Forecasting sports outcomes through machine learning

1 Introduction

Over the past two decades, data-driven decision-making has changed how sports organizations interpret performance and strategy. Early work in sports analytics was largely descriptive, focusing on collecting performance indicators and match statistics to support tactical and managerial decisions (Mondello and Kamke, 2014; Passfield and Hopker, 2017; Boulier and Stekler, 1999; Boulier et al., 2010). Later studies introduced data mining and predictive methods as tools for turning sport data into operational knowledge (Passfield and Hopker, 2017; Delen et al., 2012). Subsequent work extended analytics into performance optimization, dynamic pricing, and fan engagement (Mondello and Kamke, 2014; Boulier et al., 2006; Hubáček and Šír, 2023). The research problem now concerns the reliability of these models under small samples, uneven data access, and a sport-specific context, rather than the feasibility of modeling sports data.

Wearable sensors, tracking systems, and mobile platforms now produce continuous performance data (Morgulev et al., 2018). These systems allow continuous monitoring beyond laboratory settings and into training and competition. However, the same expansion has made it harder to evaluate model development. Wearable sensors, tracking systems, social media data, contextual metadata, and manually coded events differ in scale, reliability, and sampling frequency. Recent studies, therefore, emphasize data integration, validity checks, and interpretability as persistent barriers to useful sports analytics (Mǎnescu, 2025; Meyer et al., 2024). These barriers matter because models used by coaches, medical staff, and analysts are often judged by reported accuracy even when the underlying data pipeline is only partly visible.

Methodological development in sports research parallels work in epidemiology and performance science, where quantitative models address injury incidence and workload (Wasserman et al., 2018). Structured surveillance systems and longitudinal performance data have expanded the evidence base for identifying risk factors and optimizing athlete health, linking sports medicine and applied AI.

Recent studies have increasingly used artificial intelligence (AI) and machine learning (ML) to model performance, forecast outcomes, and estimate injury susceptibility (Chmait and Westerbeek, 2021; Leckey et al., 2025). ML methods can represent nonlinear and multivariate relationships that are difficult to capture with traditional statistical models. Deep learning has been used for video, sensor, and contextual inputs, especially when the task requires temporal or multimodal representation learning (Jia et al., 2025). As Apostolou and Tjortjis (2019) observed, ML algorithms can model interdependencies between players and teams that linear models miss.

Nevertheless, persistent methodological issues continue to limit the robustness and generalizability of current research. Many ML studies rely on small, domain-specific datasets and inconsistent validation strategies. Simultaneously, model evaluation practices often conflate the assessment of algorithms with the evaluation of the sports indicators derived from them. Davis et al. (2024) emphasizes that sports data possess challenges that require specialized evaluation protocols distinct from those used in other ML domains. Similarly, Morgulev and Lebed (2024) argues that the traditional reliance on key performance indicators is insufficient to capture the multidimensional and situational aspects of athletic behavior, calling for conceptual models that integrate context, cognition, and dynamics of sports performance.

The integration of wearable technologies and Internet of Things systems introduces new opportunities and challenges for data acquisition, privacy, and model interpretability (Seckin et al., 2023). As noted by Singh (2020), the practical value of sports analytics depends on aligning model development with athlete privacy, health, and consent.

These concerns about consent, interpretability, and athlete welfare sit at the center of how recent work frames the sport industry as a whole. Glebova and Madsen (2024) describe a twin transformation in sport management, in which digitalization and sustainability advance as a single strategic problem rather than as parallel programs, and argue that digital tools produce value only when their social and environmental consequences are planned alongside their technical performance. Madsen and Glebova (2025) extend that reasoning into Sports Industry 5.0, a framework that moves past the efficiency and automation emphasis of Sports 4.0 and treats human-centricity, resilience, and equitable access as design requirements rather than as corrections applied after deployment. Predictive models are one of the settings in which those requirements are tested, since an injury-risk model informs return-to-play decisions, a performance model can shape selection and contract negotiations, and both are built from data collected on athletes. Whether such systems can be trusted therefore depends on validation quality, uncertainty reporting, and data governance as much as on reported accuracy. Mapping how those practices are currently handled across prediction tasks, sports disciplines, and algorithm families is a precondition for judging whether machine learning in sport meets the human-centered standard that this literature proposes.

Several prior reviews have already mapped parts of this literature, but each is narrower in scope than the synthesis attempted here. Wunderlich and Memmert (2021) produced one of the earliest structured overviews of sports outcome forecasting, but that review is conceptual and pre-dates the widespread adoption of machine learning, focusing instead on classical statistical rating systems and the distinction between systematic and unsystematic variance in results. van Eetvelde et al. (2021) and Jiang et al. (2022) reviewed machine learning applications specifically for injury risk, the former cataloging tree-based, support-vector, and neural-network classifiers for injury prediction across sports, and the latter narrowing further to workload-injury relationships in professional soccer; neither considers match outcome, player performance, or in-game event prediction. Glattke et al. (2022) and Dolson et al. (2022) are clinically oriented reviews of rehabilitation protocols and wearable sensor technology, respectively, and do not evaluate predictive modeling pipelines or algorithmic benchmarking at all. To the authors' knowledge, no existing review synthesizes machine learning applications across match outcome, player performance, injury, and in-game event prediction within a single, journal-restricted evidence base spanning 2020 to 2025. The present review addresses this gap by using a shared coding scheme, comparable across the four domains, to trace how algorithm choice, dataset modality, and validation practice vary by prediction task and sport discipline (see Section 4.3, “Comparison with prior work,” for a detailed contrast with each of these reviews).

This systematic review aims to analyze recent applications of machine learning in sports forecasting, identify the predominant prediction tasks, algorithmic techniques, and represented disciplines, and highlight methodological gaps and emerging trends. The following research questions guided this synthesis:

• RQ1: What types of machine learning models have been applied to predictive tasks in sports, including match outcomes, in-game events, injury prediction, and player performance?
• RQ2: Which machine learning models demonstrate consistent predictive performance across different sports disciplines?
• RQ3: What performance metrics are most frequently used to evaluate predictive models within each assessment category?

