SynDEc: A Synthetic Data Ecosystem

Karst, Fabian Sven; Li, Mahei Manhai; Leimeister, Jan Marco Article — Published Version SynDEc: A Synthetic Data Ecosystem Electronic Markets Provided in Cooperation with: Springer Nature Suggested Citation: Karst, Fabian Sven; Li, Mahei Manhai; Leimeister, Jan Marco (2025) : SynDEc: A Synthetic Data Ecosystem, Electronic Markets, ISSN 1422-8890, Springer, Berlin, Heidelberg, Vol. 35, Iss. 1, https://doi.org/10.1007/s12525-024-00746-8 This Version is available at: https://hdl.handle.net/10419/323570 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/4.0/ Vol.:(0123456789) Electronic Markets (2025) 35:7 https://doi.org/10.1007/s12525-024-00746-8 RESEARCH PAPER SynDEc: ASynthetic Data Ecosystem FabianSvenKarst1 · MaheiManhaiLi1,2· JanMarcoLeimeister1,2 Received: 22 March 2024 / Accepted: 2 December 2024 / Published online: 25 January 2025 © The Author(s) 2025 Abstract Given the critical role of data availability for growth and innovation in financial services, especially small and mid-sized banks lack the data volumes required to fully leverage AI advancements for enhancing fraud detection, operational efficiency, and risk management. With existing solutions facing challenges in scalability, inconsistent standards, and complex privacy regulations, we introduce a synthetic data sharing ecosystem (SynDEc) using generative AI. Employing design science research in collaboration with two banks, among them UnionBank of the Philippines, we developed and validated a synthetic data sharing ecosystem for financial institutions. The derived design principles highlight synthetic data setup, training configurations, and incentivization. Furthermore, our findings show that smaller banks benefit most from SynDEcs and our solution is viable even with limited participation. Thus, we advance data ecosystem design knowledge, show its viability for financial services, and offer practical guidance for privacy-resilient synthetic data sharing, laying groundwork for future applications of SynDEcs. Keywords Synthetic data· Data sharing platform· Data ecosystem· Financial services· Data scarcity JEL classification M15 Motivation In the wake of recent global crises, the enhancement of financial services has become a crucial driver for accelerating economic recovery, particularly in developing economies where these services are essential for expanding financial inclusion and fostering socioeconomic growth (DemirgüçKunt etal., 2022; Pazarbasioglu etal., 2020; White etal., 2021). However, given the financial services industry’s reliance on information, increasing data availability is key to success. This is especially true for smaller financial institutions, which lack the necessary volume of high-quality data to leverage current AI model advancements. This lack of data results in missed opportunities, with developing countries potentially losing out on up to 5% of GDP through improvements in fraud protection, operational efficiency, and workforce allocation (White etal., 2021; Zachariadis, 2020). Although the sharing of financial transaction data could reduce risks and improve transparency (Brodsky & Oakes, 2017), thereby driving economic growth (O’Leary etal., 2021), it faces significant obstacles related to privacy regulation and information security. Existing solutions such as open banking and federated learning have significant limitations. Open banking, which enables customer-approved data exchange between financial institutions, often produces unreliable data due to selective participation (He etal., 2023) and lacks coverage of B2B transactions (Preziuso etal., 2023). Federated learning, an approach for training a model without direct data exchange, faces scalability issues, restricts Responsible Editor: Gero Strobel * Fabian Sven Karst [email protected]h Mahei Manhai Li [email protected]; [email protected] Jan Marco Leimeister [email protected]h; [email protected] 1 University ofSt.Gallen, Institute ofInformation Systems andDigital Business, Dufourstrasse 50, 9000St.Gallen, Switzerland 2 University ofKassel, Information Systems, Pfannkuchstraße 1, 34121Kassel, Germany Electronic Markets (2025) 35:77 Page 2 of 28 participants to a single shared model, and lacks adaptability (Baabdullah etal., 2024; Chatterjee etal., 2024). Therefore, research is required to explore data ecosystems that facilitate the exchange of data between financial institutions and regulatory bodies while safeguarding the privacy of individual users’ information (Assefa, 2020). In the pursuit of establishing such an ecosystem enabling financial data sharing, the application of synthetic data generation emerges as a promising solution. Synthetic data, currently primarily used in financial services to tackle class imbalance in fraud detection models by synthesizing new fraudulent samples (Charitou etal., 2021), produces artificial data that if done correctly maintains privacy while capturing and generalizing the patterns and attributes essential for the training of machine learning models. Combining this with data sharing enables the creation of a secure and robust data ecosystem. While plenty of research on synthetic data generation exists, significant gaps remain for its practical application within data ecosystems. Research has largely focused on algorithm development, leaving critical questions unanswered about how to design an ecosystem for privacy-preserving data exchange with the capability to handle complex data and achieve interoperability across institutions (Oliveira etal., 2019). Additionally, there