This report describes our synthesis of practical and methodological experiences with synthetic data generation across a diverse set of applied research contexts. We have examined how different design choices in synthetic data generation influence the usefulness of the resulting data, judged by the extent to which analyses, models, and decisions derived from synthetic data remain consistent with those obtained from observed data. In doing so, we recognise that the use of synthetic data is governed by multiple, and often competing, considerations. These include the intended objective of use (O), the structure and constraints of the original data (S), the choice of generation strategy (G), the enforcement of plausibility and domain constraints (C), the assessment of utility and validity (U), and the management of deployment and risk considerations (R).
While releasing original data would typically maximise utility, it would also entail the highest risk, whereas synthetic data occupy an intermediate position that requires careful evaluation across these dimensions. Rather than treating synthetic data as a single methodological choice, this report frames synthetic data generation as a sequence of interdependent decisions spanning O, S, G, C, U, R. The primary focus of the report is on understanding how these dimensions interact and how utility can be maximised through informed design choices, while acknowledging that risk and other constraints must be considered alongside utility. The synthesis presented here is intended to inform the development of a general decision-making framework, applicable across domains and use cases.
First, across all studies, synthetic data was generated to address clearly articulated constraints rather than being used as a generic augmentation step. Objectives included improving action and event recognition accuracy in team sports (Caruso et al., 2025; Morra et al., 2020; Hong et al., 2024), enabling learning in low-data or rare-event regimes such as jersey number recognition and small cohort studies (Bhargavi et al., 2022;Naughton et al., 2023), supporting domain adaptation and cross-domain generalisation in sensor-based human activity recognition (Khan et al., 2024), extending training intensity ranges beyond observed archives in endurance sport planning (Fister et al., 2022), predicting athlete performance attenuation under data scarcity (Cordeiro et al., 2025), and facilitating data sharing, collaboration, and exploratory analysis under privacy and confidentiality constraints (Warmenhoven et al., 2025; Naughton et al., 2023).
Second, the structure and constraints of the original data were analysed prior to generation. Across studies, the source data included structured sport training summaries (Fister et al., 2022; Cordeiro et al., 2025), annotated and labelled video frames for broadcast and tactical analysis (Caruso et al., 2025; Cabado et al., 2025 ; Hong et al., 2024), sensor-derived time-series from inertial, physiological, and wearable systems (Khan et al., 2024; Raymond et al., 2022), and tabular performance and monitoring datasets collected in applied sport science settings (Warmenhoven et al., 2025; Naughton et al., 2023; Hohl et al., 2024). Common constraints identified included small sample sizes, class imbalance, missing or noisy values, restricted variability, limited access due to privacy, and incomplete coverage of extreme or uncommon scenarios (Hohl et al., 2024; Baumgartner & Klatt, 2023; Bhargavi et al., 2022). These analyses informed the feasibility and scope of subsequent synthetic data strategies.
Third, a generation strategy aligned with both data modality and objective was selected. For vision-based tasks, computer-graphics-driven simulation and spatial synthesis were employed to generate realistic yet controllable representations of players, events, and environments (Morra et al., 2020; Caruso et al., 2025). For structured and tabular sport data, statistical, tree-based, and regression-driven synthesis approaches were adopted to preserve marginal distributions and relationships without imposing strict parametric assumptions (Shah et al., 2020; Naughton et al., 2023; Hohl et al., 2024). For complex sensor and time-series domains, deep and hybrid generative frameworks were integrated directly into learning pipelines, including GAN-based, semisupervised, and domain-adaptive architectures (Khan et al., 2024; Cordeiro et al., 2025). These choices reflected a consistent principle: the generation method was selected to match the structure and downstream use of the data rather than methodological novelty alone.
Fourth, plausibility and domain-specific constraints were enforced to ensure that synthetic data remained meaningful and realistic. Constraints were derived from sport-specific metrics and theory, such as training load and physiological stress measures (Fister et al., 2022), spatial and geometric constraints imposed by playing fields and broadcast views (Baumgartner & Klatt, 2023; Cabado et al., 2025), physical and biomechanical laws governing motion and impact dynamics (Raymond et al., 2022), and structured function representations designed to preserve interpretable relationships in high-dimensional sensor data (Khan et al., 2024). These constraints reduced the risk of generating statistically valid but practically implausible data.
Fifth, utility and validity evaluation were systematically performed. Synthetic datasets were evaluated using distributional similarity metrics, statistical equivalence testing, and downstream model performance comparisons, ensuring that analyses conducted on synthetic data led to conclusions consistent with those derived from real data (Hohl et al., 2024; Naughton et al., 2023; Cordeiro et al., 2025). In applied recognition tasks, improvements or stability in classification and prediction performance were used as indirect validation of synthetic data quality (Hong et al., 2024; Bhargavi et al., 2022; Khan et al., 2024).
