Abstract Basketball games are often defined by scoring runs, where one team gains momentum while the opponent struggles to respond. Although it is widely recognized that such runs frequently decide the outcome of games, predicting when they will occur remains unexplored. This research addresses three key areas. First, we develop an algorithm to identify patterns that lead to scoring runs. Second, we present a model capable of predicting the onset of these runs. Finally, building on these findings, we propose a generative model that, based on partial sequences, suggests full sequences likely to result in a run. Our predictor employs a novel pattern recognition method using convolutional neural networks, achieving an accuracy of 81%. To our knowledge, this is the first study to explore and predict basketball scoring runs using this approach.
Abstract This study presents a deep learning-based system designed to enhance archery performance by analyzing athletes shooting motions and providing personalized feedback. Video data of four national-level Korean archers were collected between February and May 2024, and 17-joint coordinate data were extracted using pose estimation techniques. The full shooting sequence—from ready position to release—was captured and normalized for consistent analysis. Multiple deep learning sequence models, including RNN, LSTM, GRU, Bi-LSTM, and Bi-GRU, were implemented and evaluated to determine the most effective approach for recognizing distinctive motion patterns of individual archers. The developed system enables objective quantification of motion characteristics, supporting personalized training feedback and performance enhancement. Hyperparameters were optimized using Optuna, and early stopping was applied to prevent overfitting. The system visualized motion consistency and identified joints with high error rates, allowing athletes to recognize and correct deviations in real time. By quantifying individual motion characteristics, the system facilitated the design of personalized training programs, ultimately improving technical performance. This approach offers a novel method for ongoing monitoring and performance evaluation, demonstrating significant potential not only for archery but also for other precision-based sports.
A player’s role for a team can be distinct from their playing position. Positions are generally attributed based on where the players line-up relative to their formation, whereas roles can be defined by frequency of their actions. Hence, the method presented in this research, attributed player roles based on event data. Player role feature selection involved a semi-supervised machine learning approach, that extracted feature importance in the form of Shapley values. These values helped define the KPIs for pressing attacking players. By using the proposed role similarity approach, it is possible for recruitment departments to identify players that occupy similar roles as current players. Furthermore, the evolution of player roles across time can be evaluated, which has applications with performance analysts, as they can interrogate the constituent roles of each player and its influence on overall team performance. Hence, the proposed method can help uncover the optimal KPIs for a given set of roles, while having practitioner applications within elite-level performance analysis and recruitment departments. Future methods should combine physical data sources, such as from tracking data, to enable greater specificity in player role classification.
From the observational methodology approach, this study analyses definitive errors or losing blunders, i.e. errors that result in the loss of the game, in elite players at U8 level. An ad hoc observation instrument has been designed as a combination of field format and category systems, based on a thorough theoretical review of the internal logic of chess. The games were compiled in the ChessBase 17 program and analysed using Stockfish 16 NNUE via https://lichess.org/es . The moment in the game when the error occurs is extracted and recorded and coded using Lince software. The reliability of the records from the observation system developed was guaranteed by interobserver agreement, calculated using Cohen’s Kappa coefficient. This paper’s objective is achieved by means of the decision tree analysis technique, obtained using the CHAID procedure, taking the “impact of the error” as the predicted dimension. The results obtained have allowed us to conclude that the errors that lead to the loss of the game for elite U8 players are related to short-term calculation (tactical motifs, undefended pieces or checkmate) as opposed to long-term strategic errors.
The purpose of the current investigation was to describe transfers between Icelandic youth sports and to compare drop-out from sport between those doing single sports, those doing multiple sports without transferring, and those transferring between sports. 11,382,013 youth sport invitation records sent to over 40,925 young athletes over a two-year period were analysed. Drop-out and transfers between sports were determined using the first and last attendances of players in different sports. There was net transfer from gymnastics and swimming to other sports, as well as a net transfer from soccer to handball and basketball. Girls had a net transfer from athletics and individual games to team games while boys had a net transfer from team games to athletics and individual games. The percentage of players dropping out of sport was 35.5% for those doing a single sport, 6.5% for players doing multiple sports without transferring, and 18.1% for players doing multiple sports over the two-year period and transferring between sports. These differences between drop-out rates were significant for both girls (p < 0.001) and boys (p < 0.001). Young people should be encouraged to participate in multiple sports to avoid dropping out of sport before they become adults.
