdaBoost and may warrant further investigation.
3.5 Deep Learning Models
Despite success in other domains, deep learning is underexplored in football match prediction. LSTM networks [56] have been applied by Danisik et al. [35] and Jain et al. [67] to model temporal dependencies in league data, while Rahman [94] and Malamatinos et al. [82] have tested various architectures including CNNs. However, ensemble methods (particularly CatBoost) often outperform deep models, as shown in Malamatinos et al. [82]. Notably, TabNet [3], a deep learning model tailored for tabular data, achieved strong performance in recent work by Razali et al. [97], although the comparative advantage over boosted trees remains unclear.
4 Features
Feature engineering is central to football match prediction, encompassing various data types: rating-based metrics, match and player statistics, and external influences. The quality and structure of features often determine the performance ceiling of even the most advanced models.
4.1 Ratings
Ratings quantify team strength and are critical both as standalone predictors and as model features. Their integration into machine learning models has yielded superior performance compared to direct use for prediction (Table 1.2).
4.1.1 Elo Ratings
Originally designed for chess [41], Elo ratings have been adapted to football with adjustments for goal margin [65] and home advantage [104]. The win probability of team i against team j is calculated as:
Ratings are updated via:
where Ai(t)∈{0,0.5,1}. Extensions allow for variable K based on goal difference [65].
4.1.2 Pi-ratings
Pi-ratings [30] use separate home and away ratings per team, updated based on discrepancies between expected and actual goal differences. Expected goals are given by:
with b=10, c=3, and updates modulated via:
The approach emphasizes outcome over margin, with an ordered logit model used to derive match probabilities.
4.1.3 Berrar Ratings
Berrar et al. [8] introduced a logistic function for expected goals, using offensive and defensive strengths:
Updates are applied to both attack and defense ratings for each team, modulated by learning rates.
4.1.4 GAP Ratings
GAP ratings [118] generalize pi-ratings to predict non-rare match statistics (e.g., shots). Each team maintains four ratings (attack/defense for home/away contexts), updated as:
Parameters λ, ϕ1, and ϕ2 are optimized via least squares on observed vs. expected statistics, with early and late season matches excluded to avoid instability.
4.1.5 Betting Odds
Betting odds function as opaque rating systems and can be converted into probabilities via:
Although often used as baselines, they may be included as features if betting is not the model’s purpose. Studies show that bookmaker odds are difficult to outperform, even with advanced models [5, 100].
4.2 Match Features
Match features (aggregated statistics like shots, passes, or cards) require historical preprocessing. Recency-weighted averages [22, 8], or exponential time weighting [37], are common. More granular event data, standardized by formats like SPADL [36], can capture richer context. VAEP [36] assigns value to individual actions, while VDEP [115] extends this to defensive metrics using tracking data. However, "event data focuses on on-the-ball rather than off-ball events," limiting full tactical insight. Replacing goals with expected goals (xG) is also explored “because expected goals are not as rare”.
Learn more on SPADL and VAEP:
4.3 Player Statistics
Player-level features include ratings from video games (e.g., FIFA [35, 92]), plus-minus models [64], and Hidden Markov modeling for form [88]. Arntzen & Hvattum [4] found that combining player and team ratings improved prediction. Aggregation into role-based metrics [24] and synthetic ratings like VAEP also provide utility.
4.4 Team Statistics
Team-level features extend beyond ratings to include streak and form metrics. Baboota & Kaur [5] define time-weighted streaks and recursive form updates.
Chemistry between players [12], as measured by co-action VAEP, and passing network structures [27, 66], also serve as predictors of match performance.
4.5 External Features
External data like weather [89], travel, match officials, and crowd effects can be included without preprocessing. Social media sentiment analysis [123, 72, 70] offers novel, albeit noisy, signals. Market values, transfer budgets, and managerial age are additional proxies for team quality and potential.
4.6 Feature Selection Methods
Feature selection can follow filter, wrapper, or embedded strategies [126]. Random Forests inherently support feature importance via embedded methods [79], while tools like SelectFromModel in sklearn automate this. Relief-based algorithms [73, 75] and expert-driven approaches [62] are also used. Comparative evaluation of expert vs. algorithm-selected features is a recurring theme in the literature.
5 Evaluation
Model evaluation in football match prediction depends on context: predictive models require proper probability calibration and generalization assessment, while explanatory models prioritize interpretability and feature impact clarity.