The remainder of the paper is organized as follows. Section 2 outlines the review protocol, search strategy, inclusion criteria, and data extraction process. Section 3 presents the results across four analytical dimensions: research topics, temporal publication trends, sports disciplines, and algorithmic techniques. Section 4 interprets the findings in relation to the research questions, compares them with prior literature, and discusses implications for both research and practice. Finally, Section 5 summarizes the main insights and outlines priorities for advancing reproducibility, transparency, and cross-disciplinary collaboration in sports forecasting research.

2 Methods

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) statement (Page et al., 2021). The procedure followed the PRISMA 2020 phases (identification, screening, eligibility, and inclusion) to ensure a replicable and transparent selection process. This framework was chosen for its emphasis on methodological transparency, reproducibility, and complete reporting in evidence-based research. The review protocol was defined by the authors before the searches were executed and is reported in full in Sections 2.1–2.4; it was not deposited in PROSPERO or another prospective registry.

The review aimed to identify and analyze the most frequently used machine learning algorithms applied to predictive modeling in sports across four primary categories: (a) match outcome prediction, (b) in-game event prediction, (c) injury prediction, and (d) player performance prediction. These categories were selected for their methodological distinctiveness and relevance to the field of sports analytics. Match outcome prediction is a foundational area of study within predictive modeling. Injury prediction contributes directly to athlete health management and the development of effective prevention strategies. Player performance prediction supports data-driven talent evaluation and optimization of athletic capabilities. In contrast, in-game event prediction reflects recent advancements in real-time data processing and event modeling enabled by large-scale data collection systems.

2.1 Information sources and search strategy

Two databases, Scopus and Web of Science, were selected because both function as multi-publisher citation indexes covering peer-reviewed journal literature across computer science, sports analytics, and machine learning. Scopus and Web of Science are not single-publisher repositories; both aggregate journal output from the major scientific publishers active in this field, including Elsevier, Springer Nature, Wiley, Taylor & Francis, SAGE, MDPI, Frontiers, IEEE, and ACM. Journal articles published in IEEE titles (for example, IEEE Transactions on Pattern Analysis and Machine Intelligence and IEEE Access) and in ACM titles (for example, ACM Transactions on Intelligent Systems and Technology) were therefore eligible for retrieval through the search strings reported in Table 1. The journal-only scope of this review introduced two specific exclusions. Conference proceedings indexed primarily through IEEE Xplore, ACM Digital Library, or DBLP fall outside the eligibility criteria of this review, which restricted the corpus to peer-reviewed journal articles with full methodological reporting. Preprints hosted on arXiv and clinical studies indexed only in PubMed without cross-listing in Scopus or Web of Science may also be absent. Google Scholar was not used as a primary source because it includes gray literature and duplicates that complicate reproducible retrieval. Two complementary Boolean search strings were applied, each adapted to the syntax of the corresponding database (see Table 1). The complete process of searching for articles and processing them was carried out from July 1 to September 16, 2025.

Table 1. Search strings used across databases.

2.2 Eligibility criteria

All records retrieved from the initial search were reviewed for alignment with the research questions and compliance with the predefined eligibility criteria.

The inclusion criteria for this systematic review were:

• Studies published between January 2020 and June 2025 (inclusive), reflecting all articles available in the databases at the time of the search conducted on July 1, 2025.
• Full-text journal articles written in English, allowing assessment of research methods and findings.
• Peer-reviewed publications that apply machine learning or predictive analytics techniques to forecast sports-related outcomes, including match results, in-game events, injury likelihood, or player performance.
• Studies that provide sufficient methodological detail to enable evaluation of model design, data characteristics, and validation procedures.
• Articles reporting quantitative performance results using recognized evaluation metrics (e.g., accuracy, precision, recall, F1-score, RMSE, or equivalent measures).

These inclusion standards were designed to ensure that all selected studies provided methodological transparency, quantitative validation, and empirical relevance to the predictive modeling of sports outcomes.

The exclusion criteria were:

• Studies that are not centered on predictive modeling or machine learning applications within sports contexts.
• Publications outside the defined time frame (pre-2020 or post-2025) or published in languages other than English.
• Non-peer-reviewed works, including preprints, theses, reports, book chapters, and conference abstracts, that lack complete methodological documentation.
• Studies without predictive modeling relevance, that is, those that do not implement machine learning techniques to forecast observable sports outcomes such as match results, in-game events, injury occurrences, or player performance.
• Studies focusing primarily on economic, financial, or administrative forecasting (e.g., revenues, ticket sales, or expenditures) were excluded.
• Studies lacking quantitative performance evaluation or model validation metrics, including those that do not report accuracy, precision, recall, F1-score, RMSE, or equivalent measures, or that fail to describe validation and comparison procedures in sufficient detail to support cross-study reproducibility.

2.3 Data extraction and synthesis

For each included study, relevant data were systematically extracted and organized using a predefined coding structure. Extracted variables included: (a) prediction category (match outcome, in-game event, injury prediction, or player performance), (b) machine learning algorithm(s) applied, (c) dataset characteristics, (d) performance metrics, and (e) sport discipline analyzed.

The extracted data were synthesized through descriptive and thematic analysis to identify recurring methodological patterns, algorithmic preferences, and performance trends across studies. The explanatory synthesis summarized the distribution of publications by topic, year, and sport discipline, while the thematic analysis focused on the most frequently used machine learning algorithms and evaluation metrics across categories. A quantitative meta-analysis was not conducted because the included studies differed in their prediction targets, sports disciplines, sampling units, validation designs, and metric definitions. In several cases, the same metric was calculated across different train-test splits, season partitions, or class-balance conditions, making pooled estimates difficult to interpret.

2.4 Methodological appraisal

The methodological appraisal was embedded in the eligibility and extraction process rather than implemented as a separate numerical risk-of-bias score. Each full-text article was checked for the presence of a clearly defined prediction task, sufficient description of the data source, identification of the algorithm or algorithms used, reporting of at least one quantitative performance metric, and description of validation or comparison procedures. Studies that did not provide sufficient information to assess these elements were excluded during the full-text assessment.