Wearable sensors combined with machine learning (ML) offer a promising approach for estimating joint kinetics in real-world settings, with potential applications in athlete monitoring and injury prevention. However, the variety of sensor configurations in previous studies complicates comparisons and optimal configuration selection. This study compared different wearable sensor configurations, comprising inertial measurement units (IMUs) and pressure insoles (PIs), to determine their influence on the accuracy of ML – based predictions of 3D knee moments during running. Sensor configurations ranged from one to four IMUs, with and without PIs. The dataset consisted of wearable and ground truth knee moment data from 19 recreational runners during treadmill running. Model performance of the convolutional neural networks was evaluated on an independent test set. Hyperparameter optimization (HPO) was applied to refine model architectures and training parameters. Performance gains by PIs and a greater number of IMUs were small but significant. The results after HPO confirmed similar performances between single- and multi-sensor configurations, suggesting only small benefits from additional sensors. Our findings highlight that both sensor configuration and model optimization play critical roles in achieving optimal performance. We provide practical recommendations for sensor selection, balancing accuracy and feasibility, to enable biomechanical assessments in real-world environments.
This paper presents a novel approach to analyzing basketball games. It uses image processing techniques to track player movements, evaluate passes and shots, and visualize game dynamics. The system employs player and ball detection methods, leveraging appearance embedding-based particle filters for robust tracking across consecutive frames. We generate trajectory diagrams that provide insights into team strategies and player performance by applying projective transformation to map coordinates from player feet to the basketball court. Key challenges addressed include improving tracking accuracy under dynamic conditions, minimizing over-detections in pass and shot judgment, and refining ball possession calculations. Experimental results show high tracking accuracy for players, but lower performance in ball tracking and shot detection, particularly in high-speed movements or when objects are occluded. The analysis also revealed that player and team behaviors, such as passing success rates and movement patterns, could be effectively visualized through trajectory diagrams. While the current system provides valuable insights into game strategies, further improvements are needed, particularly in enhancing the reliability of tracking, judgment of passes and shots, and clarity of trajectory in dense sequences of plays.
Abstract Computer sports methods use computational techniques to analyse and optimise athletic performance. Computer vision (CV) has emerged as a tool that offers objective data on techniques and tactics. Depth camera technology can support markerless kinematic analyses. This systematic review, following the Preferred Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, examined the integration and impact of depth camera technology in sports biomechanics over the past decade. Using databases such as PubMed, Web of Science, and Scopus, we identified and analysed 14 relevant studies. Depth cameras such as Microsoft Kinect and Intel RealSense have been used to analyse performance in various sports by providing biomechanical feedback in real time, improving athlete training, and implementing injury prevention strategies. This review highlights the technology’s cost-effectiveness and accessibility, extending from elite sports to community programs. It suggests further advancements with AI and machine learning to enhance personalised training and integrate virtual and augmented reality, which is promising for the development of sports biomechanics.
This study evaluates the predictive power of common performance indicators (PIs) in soccer for success- or scoring-related events (SREs) such as shots, corner kicks, and box entries. Using data from 102 Bundesliga matches, we applied five machine learning methods to assess how well 28 widely used PIs (e.g., passes, ball possession time, opponents outplayed) within a past time span (up to 15 minutes) predict an SRE in a future window (up to 15 minutes). We ranked PIs based on the mean Matthews Correlation Coefficient. Results show PI Dangerousity best predicts SRE Goal and SRE ShotTaken , while PI EntriesAttaThird is strongest for SRE Cornerkick , SRE EntryAttaThird , and SRE EntryOppBox . PI Dangerousity and PI SuccPassAttThird consistently rank in the Top 9, highlighting their predictive strength. Combining PI OutplayedOpp and PI TacklingsWon over a five-minute input window improves goal prediction within three minutes, outperforming random guessing by 6%. PIs based on rare events, such as goals and corner kicks, are less effective for SRE prediction, whereas those capturing frequent actions (e.g., final-third possession, Dangerousity, outplayed opponents) perform better. These findings highlight the value of in-game data for short-term event prediction and its potential applications in quantifying match momentum, optimizing live betting odds, and improving performance analysis.
Evaluating the quality of shots in basketball is crucial and requires considering the context in which they are taken. We introduce a graph neural network to process a graph based on player and ball tracking data to compute expected shot quality. We evaluate this model against other models focusing on calibration. The messages between spatial and temporal features are separated, and an attention mechanism is implemented, making the graph neural network interpretable. We use the GNNExplainer to further show the importance of node features. To demonstrate possible practical applications, we analyse the embeddings of the graph neural network concerning different situations like the mean of all player predictions or similarity between created shots and compare this to existing methods.