5.1 Evaluation Metrics
Accuracy, defined as the proportion of correctly predicted outcomes, is widely used but insufficient for probabilistic outputs. Logarithmic loss (log loss) penalizes poorly calibrated predictions and is sensitive to confident misclassifications. The Ranked Probability Score (RPS), favored in the 2017 Soccer Prediction Challenge, evaluates cumulative probability accuracy, reflecting both correctness and certainty. While log loss punishes overconfidence, RPS rewards well-distributed and well-ranked probabilities, with lower scores indicating better performance.
5.2 Comparative Evaluation Studies
Numerous studies compare model types and features across leagues and datasets. Groll et al. [53] showed that combining Random Forests with Poisson-based team ratings outperforms standard statistical models. Danisik et al. [35] reported that XGBoost surpassed both deep learning and traditional ML across multiple leagues. Razali et al. [97] found CatBoost with pi-ratings yielded the highest accuracy (55.82%) and lowest RPS (0.1925) on the 2017 Challenge dataset, confirming that "feature engineering was a critical factor". These studies reinforce ensemble methods' dominance when coupled with football-specific ratings.
5.3 Model Robustness and Generalisation
Robust evaluation necessitates attention to season structure and temporal constraints. Random sampling across seasons violates time-dependency; thus, walk-forward validation and sliding windows are recommended. Studies like Arntzen & Hvattum [4] and Danisik et al. [35] implement league-specific temporal validation. The 2017 Challenge’s holdout set of 206 matches served as a standard benchmark, but variation in evaluation procedures (e.g., filtering matches by betting odds [97]) hampers reproducibility. Generalization across leagues is rarely tested, though Baboota & Kaur [5] observed mixed results, with XGBoost generalizing better than Random Forest.
5.4 Explanatory Model Evaluation
For explanatory models, interpretability is central. Random Forests offer embedded feature importance; SHAP values [80] extend interpretability to any black-box model. Studies using SHAP (e.g., Ren & Susnjak [99]; Moustakidis et al. [84]) identify key features affecting win probabilities, such as team ratings and momentum indicators. Alignment between expert knowledge and model explanations increases trustworthiness, "agreement between experts and model feature rankings... was key to building confidence in predictive models". Hence, explanatory evaluation focuses less on numerical accuracy and more on insights generated for decision-makers.
6 Summary and Research Opportunities
This section synthesizes key takeaways from prior analyses, identifying both methodological progress and open questions in football match prediction modeling.
6.1 Lessons Learned
Machine learning models that incorporate domain-specific rating features consistently outperform traditional statistical models. Ensemble methods, particularly gradient-boosted trees (e.g., XGBoost, CatBoost), excel when paired with tailored features like Elo, pi-, or Berrar ratings. Feature engineering, notably from limited datasets, remains "a critical component in the predictive power of models". Random Forests and SHAP provide valuable interpretability, which is especially useful in applied contexts like coaching.
Nevertheless, standard benchmark datasets are rare, complicating direct model comparison. The 2017 Soccer Prediction Challenge is one of the few shared evaluation platforms. Deep learning methods have yet to consistently outperform ensembles, often due to data constraints or lack of appropriate temporal feature modeling.
6.2 Outstanding Issues
Several challenges persist. First, there's limited cross-league generalization testing; most models are trained and validated within single-league contexts. Second, there is no consensus on temporal validation practices, with many studies relying on non-temporal splits that inflate performance metrics. Third, the lack of standardized datasets and evaluation protocols "makes it difficult to compare results across studies", and many studies fail to release code or data, hampering reproducibility.
Additionally, interpretability remains underutilized. Few studies examine the alignment between model explanations and domain expertise, which could bridge the gap between predictive and actionable models. Finally, while betting odds are frequently used as benchmarks, there is little exploration of their structural integration as features, despite their strong predictive baseline.
6.3 Future Directions
Future work should explore hybrid models combining statistical and machine learning techniques to better capture football’s stochastic dynamics. Deep learning models, especially those adapted to tabular and temporal data (e.g., TabNet or LSTM architectures), warrant further study; particularly in data-rich contexts.
There is also a pressing need for publicly available, temporally-structured benchmark datasets across leagues, incorporating player-level and event data. Expanding evaluation protocols to test cross-league and out-of-season generalizability will enhance methodological robustness. Lastly, integrating explainability frameworks like SHAP into training and validation loops could improve model trust and adoption in applied settings.
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
Bunker, R., Yeung, C., & Fujii, K. (2024). Machine learning for soccer match result prediction. In Artificial Intelligence, Optimization, and Data Sciences in Sports (pp. 7-49). Cham: Springer Nature Switzerland. https://arxiv.org/abs/2403.07669
To keep this article concise, please refer to the original paper for the full list of references.