This appraisal should be interpreted as a methodological screening procedure, not as a formal risk-of-bias instrument. The review did not assign study-level quality scores, did not weight findings by risk of bias, and did not pool model performance statistically. The synthesis therefore emphasizes patterns in model use, reported metrics, and validation practices rather than definitive comparative effect estimates.

A formal, validated risk-of-bias instrument was deliberately not applied, for three reasons. First, established tools such as PROBAST were developed for clinical prediction models with a single outcome type and a comparatively homogeneous set of design choices; they do not map cleanly onto a corpus that spans four distinct prediction tasks (match outcome, player performance, injury, and in-game events), heterogeneous label types (binary, multi-class, and continuous), and data modalities ranging from season-level tabular statistics to video and sensor streams. Applying a single generic checklist across this range risked producing scores that looked precise but were not meaningfully comparable across domains. Second, no risk-of-bias instrument specifically validated for machine learning sports-prediction studies currently exists in the methodological literature, so any instrument used would itself require ad hoc adaptation and validation, which was outside the scope of this review. Third, retrospectively scoring 118 studies against even an adapted instrument would require re-reading each full text for design details, such as validation/cross-validation procedure, class balance, and external validation, that are not uniformly reported and, in many cases, are not extractable from the published text alone; a partially completed or inconsistently applied instrument would risk conveying false precision rather than genuine quality discrimination. This absence is a substantive limitation of the current evidence base as much as of this review; the issue is revisited, together with concrete recommendations for reproducibility, external validation, and bias reporting, in Section 4.

3 Results

3.1 Search results and study selection

The search and review procedures followed the predefined protocol. Table 2 summarizes the timeline of key stages in the review process, while Figure 1 illustrates the document identification, screening, and inclusion process following the PRISMA 2020 structure.

Table 2. Timeline of the search and review process.

The systematic search yielded 978 records, of which 531 were retrieved from Scopus and 447 from Web of Science. After removing 18 retracted publications and 580 duplicate records, 380 unique studies remained for screening. Nineteen review articles were excluded, leaving 361 records for title and abstract screening. During this phase, 213 records were excluded for failing to meet the inclusion criteria, leaving 148 articles for full-text assessment. After applying the methodological eligibility checks and verifying alignment with the defined research questions, 30 additional studies were excluded. Consequently, 118 articles met all eligibility standards and were included in the final synthesis.

The full extraction table for the 118 included studies is provided as Supplementary Table S1. It reports the sport discipline, prediction target, algorithm, metric, result, and dataset type for each study, with continuation rows used when a study reports multiple algorithms.

3.2 Findings and interpretation

3.2.1 Research focus areas

The analysis of research topics revealed four primary areas of focus across the included studies: match outcome prediction, player performance analysis, injury prediction, and in-game event forecasting (Table 3). Among these, match outcome prediction emerged as the dominant theme, accounting for 42% of the reviewed literature (n = 50). This prominence reflects the centrality of outcome forecasting in sports analytics, where predictive models estimate the probability of victory or defeat based on historical, contextual, and team-level variables. Studies in this category typically employ classification algorithms on tabular or event-based datasets, often emphasizing model accuracy, interpretability, and adaptability across different leagues and seasons.

Table 3. Distribution of research topics across the reviewed studies (n = 118).

Categories follow the coding scheme described in Section 2. Match outcome prediction covers studies that forecast binary or multi-class match outcomes. Player Performance covers studies estimating individual indicators such as workload, efficiency, fatigue, or expected contribution. Injury prediction encompasses studies that estimate injury risk or occurrence using biomechanical, physiological, or workload variables. In-game events covers studies that forecast within-match events, such as scoring actions, possession changes, or tactical sequences, using spatiotemporal or video data.

Player performance analysis was the second most prevalent topic, accounting for 29% of the included studies (n = 34). This category explores the use of machine learning to assess or forecast athlete-specific indicators such as physical workload, positional efficiency, tactical contribution, or expected performance rating. Such models frequently incorporate multimodal data streams, including physiological, biomechanical, and positional features, to identify talent, predict fatigue, and guide individualized training interventions. The growing use of player tracking and sensor-based systems has accelerated this research direction, enabling more granular, personalized analytics.

Injury prediction accounted for 21% of the reviewed corpus (n = 25), a share that shows sustained attention to athlete safety and workload management. These studies integrate variables such as biomechanical load, heart rate, acceleration metrics, or match frequency to identify risk patterns and support preventive interventions. Despite its growing consistency, methodological diversity remains high, reflecting differences in data quality and labeling standards across sports disciplines.

Finally, in-game event forecasting was the least represented topic (8%, n = 9). This area focuses on predicting dynamic events in a match, such as goal events, possession changes, or tactical sequences, using spatiotemporal or video-derived data. The scarcity of studies in this category likely reflects the computational demands and real-time constraints of sequence modeling, which require extensive labeled datasets and specialized architectures.

Table 4 shows that sports analytics research remains primarily focused on performance and outcomes. At the same time, emerging work on in-game events and player-centered models indicates a shift toward frameworks that link contextual, tactical, and physiological information to performance and strategic decision-making.

Table 4. References by topic.

3.2.2 Temporal distribution of publications

The temporal distribution of publications between 2020 and 2025 revealed a consistent upward trajectory in research output, culminating in a clear peak in 2024 (Figure 2). Overall, the number of studies increased from 12 in 2020 to 26 in 2025, with moderate fluctuations reflecting evolving research priorities, data accessibility, and funding trends within the global sports analytics community.

In 2020, publications were primarily concentrated on match outcome prediction (n = 6), with smaller contributions addressing player performance (n = 2), injury prediction (n = 2), and in-game events (n = 2). Exploratory studies and baseline comparisons of traditional machine learning models characterized the early phase of the decade. By 2021, research activity expanded to 17 total studies, driven mainly by outcome prediction (n = 8) and a growing number of injury modeling investigations (n = 4). These studies began incorporating contextual and workload-based variables, reflecting an early convergence of machine learning and sports science.

