The Application of Machine Learning Techniques for Predicting Results in Team Sport: A Review

1. Introduction

• Sport-result prediction is challenging because outcomes are inherently unpredictable and many potential factors can affect them.The review focuses on final results rather than specific events occurring within a match.Source: §1 · p. 2
• The review focuses on machine-learning studies of team-sport results, while excluding individual sports and methods outside its stated focus.Prior reviews covered the area, and this review aims to provide more critical insight and broader paper coverage.Source: §1 · pp. 2–3
• The review asks whether ANNs outperform other ML models and examines successful-study characteristics, cross-sport predictability, accuracy evaluation, and future research themes.These questions frame its survey of machine learning for team-sport result prediction.Source: §1 · p. 3
• The review collates studies, identifies preferred and emerging algorithms and characteristics of successful applications, and develops future-research recommendations.Its stated contribution includes identifying research trends and opportunities across the surveyed work.Source: §1 · pp. 3–4

2. Methodology

• The review considered 31 papers from 1996–2019, covering nine sports and 35 sport studies.Some papers studied multiple sports; Figure 1 shows publication-year distribution, and Figure 2 shows the distribution of studies by sport.Source: §2 · p. 4, Figure 1 · p. 4, Figure 2 · p. 5
• Included studies used at least one ML technique, focused on team sports, and generally appeared in peer-reviewed journals or conference proceedings.Statistical, mathematical, or operations-research studies without ML were excluded; fuzzy methods and deep learning were outside the review’s focus.Source: §2 · pp. 4–5
• The review grouped algorithm variants into families to identify usage trends, including ANN, Decision Tree, Rule Sets, kNN, Boosting, SVM, Ensemble, Naive Bayes, and Bayesian Networks.Variants such as CART, C4.5, and J48 were grouped as Decision Trees; SMO was grouped with SVMs, while Naive Bayes and Bayesian Networks remained separate.Source: §2 · pp. 5–6, Figure 3 · p. 6
• Accuracy was the primary comparison metric, but binary- and multiclass results were not directly compared because accuracy becomes harder to achieve as class count increases.Three-class accuracies were excluded from comparative charts but retained in summary tables, as were results reported with other metrics.Source: §2 · pp. 6–7
• The authors extracted recurring future-research themes, coded each paper’s future-direction text against them, and counted theme frequencies in a histogram.The listed themes include richer features, improved feature selection and training, expert benchmarking, causal investigation, cross-sport application, and other ML algorithms.Source: §2 · p. 7

3.1. 1990s

• In the 1990s NFL study, an ANN predicted 11 of 14 week-16 games (78.6%) and 10 of 14 week-15 games (71.4%).Betting-line odds improved initial results; the ANN generally outperformed the unsupervised methods, including the best unsupervised method, SOM.Source: §3.1 · p. 8

3.2. 2000s

• Across other 2000s studies, reported best-model accuracies included 75% for NFL, 59.2% for EPL Soccer, 67.5% across four sports, and 74.3% for NBA.Kahn’s 75% was slightly better than expert predictions; Joseph et al.’s expert Bayesian Network achieved 59.2% on a three-class problem. McCabe & Trevathan’s BP-trained ANN averaged 67.5%, above expert predictions of 60%–65%; Loeffelholz et al.’s FFNN achieved 74.3%, above USA Today experts’ 68.7%; Zdravevski & Kulakov reported 72.8% with Logistic Regression.Source: §3.2 · pp. 9, 11, 12
• In 2000s Rugby studies, computer models matched or exceeded expert predictions, while Reed and O’Donoghue found Soccer predictions more accurate than Rugby predictions.O’Donoghue & Williams’s computer models correctly predicted 39–44.5 of 48 matches, versus a 40.7-match human-expert average; the Simulation Model performed best. Reed & O’Donoghue reported 57.9% accuracy for Soccer and 46.1% for Rugby, with their models outperforming human experts and an ANN best-performing.Source: §3.2 · pp. 10–11