is limited guidance on which algorithms are most effective in a context where synthetic data is leveraged to be shared between institutions and not merely used to increase the amount of training data (Langevin etal., 2022). Practical strategies for integrating shared synthetic data within machine learning models are also sparse, though such strategies are essential for realizing synthetic data’s potential in AI applications (Sattarov etal., 2023; Strelcenia & Prakoonwit, 2023). Finally, incentives, for big as well as small players, necessary to encourage participation in a synthetic data-sharing ecosystem remain underexplored, despite being vital for fostering the cooperative engagement on which such ecosystems depend (Gelhaar & Otto, 2020). In response, our research seeks to answer the following questions: What architecture is best suited for secure data exchange? Which algorithms are most effective for data generation? What are the optimal strategies for utilizing shared synthetic data within individual institutions? And do the incentives within such an ecosystem effectively encourage participation? Furthermore, there is a need for specialized engineering and management methodologies tailored to the unique demands of financial services, where stringent privacy regulations and the complex nature of transaction data introduce distinct challenges (Oliveira etal., 2019). Our research goal is to provide design knowledge for a synthetic data ecosystem that enables financial institutions to share financial transaction data and generate utility from doing so. Our study contributes to the existing literature in two significant ways. First, it advances the field of data ecosystems by addressing privacy challenges and exploring the use of data from multiple institutions for machine learning (Brée etal., 2024). Second, it offers practical guidance for financial institutions on generating and utilizing synthetic data, including benchmarking different algorithms, setups, and training schemes. Given the current lack of guidance on the conceptualization and implementation of such systems, this leads us to the following research question: RQ: How to design a financial data ecosystem (SynDEc) based on synthetic data sharing? To address the RQ, the paper adopts a multifaceted approach to investigate architectural design decisions. It encompasses an examination of synthetic data generation techniques within the ecosystem, explores its implications for training predictive models, and seeks to identify and mitigate potential challenges to the ecosystem’s stability and functionality. Additionally, it assesses the generalizability of the derived principles beyond the domain of financial fraud detection. The paper is organized as follows: In the next section, we present an overview of data ecosystems in financial services and synthetic data generation. Next, we outline, the Design Science Research Methodology by Peffers etal. (2007), combining context-driven innovation and iterative development, which we use as our methodological foundation. In the first of our four design cycles, we diagnose the problem space through the meta (MR) and design requirements (DR) based on both literature and expert interviews. Based on this, our initial set of design principles (DP) is derived and instantiated as a system architecture. Building on this the second design cycle evaluates the feasibility of different synthetic data generation and integration methods. The following design cycle extends this by evaluating the proposed approach in new domains while also investigating improvements to the ecosystem based on data generation and exchange. Lastly, design cycle four takes a network view, investigating design elements to ensure early challenges frequently seen in data ecosystems can be overcome. Finally, we discuss the findings, outline limitations, provide a perspective for future work, and conclude with a brief summary. Related work Data ecosystems infinancial services The growing recognition of data as a critical asset for innovation, growth, and value creation has led firms to increasingly seek external sources to enhance their data capabilities (Bagad etal., 2021; Gelhaar & Otto, 2020). One promising Electronic Markets (2025) 35:7 Page 3 of 28 7 approach is the formation of inter-organizational networks, where organizations collaborate to share resources and knowledge (Gray & Sites, 2013). Within this context, data ecosystems have emerged as an effective framework for data exchange (Abbas etal., 2021; Heinz etal., 2022; Zuiderwijk etal., 2014). Defined as “a set of networks composed of autonomous actors that directly or indirectly consume, produce, or provide data and other related resources” (Oliveira & Lóscio, 2018, p. 4), data ecosystems are built around four key constructs: (1) actors, (2) their roles, (3) relationships among them, and (4) the resources they require. Actors in these ecosystems—whether organizations, individuals, or institutions—take on roles such as data consumers, providers, and intermediaries, each contributing uniquely to the ecosystem's function (Oliveira & Lóscio, 2018; van Schalkwyk etal., 2016). The roles they assume drive specific tasks, such as data intermediaries connecting various actors and data consumers analyzing and providing feedback to data providers. These interactions, and the dependencies that arise from them, form the relationships that underpin the ecosystem (Heimstädt etal., 2014; Oliveira & Lóscio, 2018). At the core of a data ecosystem, data platforms provide