Finally, deployment and risk considerations were explicitly addressed. Several studies emphasised the importance of interpretability, disclosure risk management, and responsible reuse, particularly when synthetic data were intended for sharing beyond the original data custodians (Warmenhoven et al., 2025; Naughton et al., 2023). Practical accessibility was also prioritised through the release of datasets, tools, and applications that lowered technical barriers for practitioners and researchers (Cabado et al., 2025 ; Naughton et al., 2023; Warmenhoven et al., 2025). These considerations positioned synthetic data not only as a modelling tool, but as an enabling mechanism for collaboration, education, and reproducible sport science.
2025-12-03
This report provides an overview of how the proposed research questions are developed using the SPIDER framework, based on the patterns and insights identified in the analysis.
Each question is aligned with specific issues reported in the literature, and each issue is linked to practical challenges that appear when synthetic data is used in sport settings.
The first question arises from recognising that real-sport datasets vary across sports, competition levels, and measurement systems. When models trained on one dataset are applied to another, the underlying distribution often changes. This challenge is directly noted in:
“As a result, models trained exclusively on synthetic data may exhibit limited generalizability when deployed in unconstrained or heterogeneous settings, unless domain adaptation techniques such as adversarial alignment or feature disentanglement are employed.” (Caruso et al., 2023)
This point is reinforced by the observation that:
“The domain shift is a common problem to all synthetic datasets and is being addressed by extensive research in the field of domain adaptation [23]; however, current research is mostly focused on images (as input), convolutional neural networks (as models/detector) and may not apply directly to spatio-temporal data. Developing domain adaptation techniques for such data is an interesting research question.” (Morra et al., 2020)
“In order to reduce the need for annotating large dataset as well as account for the size and imbalance of the real-world data, we generated two double-digit synthetic datasets—Simple2D and Complex2D with different levels of complexity as described in section 3.2.2. This helps to circumvent the domain shift when only synthetic data is used and improves generalization on real-world data for fine-tuning.” (Bhargavi et al, 2022)
“The novelty of the proposed model lies in leveraging the KA representation with the GAN, thereby leading to better handling of multidimensional data, better generalization, and improved interpretability through structured function approximation. In addition, the ability of domain adaptation makes the proposed GAN-based model well-suited for HAR where labeled data is scarce.” (Khan et al., 2024)
“Second, for cases when there are variables with few observations or with extreme outliers, users should be aware that these may increase the disclosure risk, even in the synthetic data sets. An example of this may be a data set that includes a world-record holder or world-leading performance, which would be identifiable in a synthetic data set if the data set containing the original performance was able to be linked.” (Naughton et al., 2023)
“Currently, the model’s training was exclusively conducted using synthetic augmented data. The collection and labeling of real-world data is ongoing. Based on our findings, we assert that the model exhibits a high degree of resilience to substantial noise levels, a feature we anticipate will enable effective generalization to real-world datasets. Nonetheless, evaluation with real-world data may necessitate refinements to the data augmentation techniques employed during the training…” (Hong et al., 2024)
“Also, while the FE model was used to great effect in generating synthetic data, it is likely that the scope of impact types that can occur is much larger than this modest setup can reproduce. This would require a more complex model with more motor functional control to generate truer to life impacts. Finally, this model was only trained and tested on the same athletes in the same sport. It is possible that the model would not extend to all other athletes playing American football, or the same people playing another sport with potential head impacts. While the main differences in the strategies were the performance characteristics, it’s possible with more varied data that certain kinds of impacts might not be well modeled with insufficient balance of real and synthetic data.” (Raymond et al.,2022)
These statements support the inclusion of the first question, which investigates how domain-shift severity relates to predictive accuracy across real-sport case studies.
Attention then moved to the link between synthetic and real data. Several studies described the need to evaluate how closely synthetic data matches real data and how well it supports modelling. The statement that:
“To evaluate the quality of the generated data we focused on assessing both, the fidelity of synthetic data by comparing diverse measures of similarity of data distributions, and the utility of these data by measuring the generalization capabilities of models trained on them. … Data fidelity refers to the degree of similarity between synthetic and real data. … Besides data fidelity, another important aspect to consider when generating synthetic data is data utility.” (Hohl, et al., 2024), highlights this issue.