A major challenge for sports coaches and analysts is to identify critical elements of athletes’ movement patterns. A potentially relevant tool is machine learning, useful because of its ability to extract patterns from data. In the current study, we employed various deep learning frameworks, including Gated Recurrent Unit networks (GRUs), Convolutional Neural Networks (CNNs), and Multi-Layer Perceptrons (MLPs), to search for differences between elite and non-elite rowers using a rowing ergometer. The MLP model achieved an accuracy of 100% when using all input features, indicating that the problem is suitable as a machine learning task. Our research focused on using a limited amount of the data. Despite using fewer input features, the models managed to classify skill levels with reasonable precision, reaching a best performance of 77% accuracy for the model combining GRU and CNN architectures, 78% for the GRU model, and 94% for the MLP model. From a rowing perspective, the results suggest that movement coordination between upper and lower body limbs, as represented by different feature combinations, is informative in distinguishing between elites and non-elites. The current work suggests that machine learning may supplement human experts in sports coaching, analytics, and talent identification.
Fantasy sports have become increasingly popular, with millions of players engaging in strategic team management and competition. In the realm of Fantasy Premier League (FPL), effective player analysis and performance prediction are crucial for success in each game. This paper presents an innovative approach to enhance FPL analysis and performance prediction by integrating news sentiment and players’ injury with statistical data sources. A dataset of weekly news articles was enriched through pretrained transformer-based sentiment analysis toolkit and combined with different boosting and neural network algorithms for prediction tasks. Our findings demonstrate that integrating these features enhances model performance, with the CNN architecture achieving a reduction in MSE from 6.27 to 5.63 outperforming the state of the art model. These results highlight the potential of leveraging diverse data sources for more accurate predictions and informed decision-making in FPL.
This study presents a machine learning-based approach to predicting the outcosmes of NBA games, with the aim of enhancing decision-making in sports betting and performance analysis. Using a dataset spanning 20 NBA seasons (2003–2023), we incorporated key features such as team statistics, player performance metrics, and external factors like team fatigue and rankings. The methodology followed the CRISP-DM process, involving data preprocessing, feature selection, and model evaluation.
Football match result prediction is a challenging task that has been the subject of much research. Traditionally, predictions have been made by team managers, fans, and analysts based on their knowledge and experience. However and recently there has been an increased interest in predicting match outcomes using statistical techniques and machine learning. These algorithms can learn from historical data to identify complex relationships between different variables, and then make predictions about the outcome of future matches. Accordingly, forecasting plays a pivotal role in assisting managers and clubs in making well-informed decisions geared toward securing victories in leagues and tournaments. In this paper, we presented an approach, which is generally applicable in all areas of sports, to forecast football match results based on three stages. The first stage involves identifying and collecting the occurred events during a football match. As a multiclass classification problem with three classes, each match can have three possible outcomes. Then, we applied multiple machine learning algorithms to compare the performance of those different models, and choose the one that performs the best. As a final step, this study goes through the critical aspect of model interpretability. We used the SHapley Additive exPlanations (SHAP) method to decipher the feature importance within our best model, focusing on the factors that influence match predictions. Experiment results indicate that the Multilayer Perceptron (MLP), a neural network algorithm, was effective when compared to various other models and produced competitive results with prior works. The MLP model has achieved 0.8342 for accuracy. The particular significance of this study lies in the use of the SHAP method to explain the predictions made by the MLP model. Specifically, by exploiting its graphical representation to illustrate the influence of each feature within our dataset in predicting the outcome of a football match.
Analyzing dual-lane speed climbing videos provides critical insights into data-driven performance evaluation in sports climbing. This study introduces an enhanced deep learning approach based on 3D ResNets to classify and analyze speed climbing states. Leveraging an annotated dataset of 872 high-resolution videos covering 15 state combinations, the model integrates optimized 3D convolutions and residual connections, achieving significant improvements in classification accuracy and computational efficiency. With a test accuracy of 92.78%, the model significantly outperforms 2D CNNs and C3D models. Additionally, its lightweight architecture and reduced computational complexity equip it with the potential for real-time deployment in controlled environments. While challenges such as data imbalance and limited generalization remain, this research provides a robust technical framework for speed climbing video analysis and lays the groundwork for broader applications in spatiotemporal modeling and intelligent sports analytics.