3.3.1. 2010 - 2014

• Soccer studies reported 76.9% accuracy without predicting draws and 68% classification accuracy with no gain from expert-selected features; odds models reached around 70% ignoring draws versus around 46% for three-class prediction.For NHL outcomes in the 2012/2013 season, traditional statistics outperformed newer performance metrics using 10-fold CV, while the ANN achieved 59%. A cascading ensemble using statistical features achieved 60.25%.Source: §3.3.1 · pp. 13, 16, 18, 19, 20
• A study’s accuracy rose from 66.4% to just under 81% after adding rebounds, assists, steals, turnovers and blocks; NBA Naive Bayes reached 67% using 10-fold cross-validation, while NCAAB experiments reported an accuracy upper limit of 74%.Source: §3.3.1 · pp. 14, 18, 19
• For 2010–2014 American football, CART achieved 86% accuracy in NCAA Bowl classification and 82.9% in a later-season test, while an NFL ANN ensemble matched bookmakers.The NCAA study found classification better than numeric points-margin prediction; David et al.’s NFL approach used committees of committees and found 99.7% of data variance attributable to passing and rushing differentials.Source: §3.3.1 · pp. 15, 16, 18

3.3.2. 2015 - 2019

• For Dutch Eredivisie Soccer, combining bookmaker odds and public match data produced 56.1% accuracy, but the improvement was not statistically significant.Naive Bayes and ANN with PCA achieved 54.7% on public-data features; FURIA reached 55.3% on bookmaker odds, also not statistically significant.Source: §3.3.2 · p. 20
• In 2015–2019, SVM averaged 61.67% balanced accuracy across ODI Cricket teams, while an EPL model reached 69.5% by omitting draws.The cricket study used team-specific models and evaluated them with balanced accuracy and Kappa; the EPL study used binary Logistic Regression, so its result excluded draws.Source: §3.3.2 · pp. 21–22
• For Baseball, SVM classification achieved around 59% accuracy and significantly outperformed numeric prediction, but was not significantly more accurate than betting odds.Valero (2016) found no improvement from adding features beyond the top three ranked variables.Source: §3.3.2 · p. 22
• In international Rugby, the best simulated model achieved 74% accuracy, suggesting outcomes were harder to predict than in previous years.It used all prior World Cup data while violating Linear Regression assumptions, with one independent variable, no relative-home-advantage variable, and separate pool- and knockout-stage models.Source: §3.3.2 · pp. 22–23
• In soccer, the LSTM Regression model achieved 52.5% accuracy for three-class results and 70.2% for two-class results excluding draws.The study used player-level and match-history data, with 139 features overall and four EPL seasons comprising 1,520 matches.Source: §3.3.2 · p. 23
• The LMT model achieved 83% accuracy on NBA finals matches using eight RIPPER-selected features; defensive rebounds were selected by all three feature-selection methods.The study compared Naive Bayes, an ANN and LMT using 430 matches and 21 features.Source: §3.3.2 · pp. 23–24
• Ensemble methods achieved 91.8% test accuracy predicting NHL match outcomes, the highest accuracy among the tested algorithms and ensembles.Gu et al. (2019) used data from 1,230 matches across ten seasons, combining historical results, opposition information, player performance and rankings, and 26 team-performance variables.Source: §3.3.2 · p. 24
• The Open International Soccer Database enabled a 2017 challenge on 206 future matches; competition studies commonly combined ratings with machine learning, and accuracies levelled off.The open database contained over 216,000 matches across 52 leagues and 35 countries; its emphasis on widely available data meant it lacked highly specialized fields. The review suggests broad predictive-feature coverage may be important.Source: §3.3.2 · pp. 24, 25, 27
• In the soccer challenge, Hubáček et al. (2019b)’s feature-based XGBoost classification achieved 52.4% test accuracy; Constantinou’s Bayesian Network scored 51.5%, and Berrar et al.’s competition kNN rating-feature model scored 51.9%.Constantinou used four home- and away-team rating features as Bayesian Network inputs and showed predictions could be made when historical match data involved neither team.Source: §3.3.2 · pp. 25–26

4.1. Does the evidence from the literature suggest that ANNs produce better performance in terms of predictive accuracy than other ML models?