Concerns about the risks of alternative modelling frameworks without validation are also raised:
“Future research should explore the potential of other such model frameworks across different data contexts, but it should be stated that use of these alternative models without prior validation poses obvious risks.” (Warmenhoven et al. 2025)
These points informed the second question, which examines the relationship between fidelity, utility, and predictive accuracy.
A third issue emerged from understanding the conditions under which synthetic data is produced. Several papers described situations where unconstrained synthetic data may fail to represent sport-specific structure. For example, it is reported that:
“Our validation focused on statistical properties and predictive performance, rather than incorporating biological plausibility into the SDG process. This limitation emphasizes the importance of data-centric AI frameworks that enforce physiological validity during synthesis. Thus, future implementations should integrate constraint-based mechanisms to ensure biologically plausible conditions, particularly when generating synthetic data for rare or underrepresented performance profiles.” (Cordeiro et al.,2025)
“When the AST generates the training plan based on the existing training sessions, it is confronted with the same problem as the other ML methods, i.e., how to handle objects of qualities beyond those presented in the testing database. Moreover, in the theory of Sport Training, this limitation even violates the training principle of progressive overload, stating that the progressively harder training plan needs to be performed by athletes in training to increase their fitness. As a result, a data augmentation method is proposed in the paper to overcome the problem. The method consists of the following steps: 1. identification of features in an archive of sport activities, 2. identification of interdependence among features, 3. identification of the most intensive training sessions, 4. generation of uncommon training sessions based on synthetic data, 5. data enrichment of the existing archive with the uncommon training sessions.” (Fister et al., 2022)
A related challenge appeared in sport training research, where it is described that models struggle to handle multiple scenarios:
“Of course, our comparisons are highly unfair and skewed toward our exact scenario. That is the point we are trying to make in this investigation. … Since there is no ground truth and it is unlikely that there could ever be an actually recorded ground truth of the magnitude and variation of our data, we demonstrate qualitatively, what the developed improvements could mean in practice.” (Baumgartner et al., 2023)
These statements supported the third question, which compares constrained and unconstrained synthetic data generation methods.
The final question is motivated by imbalances that appear across many sport datasets. This issue is highlighted in a team sport data:
“Finally, it should be noted that due to the nature of indoor sports games (handball and basketball), there is a clear class imbalance. Certain game situations such as regular plays and attacks are much more common than specific events like free throws or penalties.” (Cabado et al., 2025)
Concerns about extreme values and limited observations are also noted, as stated:
“The data generated by the model is currently not “pure” enough to be considered truly pseudo-real, and hence, cannot be used to generate pseudo-real test data for testing and validation but by using state of the art random number generator for initial state and if used in apropos scenarios will yield promoting results.” (Shah et al., 2020)
These observations guided the fourth question, which examines the need of controlled-class imbalance models in synthetic data.
| Prompt | R1 | R2 | R3 | R4 |
|---|---|---|---|---|
| Sample | Real-sport datasets | Real and synthetic versions of sport datasets | Synthetic datasets generated with vs. without specific constraints | Case-study datasets from multiple sports with known imbalance issues. |
| Phenomenon of Interest |
“As a result, models trained exclusively on synthetic data may exhibit limited generalizability when deployed in unconstrained or heterogeneous settings, unless domain adaptation techniques such as adversarial alignment or feature disentanglement are employed.” (Caruso et al., 2023) “The domain shift is a common problem to all synthetic datasets and is being addressed by extensive research in the field of domain adaptation [23]; however, current research is mostly focused on images (as input), convolutional neural networks (as models/detector) and may not apply directly to spatio-temporal data. Developing domain adaptation techniques for such data is an interesting research question.” (Morra et al., 2020) “In order to reduce the need for annotating large dataset as well as account for the size and imbalance of the real-world data, we generated two double-digit synthetic datasets—Simple2D and Complex2D with different levels of complexity as described in section 3.2.2. This helps to circumvent the domain shift when only synthetic data is used and improves generalization on real-world data for fine-tuning.” (Bhargavi et al, 2022) “The novelty of the proposed model lies in leveraging the KA representation with the GAN, thereby leading to better handling of multidimensional data, better generalization, and improved interpretability through structured function approximation. In addition, the ability of domain adaptation makes the proposed GAN-based model well-suited for HAR where labeled data is scarce.” (Khan et al., 2024) “Second, for cases when there are variables with few observations or with extreme outliers, users should be aware that these may increase the disclosure risk, even in the synthetic data sets. An example of this may be a data set that includes a world-record holder or world-leading performance, which would be identifiable in a synthetic data set if the data set containing the original performance was able to be linked.” (Naughton