Finite Markov chain modelling is a commonly used type of stochastic modelling employed in performance analysis of net games. Finite Markov chains are based on a state transition model which can be used to depict the game structure of net games as a succession of states which are defined as equivalence classes for game situations, e.g. service and return. Furthermore, the theory of finite Markov chains allows for the calculation of model variables which are of significant interest not only for validation but also for performance analysis, like wining probabilities or expected rally lengths starting from different states. By simulation, of a more-or-less of tactical behaviors one may study the impact of these tactics on overall success. A novel state transition model for table tennis is introduced in this study as extension of an existing model in the literature containing only the first offensive shot. The new model additionally contains subsequent shots since they may be perceived as being influenced by the first offensive shot. A sample of 105 single matches (49 female, 56 male) at the 2020 Tokyo Olympics was examined. The validation of the Markov property resulted in satisfactory results. The relevance of 26 transitions denoting specific tactical behaviors was obtained using simulation and subsequently compared between sexes. Results provide insights concerning the game structure of table tennis with a particular emphasis on the transition from the initial phase of rallies to the first offensive shot.
Studies to understand the shooting preferences of basketball players relied exclusively on data on shot location, which did not lead to concrete understandings because they contained no information on how they moved to that location. Therefore, this study tried to cluster the players' shooting tendencies using the tracking data of the players' movements during the game. To do this, we first created hand-crafted shot features that included information on the pre-shot movement. Using those features, the dissimilarity of shooting tendencies between players was computed by considering the shot set of each player as a probability distribution and calculating the Wasserstein distance between them. The clustering based on their dissimilarity resulted in more clusters than in previous studies and allowed for specific shooting styles to be defined. Clustering using Gower distance as a dissimilarity measure for shot features, including a categorical feature, extracted clusters of shots that are useful for understanding players' more detailed shooting tendencies. These results prove that it is not only the shot location but also how the player moved before the shot that is important to capture the player's shooting preferences.
Artistic Swimming (AS) requires complete execution and synchronization of movements for performance evaluation. The interest in objective and subjective performance analysis worldwide in sports via valid and reliable Artificial Intelligence (AI) tools is spreading depending on the required analysis parameters to design a novel system. This study investigated a novel application of the MediaPipe-based computer vision tool validation by examining biomechanical aspects and the objective performance impact in ballet leg and barracuda AS techniques. Twenty experienced AS athletes participated and executed these techniques under controlled conditions. Thirty-six recorded video trials were captured and analyzed via computer vision using MediaPipe, Kinovea, and AutoCAD (gold standard), with correlations calculated to assess the reliability of measurements and tools. The results indicated a non-significant difference (p<0.05) among the software tools, supported by one-way ANOVA and Bland-Altman tests. Notably, in ballet leg technique, maintaining alignment between the upper body trunk and knee in a line had a small correlation with other leg deviations; however, this aspect had a moderate negative correlation in scoring. Overall, this study suggests MediaPipe efficiency in computer vision for AS officiating and performance analysis, offering a reliable, real-time alternative to traditional methods and providing perceptions of AS techniques.
The role of firefighters has evolved from traditional tasks like rescuing cats from trees and extinguishing house fires to more complex land, sea, and air rescues. The increasing demands for public safety necessitate rigorous training and high fitness levels for firefighters to manage their daily tasks effectively. In this study, final assessments of fitness and anthropometric parameters were gathered from 746 Malaysian firefighter recruits. A k-means clustering algorithm was utilized to group the performance levels of the firefighters whilst a quadratic discriminant analysis model was employed to predict the grouping of firefighters based on these parameters. Feature importance analysis was used to identify the most significant parameters contributing to model performance. Concurrently, the Mann-Whitney test was used to determine the essential anthro-fitness parameters differentiating between the groups of firefighters. The k-means clustering identified two performance groups: excellent and average anthro-fitness readiness (EFR and AFR) groups. The model demonstrated a mean performance accuracy of 91% for training and 87% for independent tests. Feature importance analysis revealed that inclined pull-ups, standing broad jump, shuttle run, 2.4 km run, age, and sit-ups were the most significant parameters. The Mann-Whitney test showed that the EFR group outperformed the AFR group in all anthro-fitness parameters except for height, weight, and age, which showed no significant difference. This study highlights the critical role of specific fitness and anthropometric parameters in distinguishing high-performing firefighters. By identifying the most significant contributors to overall fitness, fire departments can better prepare their personnel to meet the increasing public safety demands. The high accuracy of the predictive model also suggests its potential application in ongoing firefighter assessments and training optimization.