• ANNs appeared in 65% of reviewed studies, but the evidence does not show they consistently outperform other algorithms; alternatives achieved top results in some sports.Decision trees offer interpretability, while ensemble methods achieved top accuracy results in Ice Hockey; Bayesian networks can incorporate expert knowledge and work with minimal data.Source: §4.1 · pp. 27–30

4.2. What are some of the defining characteristics that can be drawn from successful studies?

• Successful studies increasingly used data-driven feature selection, and accuracy improvements coincided with larger feature sets in Ice Hockey, Soccer, and Basketball.Researchers also compared feature subsets—such as betting, player-level, expert-selected, and match-related features—with each other and with the full set.Source: §4.2 · p. 30
• Robust studies tested multiple training and validation splits, but randomly shuffled cross-validation can let future games predict past games; sport-specific work requires a new experimental process.Results comparing points-margin prediction with win/loss/draw classification were mixed, supporting testing both formulations where possible; the review recommends structured frameworks such as SRP-CRISM-DM.Source: §4.2 · pp. 31–33

4.3. Are the results of some sports inherently more difficult to predict than in others?

• Prediction difficulty cannot be attributed to sport alone: Soccer’s best recorded accuracy was 78%, while Rugby Union’s was 93% in a highly predictable World Cup context (O’Donoghue & Williams, 2004).Lower scores and draws can increase randomness and make multi-class prediction harder, while competition depth also matters; Rugby Union accuracy was lower in more competitive settings, at 46% and 74%.Source: §4.3 · pp. 33–35
• Ice Hockey reached 92% accuracy, compared with 86% in American Football and 83% in Basketball; its improvement was attributed to more features, more data, and ensemble methods.The review notes that difficult conditions include highly competitive, low-scoring matches with fewer possible score increments, while richer features, substantial datasets, and ensemble training can help; Cricket and Baseball remain under-researched.Source: §4.3 · pp. 35–37

4.4. What are the best ways to evaluate accuracy results in this application domain?

• Because studies differ in sport, data, predictors, labels, and competitions and benchmark datasets are scarce, the review recommends comparisons with betting, expert, majority-class, or random baselines.Betting odds can also serve as a sole predictor, but should not be included when the model’s goal is to beat bookmakers.Source: §4.4 · p. 37

4.5. What are the common themes of future research directions that can be identified from surveyed papers?

• Nearly 90% of papers identified richer feature engineering as a future priority, and the review encourages collaboration between sport-performance analysts and prediction researchers.Domain expertise may help create descriptive features, including performance indicators and external factors such as weather, venue, travel, and player availability.Source: §4.5 · pp. 38–39, Figure 10 · p. 39
• Nearly 40% of studies planned to test alternative algorithms and improve training methods, while a quarter planned larger datasets or applying methods to other sports.Applying sport-tailored methods across codes may conflict with plans to develop custom features for individual sports.Source: §4.5 · pp. 39–40
• Only 10% of surveyed studies identified feature selection as a future research priority, despite over-fitting risks from large feature pools and generally small datasets.The review recommends giving feature selection or dimensionality reduction attention alongside feature engineering and generation.Source: §4.5 · p. 40

4.6. Limitations

• The review included studies using at least one ML technique, potentially overlooking useful work in statistics, operations research, or mathematics, especially studies of the same sport.Researchers are encouraged to consult those literatures for insights into experimental approaches and results.Source: §4.6 · p. 40
• Because the review covered team sports only, research on individual sports may offer useful insights, particularly for incorporating player-level data into team-sport prediction models..Source: §4.6 · p. 40
• Accuracy comparisons within a sport are limited in reliability because studies often used different seasons, time frames, and predictive features.The review notes substantial variability even among studies of the same sport.Source: §4.6 · p. 40
• Predicting match outcomes does not by itself establish a profitable betting strategy, which also requires assessing whether the odds justify the risk of betting.The review discusses betting applications only briefly.Source: §4.6 · pp. 40–41
• Coaches may value how a team plays beyond win/loss/draw results, so future models could predict coach-defined categories such as good or poor wins and losses.A team may win while playing poorly against weak opposition or play well while losing to strong opposition.Source: §4.6 · p. 41

• A wide range of candidate algorithms and ensembles should be tested, continuing the shift away from relying on a single model such as ANN.Source: §5 · p. 41
• Rich feature sets matter more than large instance counts, and feature subsets should be compared robustly against full feature supersets.Source: §5 · p. 41
• Use order-preserving training-validation folds and compare predictions with betting odds, expert forecasts, and simple outcome baselines.The review notes that randomly shuffled k-folds can use future matches to predict past matches; benchmark datasets remain limited.Source: §5 · pp. 41–42
• Sport competitiveness and predictability vary with scoring patterns, score increments, and the probability of draws.Source: §5 · p. 42
• Sport-science and performance-analysis knowledge may improve match-feature identification, while Baseball and Cricket could benefit from more quality research.Source: §5 · p. 42