et al., 2023) “Currently, the model’s training was exclusively conducted using synthetic augmented data. The collection and labeling of real-world data is ongoing. Based on our findings, we assert that the model exhibits a high degree of resilience to substantial noise levels, a feature we anticipate will enable effective generalization to real-world datasets. Nonetheless, evaluation with real-world data may necessitate refinements to the data augmentation techniques employed during the training…” (Hong et al., 2024) “Also, while the FE model was used to great effect in generating synthetic data, it is likely that the scope of impact types that can occur is much larger than this modest setup can reproduce. This would require a more complex model with more motor functional control to generate truer to life impacts. Finally, this model was only trained and tested on the same athletes in the same sport. It is possible that the model would not extend to all other athletes playing American football, or the same people playing another sport with potential head impacts. While the main differences in the strategies were the performance characteristics, it’s possible with more varied data that certain kinds of impacts might not be well modeled with insufficient balance of real and synthetic data.” (Raymond et al.,2022) |
“To evaluate the quality of the generated data we focused on assessing both, the fidelity of synthetic data by comparing diverse measures of similarity of data distributions, and the utility of these data by measuring the generalization capabilities of models trained on them. … Data fidelity refers to the degree of similarity between synthetic and real data. … Besides data fidelity, another important aspect to consider when generating synthetic data is data utility.” (Hohl, et al., 2024) “Future research should explore the potential of other such model frameworks across different data contexts, but it should be stated that use of these alternative models without prior validation poses obvious risks.” (Warmenhoven et al. 2025) |
“Our validation focused on statistical properties and predictive performance, rather than incorporating biological plausibility into the SDG process. This limitation emphasizes the importance of data-centric AI frameworks that enforce physiological validity during synthesis. Thus, future implementations should integrate constraint-based mechanisms to ensure biologically plausible conditions, particularly when generating synthetic data for rare or underrepresented performance profiles.” (Cordeiro et al.,2025) “When the AST generates the training plan based on the existing training sessions, it is confronted with the same problem as the other ML methods, i.e., how to handle objects of qualities beyond those presented in the testing database. Moreover, in the theory of Sport Training, this limitation even violates the training principle of progressive overload, stating that the progressively harder training plan needs to be performed by athletes in training to increase their fitness. As a result, a data augmentation method is proposed in the paper to overcome the problem. The method consists of the following steps: 1. identification of features in an archive of sport activities, 2. identification of interdependence among features, 3. identification of the most intensive training sessions, 4. generation of uncommon training sessions based on synthetic data, 5. data enrichment of the existing archive with the uncommon training sessions.” (Fister et al., 2022) “Of course, our comparisons are highly unfair and skewed toward our exact scenario. That is the point we are trying to make in this investigation. … Since there is no ground truth and it is unlikely that there could ever be an actually recorded ground truth of the magnitude and variation of our data, we demonstrate qualitatively, what the developed improvements could mean in practice.” (Baumgartner et al., 2023) |
“The data generated by the model is currently not “pure” enough to be considered truly pseudo-real, and hence, cannot be used to generate pseudo-real test data for testing and validation but by using state of the art random number generator for initial state and if used in apropos scenarios will yield promoting results.” (Shah et al., 2020) “Finally, it should be noted that due to the nature of indoor sports games (handball and basketball), there is a clear class imbalance. Certain game situations such as regular plays and attacks are much more common than specific events like free throws or penalties.” (Cabadoa et al., 2025) |
| Design | Quantitative relational design evaluating domain-shift metrics versus model performance. | Quantitative correlation analysis relating fidelity/utility scores to predictive outcomes. | Comparative predictive modelling. | Explanatory quantitative modelling comparing prediction quality across imbalance levels. |
| Evaluation | Predictive accuracy, stability across datasets, generalisation patterns. | Model accuracy, error rates, calibration. | Prediction of rare, uncommon events, outlier behaviour or specific scenarios. | Sensitivity to multiple distributions, false-negative injury predictions, bias patterns. |
| Research Type | Relational | Relational | Prediction of rare injuries, uncommon performance extremes, outlier behaviour or specific scenarios. | Explanatory |
| Research Question | What is the relationship between domain-shift severity and the predictive accuracy of performance and injury models across different real-sport case-study datasets? | How do differences in fidelity and utility between synthetic and real data affect model accuracy when predicting athlete performance and injury outcomes? | Do specific constrained synthetic data generation methods improve prediction accuracy for performance and injury events compared with unconstrained methods across different sports datasets? | What effect does class imbalance have on the performance of models trained with synthetic data when predicting performance and injury across sport-specific case-study datasets? |