Output decreased slightly in 2022 (n = 13), but the range of prediction tasks diversified, suggesting that the field was transitioning from conceptual exploration to methodological specialization. In 2023, the number of studies increased again (n = 19), with match outcome prediction (n = 9) and injury forecasting (n = 4) remaining the most prominent areas of focus. The most notable surge occurred in 2024 (n = 31), with expansions in player performance studies (n = 14) and match outcome modeling (n = 11). This sharp rise coincided with the proliferation of accessible open datasets, the growth of sports data competitions, and the increased adoption of athlete-tracking technologies.

Although publication frequency slightly declined in 2025 (n = 26), the relative balance among categories stabilized. Player performance (n = 6), match outcome (n = 11), and injury prediction (n = 7) continued to show strong representation, while in-game event forecasting sustained a modest but steady presence. Across all years, injury prediction showed the most consistent trend, with an average of four studies per year, reflecting ongoing concern about athlete health and preventive modeling. Table 5 illustrates a research domain that has moved from outcome-centered modeling toward broader applications of machine learning in performance, health, and in-game analysis.

Table 5. References by year published.

3.2.3 Historical evolution of predictive methods

Beyond the volume of publications, the reviewed corpus also shows a discernible progression in which algorithms were used to address match outcome and related prediction tasks, consistent with broader trends in applied machine learning. In the earliest years covered by this review (2020–2021), studies concentrated on classical statistical and simple supervised-learning approaches: Logistic Regression, Naive Bayes, and single Decision Trees appear disproportionately among the studies published in this period (Table 5), often used as transparent, low-data-requirement baselines for outcome classification, in line with the rating-based and regression-oriented forecasting tradition described by Wunderlich and Memmert (2021). As larger and better-curated match- and event-level datasets became available from 2021 onward, ensemble methods, particularly Random Forest and, increasingly from around 2022, Gradient Boosting and XGBoost, became the dominant choice, reflecting their favorable balance of predictive accuracy, tolerance for mixed and noisy tabular predictors, and moderate computational cost (Section 3.2.5). This shift coincides with the movement from conceptual exploration toward methodological specialization already noted for the 2022 literature above. From 2023 onward, the corpus shows the clearest uptake of Neural Networks and their specialized variants, convolutional networks for video and image data, LSTMs and temporal convolutional networks for sequential data, and Transformer-style attention architectures for multimodal fusion, concentrated in studies with access to tracking, video, or sensor data rather than season-level statistics alone (Section 3.2.5).

This progression, from statistics-based and simple classifiers, to tabular ensembles, to deep and multimodal architectures, tracks the growing availability of high-resolution sports data at least as closely as it tracks advances in modeling technique. Ensemble methods did not disappear as deep learning emerged; rather, they remained the default for tabular, season-level problems while deep learning was adopted selectively where richer positional, visual, or sensor data justified the added complexity (Section 3.2.5). The most recent development visible at the boundary of this review's eligibility window is the emergence of large language models and multimodal foundation models applied to sports contexts. Because peer-reviewed journal studies applying these models to the four prediction domains examined here were not yet indexed in sufficient number in Scopus or Web of Science by the close of the search window (June 2025), this emerging direction is discussed qualitatively as a forward-looking trend in Section 4 rather than quantified within the main synthesis.

3.2.4 Sports disciplines analyzed

The disciplinary distribution revealed that sports analytics research remains concentrated in a limited number of sports, with soccer emerging as the most extensively studied discipline (n = 28), accounting for 24% of the total sample (Table 6). Basketball followed closely with 21 studies (18%), while cricket ranked third with 14 reports (12%). Together, these three sports account for more than half of all included studies, a concentration that reflects their status as data-rich, globally popular disciplines.

Table 6. Distribution of sports disciplines across the reviewed studies.

A smaller body of literature addressed esports (n = 10, 8%), which indicates growing academic recognition of competitive gaming within sports analytics. Multi-sport analyses (n = 8, 7%) often compared model transferability across disciplines and offered evidence on algorithmic generalization and cross-domain feature relevance. Tennis accounted for seven studies (6%), while running and American football each contributed four (3%). The representation of these individual and niche sports highlights expanding efforts to apply predictive frameworks beyond traditional team contexts.

Less frequently represented disciplines included baseball, volleyball, handball, hockey, badminton, weightlifting, combat sports, rugby, horse racing, archery, and cycling, each represented by one to three studies. This distribution reveals a clear imbalance in favor of team-based sports, which generate extensive, structured datasets amenable to supervised learning and performance benchmarking. The scarcity of research in individual sports likely reflects limited data availability, smaller sample sizes, and inconsistent performance metrics across events or athletes.

These discrepancies show how data accessibility and standardization shape research patterns. Team sports such as soccer and basketball benefit from event-capture systems, publicly available databases, and commercial analytics partnerships, which support reproducible model development. In contrast, individual sports often rely on proprietary sensor systems or manually curated datasets, which restrict reproducibility and cross-sport generalization. The growing inclusion of esports provides an illustrative counterexample, as standardized digital telemetry and open-access data pipelines have accelerated algorithmic experimentation and benchmarking in this domain.

Table 7 shows that the disciplinary distribution of sports analytics mirrors broader differences in data infrastructure and technology adoption. If dataset access improves and tracking systems become standardized across more sports, the methodological diversity currently concentrated in soccer and basketball may extend to a broader range of athletic domains.

Table 7. References grouped by sports discipline, sorted by number of reports.

3.2.5 Machine learning algorithms applied

The analysis of machine learning algorithms revealed a clear concentration around a small set of dominant methods (Figure 3). Random Forest was the most frequently employed technique, appearing in 45 studies (38% of the total). Gradient Boosting and XGBoost followed, reported together in 36 studies (31%). Support Vector Machines (SVMs) ranked third, used in 33 studies (28%), followed by Logistic Regression, with 31 studies (26%), and Neural Networks, with 30 studies (25%). Traditional classifiers such as Decision Trees (n = 19), Naive Bayes (n = 15), and k-Nearest Neighbors (k-NN; n = 13) were moderately represented. In contrast, Linear Regression (n = 4) and time series models such as ARIMA (n = 2) were infrequently used. Table 8 shows that ensemble-based methods such as Random Forest, Gradient Boosting, and XGBoost accounted for more than half of all implementations, mainly because they handle structured sports data with limited preprocessing and provide interpretable feature-level summaries.

Table 8. References grouped by algorithm (top algorithms).

Algorithm names follow common machine learning usage.

CNN, convolutional neural network; GBM, gradient boosting machine; k-NN, k-nearest neighbors; LSTM, long short-term memory; SVM, support vector machine; SVR, support vector regression; XGBoost, extreme gradient boosting.

A comparative breakdown by prediction category reinforces this dominance (Figure 4). Random Forest led player performance prediction (n = 14) and ranked second in match outcome forecasting (n = 19), behind Logistic Regression (n = 20), with additional applications in injury prediction (n = 6) and in-game event modeling (n = 6). Gradient Boosting and XGBoost exhibited similar trends, with prominent patterns in analyses of match outcomes (n = 17), player performance (n = 10), and injuries (n = 5). Neural Networks were used most often in tasks involving high-dimensional or temporal data, especially injury prediction (n = 6) and match outcome modeling (n = 16). SVMs and Logistic Regression remained common baselines because they are comparatively transparent and useful for benchmarking.

The main algorithm families can be distinguished by the type of structure they impose on the data. Logistic Regression models the log-odds of a class as a linear combination of predictors and is commonly used as an interpretable baseline, as in Madan et al. (2025) for soccer outcome prediction. SVMs separate classes by maximizing the margin between decision boundaries and can use kernels for nonlinear classification, illustrated by the multi-kernel comparison reported in Lee and Kim (2025) for esports outcomes. Random Forest combines many decision trees trained on resampled data and is used, for instance, by Knoll and Stübinger (2020) in soccer outcome modeling. Gradient Boosting and XGBoost build trees sequentially to correct previous errors and were applied by Wunderlich et al. (2025) for tennis match-result regression and by Hecksteden et al. (2023) for injury risk in team sports. Neural Networks learn layered nonlinear representations and are often extended into CNNs for image or video features, LSTMs and temporal convolutional networks for sequences [see Wu et al. (2022) for time-resolved football modeling], and Transformer-style attention architectures for multimodal inputs (e.g., Bendiaf et al., 2025; Sudasinghe et al., 2025). Graph-based models, including graph convolutional networks, appear less frequently in the reviewed corpus but are relevant when relationships among players, teams, or events are represented as networks.

This pattern suggests that algorithmic choice in sports analytics is more strongly influenced by data scale and structure than by the type of sport. Ensemble and boosting techniques are well-suited to tabular or event-level datasets because they tolerate mixed predictors and can perform competitively with limited feature engineering. Neural and hybrid architectures such as CNNs, LSTMs, Transformers, and graph convolutional models are more visible in data-rich contexts involving video, positional, relational, or sequence data, but they remain less common than tree-based methods in the reviewed journal literature.

Overall, ensemble-based algorithms currently form the main empirical baseline for predictive sports analytics. Deep learning is present, but its use is concentrated in settings where the available data justify the added complexity. Future work should therefore compare newer architectures against strong tabular baselines under shared validation protocols before attributing performance gains to model novelty alone.

The preceding paragraphs describe how algorithm families differ mechanically. Table 8, together with its continuation panel, provides the underlying reference list for readers who want to trace a given algorithm to the specific studies that used it, and Table 9 translates the mechanical differences into practical selection guidance by summarizing, for each family, its main strengths, its main limitations, and the prediction scenarios in which it was most frequently and most successfully applied within the reviewed corpus.

Table 9. Comparative summary of the main algorithm families: strengths, limitations, and the prediction scenarios in which each was most applicable within the reviewed corpus.

“Most applicable scenarios” reflects the settings in which each family was most frequently and most successfully reported within the 118 included studies, not a universal ranking; see Table 8 and its continuation panel ABOVE for the complete reference list per algorithm and Section 4 for cross-domain performance caveats regarding confounding by task, dataset, and validation design.

4 Discussion

This section interprets the review findings in relation to the research questions and the limits of the evidence base. The review mapped journal literature on ML for sports prediction across topics, sports disciplines, algorithms, and validation practices. The results show that predictive modeling is most frequently applied in structured, data-rich settings, particularly match outcome and player performance prediction, and that ensemble models, notably Random Forest and Gradient Boosting, dominate model selection. These results follow from the wide availability of tabular event data and the practical demand for interpretable models.

Across applications, models typically combine historical performance, contextual variables, and tracking features to forecast results, player efficiency, or injury risk. Some studies combine physiological, tactical, and spatiotemporal information, but multimodal modeling is not yet evenly distributed across sports or tasks. Single-source datasets are gradually being replaced by integrated pipelines, though local data availability, custom preprocessing, and study-specific validation still drive most decisions.

4.1 Principal findings in relation to the research questions

4.1.1 RQ1: What types of machine learning models have been applied to predictive tasks in sports, including match outcomes, in-game events, injury prediction, and player performance?

Research remains concentrated in four main categories: match outcome prediction, player performance prediction, injury prediction, and in-game event analysis. Match outcome prediction accounts for the largest share of studies (Table 4), confirming its foundational role. These studies typically use historical match data, team statistics, and contextual variables to model probabilities of winning or scoring differences. Player performance prediction reflects growing interest in individualized modeling, where machine learning is used to estimate workload, fatigue, or expected contribution (Table 4). Injury prediction emphasizes biomechanical and physiological variables to assess risk and inform preventive strategies (Table 4). In-game event prediction remains less common but is an emerging area that draws on positional and video data for spatiotemporal forecasting (Table 4).

A relevant observation from the synthesis is that predictive difficulty varies substantially across sports, and these differences cannot be attributed solely to the number of available observations per season. Sports with larger schedules, such as baseball, might be expected to yield more stable predictive models than sports with fewer observations, such as American football. However, this pattern does not consistently appear in the reviewed literature. Basketball and soccer, which often fall within an intermediate range of 60–80 matches per season, frequently report higher or more stable predictive performance. This suggests that prediction difficulty is shaped by event randomness, contextual volatility, competitive parity, and the degree to which individual actions accumulate toward outcomes. Apparent differences in model accuracy across disciplines should therefore be interpreted in relation to sport-specific dynamics rather than as sample-size effects alone (Table 7).

4.1.2 RQ2 and RQ3: model performance across disciplines and the metrics used to evaluate it

Ensemble-based methods, particularly Random Forest and Gradient Boosting, frequently outperformed or matched deep learning approaches on small-to-medium structured datasets (Table 8). Logistic Regression and SVMs remained useful baselines for transparent classification tasks, while Neural Networks and hybrid architectures were used mainly in data-intensive sports where tracking, video, or sequence data were available. Across the reviewed studies, model performance appeared to depend less on algorithmic novelty alone than on data scale, signal complexity, validation design, and task definition. Evaluation practices reflected this same heterogeneity. Accuracy, precision, recall, and F1-score dominated classification studies, whereas regression tasks relied on RMSE, MAE, or R-squared. The choice of metric generally aligned with task type, but validation protocols and data partitions remained inconsistent: only a subset of studies used standardized benchmarks or reported enough detail to support exact replication. Publications after 2023 more often used multi-metric evaluation and reproducible workflows, but this practice is not yet universal.

4.2 Interpretation and implications

Machine learning in sports analytics is shaped primarily by data structure and accessibility. Ensemble and tree-based algorithms dominate because they accommodate mixed data types, small sample sizes, and limited feature engineering while remaining interpretable. Deep learning approaches are increasingly adopted in data-rich contexts involving tracking or video-based tasks. The concentration of research in soccer, basketball, and cricket reflects both dataset maturity and the global reach of these sports. Conversely, individual sports remain underrepresented due to limited open data. Esports show how standardized telemetry can improve reproducibility and support cross-domain comparison.

Predictive modeling in sports is no longer limited to proof-of-concept studies, but the evidence base remains fragmented. Validation practices, dataset transparency, and benchmarking vary widely. These issues limit reproducibility and make it difficult to judge whether a reported performance gain would transfer to another season, league, population, or sport.

4.3 Comparison with prior work

Previous systematic and narrative reviews have examined specific facets of predictive analytics and injury modeling in sports science. The present review differs in that it organizes recent journal literature across four application domains and compares algorithms, data types, metrics, and sport disciplines within a single synthesis. For example, Wunderlich and Memmert (2021) offered one of the earliest structured overviews of forecasting methods in sport, emphasizing rating-based models and outcome prediction across traditional team sports. Their work introduced methodological taxonomies but remained largely conceptual, focusing on classical statistical forecasting and on the distinction between systematic and unsystematic effects. In contrast, the present review extends this scope by incorporating machine learning and deep learning algorithms applied to outcome forecasting, performance estimation, and in-game event prediction across multiple sports and data sources.

Injury-related reviews, such as van Eetvelde et al. (2021) and Jiang et al. (2022), narrowed their scope to the predictive modeling of injury risk. van Eetvelde et al. (2021) cataloged early machine learning efforts for injury prediction, reporting primarily tree-based, support vector, and neural network classifiers, but that review excluded non-injury performance prediction models. Similarly, Jiang et al. (2022) synthesized evidence on the relationship between workload and injury risk in professional soccer, highlighting inconsistencies in GPS-derived load metrics and the role of artificial intelligence in risk estimation. The current review complements these works by placing injury prediction alongside performance and tactical prediction, thereby enabling cross-domain comparisons of model choice and data types (e.g., match statistics, sensor data, and video-derived data).

From a methodological standpoint, earlier reviews such as Glattke et al. (2022) and Dolson et al. (2022) remained clinically focused, evaluating rehabilitation protocols and wearable sensor technologies, respectively, and did not assess predictive modeling pipelines or algorithmic benchmarking. Including clinical and physiological data streams in the current review helps contextualize the use of sensor- and workload-based predictors in injury and fatigue models. Unlike prior works that examined medical rehabilitation (Glattke et al., 2022) or environmental physiology (Dolson et al., 2022) in isolation, this analysis situates these domains within a broader machine learning workflow spanning data acquisition to prediction.

The current review provides a cross-domain perspective that complements earlier discipline-specific analyses. By evaluating 118 studies spanning outcome forecasting, performance analysis, injury prediction, and in-game event recognition, it identifies how algorithm choice, data modality, and validation practice vary by problem type. The synthesis remains descriptive, but it provides a clearer basis for comparing where model evidence is strongest and where methodological reporting remains weak.

4.4 Strengths and limitations

This review provides a cross-disciplinary synthesis of machine learning applications in sports prediction. It follows PRISMA 2020 reporting principles, applies explicit inclusion and exclusion criteria, and codes studies by topic, algorithm, metric, dataset type, and sport discipline. The combination of descriptive and thematic synthesis offers quantitative mapping and qualitative interpretation of trends.

Several limitations must be acknowledged. The search was restricted to English-language journal articles indexed in Scopus and Web of Science. Although both databases function as multi-publisher citation indexes that include IEEE and ACM journal output, conference proceedings indexed primarily in IEEE Xplore, ACM Digital Library, and DBLP fall outside the journal-only eligibility criterion of this review. Preprints on arXiv and clinical studies indexed only in PubMed without cross-listing in Scopus or Web of Science may also be absent. The journal-only scope was selected to keep the corpus consistent and peer-reviewed, but it likely underrepresents fast-moving AI work that first appears at conferences or in preprint repositories, which is a particularly relevant gap in a field, such as machine learning, where methodological advances frequently appear at venues such as NeurIPS, ICML, or sport-analytics workshops before journal publication. In addition, the review used methodological eligibility checks but did not apply a formal study-level risk-of-bias or quality-scoring instrument, for the reasons detailed in Section 2.4. This limits the strength of any claim about comparative model performance and means that methodologically stronger and weaker studies contribute equally to the descriptive patterns reported here.

The heterogeneity of reporting practices also limited direct comparison of metrics and validation methods across studies. Accuracy, AUC, F1-score, RMSE, and related metrics were often reported under different sampling designs, class distributions, season splits, and benchmark conditions. For this reason, the review used descriptive and thematic synthesis rather than a quantitative meta-analysis. Most studies relied on structured tabular datasets, which may restrict the generalizability of conclusions to multimodal or unstructured data. These limitations do not invalidate the patterns identified, but they mean the findings should be read as a map of current journal evidence rather than a pooled estimate of model superiority.

Reproducibility and external validation deserve specific comment because they bear directly on how much weight the reported performance figures should be given. Across the included studies, external validation, testing a model on data from a different season, league, or population than the one used for training, was reported far less often than internal validation (a single train-test split or in-sample cross-validation), and code or trained-model availability was inconsistently reported and was not systematically extracted as a coded variable in this review. This asymmetry means that many of the accuracy figures summarized in Section 3 describe how well a model fit the data it was tuned on, rather than how well it would transfer to a new competitive context; where the two are conflated in the primary literature, this review reproduces that conflation rather than resolving it. Dataset heterogeneity compounds this problem: studies varied in sample size (from single-season, single-league datasets of a few hundred matches to multi-season, multi-league corpora), in class balance (particularly for injury and rare in-game events, where positive cases are a small minority of observations), and in whether performance was reported on held-out data at all. Because these design characteristics were not uniformly reported, it was not possible to test systematically whether they explain the apparent performance differences between algorithm families discussed in Section 4.1; readers should therefore treat cross-study performance comparisons in this review as descriptive rather than causal or definitive. Sources of bias likely include publication bias toward studies reporting favorable accuracy, selective metric reporting (favoring the metric on which a proposed model performs best), and inconsistent handling of class imbalance, none of which this review could quantify given the extraction fields available. These are flagged as priorities for the field's own reporting standards, not only for future reviews, in Section 4.5.

4.5 Future research recommendations

Four directions stand out for future work. The first three are grounded directly in patterns observed across the 118 included studies; the fourth is a forward-looking direction that falls outside the journal-only, 2020–2025 eligibility criteria applied in this review (Section 2.2) and is therefore discussed qualitatively rather than quantified.

The first direction is multimodal data integration combined with stronger external validity. Most of the studies in this review relied on structured match statistics, yet the most promising applications for coaching and sports medicine require models that can process continuous sensor data, event streams, and annotated video. This implies methods that align data collected at different sampling rates, handle missing or noisy observations, and adapt to domain shifts across competitions and seasons. Many existing studies evaluate models on a single league or a limited time window, which restricts the ability to judge whether reported performance persists under different competitive conditions. Future designs should therefore combine multimodal pipelines with explicit tests across multiple seasons, leagues, levels of play, and subgroups such as age category, sex, or playing position. Transfer learning, domain adaptation, and hierarchical modeling are appropriate tools for this goal because they can draw on related contexts while accounting for sport- and team-specific effects. Multimodal benchmarks require shared annotation protocols across clubs, leagues, and technology providers; publicly available resources such as event- and tracking-data releases from providers like StatsBomb and Wyscout, together with sport-specific competitions on platforms such as Kaggle, already provide a partial foundation for this kind of shared benchmarking in soccer and basketball, but comparable open resources remain scarce for injury prediction and for most individual sports. Attention to generalization is particularly important in underrepresented sports, where sample sizes are smaller, and the risk of overfitting is higher.

The second direction is the standardization of evaluation practices. Current work often uses custom train-test splits, different performance metrics, and non-comparable experimental settings, which complicate cross-study synthesis. Future studies should report data partitioning strategies in detail, employ repeated or nested cross-validation when sample sizes are small, and adopt evaluation schemes that mirror the intended deployment scenario, such as rolling-origin validation for temporal data. Shared reference datasets, baselines, and open-source code would distinguish methodological progress from dataset-specific tuning. The systematic use of multiple metrics, including calibration and probabilistic scoring rules, would improve transparency of model behavior beyond a single accuracy value. Journals and conference venues in sports analytics could accelerate this standardization directly, for example by adopting reporting checklists modeled on the TRIPOD+AI statement for regression- and machine-learning-based prediction models (Collins et al., 2024), adapted to the sport-specific validation designs (rolling-origin, leave-one-season-out) that this review's descriptive synthesis suggests are still inconsistently used; this would give future systematic reviews the structured fields needed to support the kind of formal risk-of-bias appraisal that the present review could not retrospectively apply (Section 2.4).

Beyond these two methodological directions, the practical implications for sports organizations are direct. Coaches and performance staff gain the most from models that report uncertainty alongside a point prediction, since a single accuracy or RMSE value does not indicate whether a specific upcoming match or athlete falls inside or outside the conditions under which the model was validated. Medical and conditioning staff evaluating injury-risk models should treat within-sample accuracy figures cautiously in the persistent absence of external validation, given the pattern documented in Section 4.4, and should prefer models, or model reports, that specify sensitivity to class imbalance and that have been tested on athletes outside the original training cohort. Analytics departments deciding between an interpretable baseline (Logistic Regression, Decision Trees) and a higher-capacity model (Gradient Boosting, Neural Networks) can use Table 9 as a starting point for that trade-off, but should weight the choice by data volume and by whether a human decision-maker needs to audit individual predictions, rather than by benchmark accuracy alone.

The third direction concerns ethical, legal, and interpretability requirements, which belong in the model design phase rather than as post-hoc additions. The increasing use of tracking technologies and wearable sensors raises concerns about player privacy, data ownership, and the potential misuse of predictive models in contract decisions, team selection, and return-to-play timelines. Future studies should therefore include clear statements regarding consent procedures, data governance, and compliance with privacy regulations. At the methodological level, research should prioritize explanatory techniques that allow coaches, medical staff, and athletes to understand why a prediction is made, which factors drive risk estimates, and the uncertainty associated with each recommendation. Explainable AI tools, model auditing, and fairness assessments help preserve practitioner trust in model outputs.

Causal inference and uncertainty quantification deserve more attention across all three directions. Causal machine learning and counterfactual modeling help distinguish correlational patterns from mechanisms that are relevant to intervention in injury prevention, training load management, and tactical decision-making. Bayesian methods and simulation-based estimation techniques, including Markov Chain Monte Carlo, offer ways to explicitly represent uncertainty and work effectively with the small and imbalanced datasets typical of elite sport.

The fourth direction is the emerging use of large language models (LLMs) and multimodal foundation models in sports prediction and understanding tasks. This literature falls outside the formal search protocol and eligibility criteria of the present review, which restricted the corpus to peer-reviewed journal articles indexed in Scopus or Web of Science through June 2025 (Sections 2.1, 2.2); consequently, none of the 118 included studies applies an LLM or foundation model to the four prediction domains synthesized here, and no claim of prevalence or comparative performance is made. It is nonetheless a direction worth flagging for future systematic work. Xia et al. (2024b) survey the broader integration of natural language processing and multimodal models in sports, cataloging applications from game summarization and tactical question-answering to fan engagement, and note that dedicated benchmarks for sports-specific reasoning, such as Xia et al. (2024a), show that general-purpose LLMs can approach expert-level performance on some sports-knowledge tasks while still lagging on complex, scenario-based reasoning. Closer to the predictive tasks reviewed here, Felice (2024) combined a conventional deep learning model with an LLM layer to generate human-readable explanations of match predictions for the 2024 Olympic handball tournament, illustrating a plausible role for foundation models as an interpretability and communication layer on top of, rather than a replacement for, the tabular and deep learning methods that dominate the current corpus. Two implications follow for the domains covered by this review. First, LLMs may offer a low-training-cost route to incorporating unstructured, text-based context, such as team news, injury reports, or commentary, that structured tabular models in this corpus rarely use, but this comes with open questions about data leakage, hallucinated statistics, and evaluation protocols that remain unresolved even in the exploratory literature identified here. Second, multimodal foundation models trained across video, tracking, and language may eventually reduce the need for the sport-specific feature engineering that currently limits cross-sport generalization (Section 3.2.4), though evidence for this within peer-reviewed, journal-indexed sports-prediction research is not yet available. Future systematic reviews with a search window extending beyond 2025, and one that is not restricted to journal articles, would be needed to evaluate this direction with the same rigor applied to the four domains synthesized here.

5 Conclusion

This systematic review synthesizes journal literature on machine learning applications in predictive sports modeling from 2020 to 2025. The findings show the dominance of ensemble-based algorithms such as Random Forest and Gradient Boosting, the continued use of Logistic Regression and SVMs as transparent baselines, and the selective adoption of deep learning in data-rich contexts. The review identifies four central application domains: match outcome prediction, player performance, injury prediction, and in-game event forecasting. Across these domains, data accessibility and validation quality appear to shape model choice as much as algorithmic novelty.

Future work should place greater emphasis on causal machine learning. Correlational models estimate associations, but they are often less useful when practitioners need to assess whether changes in training load, tactical structure, or recovery protocols would alter an outcome. Causal inference and counterfactual modeling connect prediction with intervention, especially in injury prevention and tactical evaluation.

Sampling and generalization challenges also require stronger uncertainty estimation. Markov Chain Monte Carlo and related simulation-based techniques can support uncertainty quantification, model calibration, and sensitivity analysis in settings with limited, imbalanced, or biased data. These methods are particularly relevant for injury prediction and rare in-game events, where class imbalance can make single accuracy values misleading.

Methodological progress in sports analytics is bounded as much by data openness and reproducibility as by algorithmic novelty. Open multimodal benchmarks, clearer reporting of validation protocols, and privacy-aware auditing would make it easier to determine which models transfer across seasons, leagues, sports, and athlete populations. Without those conditions, predictive accuracy will remain difficult to compare across studies.

These priorities align with the direction that recent work sets for the sport industry more broadly. The twin transformation framework of Glebova and Madsen (2024) treats digital capability and sustainability as one strategic problem rather than as separate initiatives, and Madsen and Glebova (2025) carry that reasoning into Sports Industry 5.0, where human-centricity, resilience, and equitable access define what technology in sport is for. Predictive models are one of the places where that definition is tested, since they convert athlete data into decisions about selection, workload, and return to play. Read against that agenda, the gaps documented in this review are governance problems as much as statistical ones. A model validated only on the season it was trained on has not been shown to transfer to the athletes or competitions where it would be used, and a model reporting a single accuracy value gives medical and coaching staff no way to recognize when it is operating outside the conditions it was tested under. The concentration of published work in a small group of data-rich professional leagues also leaves grassroots participation, women's competitions, and most individual sports without comparable evidence, which is difficult to reconcile with a commitment to inclusive sport. Future research directions in sports forecasting therefore extend past model accuracy to open benchmarks, external validation, explicit uncertainty, and privacy-aware data governance. Those conditions determine how much weight a published accuracy figure can carry for the athlete, coach, or clinician who has to act on it.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative AI (Claude, Anthropic; Grammarly, Grammarly Inc.) was used to improve the text's grammar and readability. All AI-assisted edits were reviewed and approved by the author(s), who retain full responsibility for the content, accuracy, and conclusions of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcomp.2026.1883327/full#supplementary-material

Full extraction table for the 118 studies included in the systematic review - sport discipline, prediction target, algorithm, evaluation metric, reported result and dataset type for each study. Supplementary material at the source

Footnotes

Edited by: Ekaterina Glebova, Université Paris-Saclay, France

Reviewed by: Yan Zhou, Hohai University, China

Dejan Ratomir Dašić, Union Nikola Tesla University, Serbia