the technical infrastructure for processing and managing data from diverse sources, enabling various data applications. These platforms often incorporate data marketplaces, which serve as self-service platforms that connect data producers and consumers (Gröger, 2021). Another closely related concept is data spaces, which are frequently used to describe data-sharing ecosystems across organizations and thus will be used as synonyms in this paper (Otto etal., 2019). Building on this foundation, recent research has shifted its focus to the governance and operationalization of data ecosystems, particularly in the areas of data sovereignty (Jarke, 2017) and trust (Gelhaar & Otto, 2020; Schäfer etal., 2023), which are critical for ensuring secure and reliable data exchange. However, in their comprehensive review of data ecosystems, Brée etal. (2024) identified several gaps within the literature that are currently under-researched, among them data security and the integration of artificial intelligence and machine learning within data ecosystems. On the one hand, data security deals with ways data can be stored and shared within data ecosystems while remaining protected as well as the influence of such measures on the utility of data ecosystems (Brée etal., 2024). On the other hand, machine learning and artificial intelligence have become central to the formation of data ecosystems, yet there is a need for a deeper understanding of the requirements for sharing AI training data and how training on shared data should be conducted (Brée etal., 2024). Our research seeks to address these challenges by proposing a new type of data ecosystem centered on synthetic data, which offers a means to mitigate privacy risks while maintaining the benefits of data sharing. Additionally, we investigate strategies for maximizing the utility of shared data to enhance individual organizational performance, thereby contributing to both the theoretical and practical development of data ecosystems. With current research on data ecosystems, predominantly concentrating on applications within healthcare, Industry 4.0, and smart cities (Cappiello etal., 2020), this study tries to extend this focus to the financial services industry. Given the sector’s significant dependence on highly sensitive data and its advanced application of machine learning technologies, this context provides a suitable setting to address previously identified research gaps in data security and the implementation of AI models within data ecosystems. Current research on data ecosystems within the financial services industry can be broadly categorized into two research streams. The first stream centers on open banking, a customer-focused ecosystem where established standards facilitate the secure sharing of banking data with various actors within the financial services ecosystem, based on customer requests (Cosma etal., 2023). While this approach grants consumers greater control over their data, it also raises significant data security concerns due to the decentralized nature of data storage across multiple providers—a critical issue given the heightened sensitivity of financial transaction data (Y. Wang etal., 2018). Furthermore, open banking does not provide institutions with an efficient and secure mechanism for large-scale data exchange, which is essential for applications such as fraud detection and anti-money laundering (Asrow, 2021). The second stream of research revolves around federated learning, a methodology that completely eliminates data sharing by enabling distributed training of shared models, thereby ensuring compliance with privacy protection regulations (Awosika etal., 2024; Lei etal., 2023; Perez etal., 2023). However, federated learning presents significant challenges, including computational overhead, scalability issues, and still privacy risks, as malicious actors might be able to infer sensitive data from the model parameters shared during the training process (Baabdullah etal., 2024; Chatterjee etal., 2024). Additionally, the necessity for participants in a federated learning ecosystem to agree on a single model architecture, which is difficult to modify once established, further complicates its implementation. The constraints of existing solutions, coupled with the fact that data ecosystems do not emerge organically but instead necessitate strategic planning around a shared value proposition, have resulted in the lack of a comprehensive financial data ecosystem to date (Adner, 2017; Immonen etal., 2014). This is aggravated by a research gap in the development of specialized engineering and management methodologies tailored to the needs of such an ecosystem (Oliveira etal., 2019) which are especially critical in the financial services sector, where stringent privacy requirements and the complex nature of financial transaction data introduce distinct challenges. Consequently, further research is essential to Electronic Markets (2025) 35:77 Page 4 of 28 address these challenges and to delineate the architectural frameworks necessary for the creation of robust and secure data ecosystems within the financial industry. Synthetic data generation andits application Synthetic data can be defined as “data that has been generated using a purpose-built mathematical model or algorithm, with the aim of solving a (set of) data science task(s)” (Jordon etal., 2022, p. 5). This generation process can take many forms as comprehensively categorized by Bauer etal. (2024) into 20 distinct method types. Among these, generative adversarial networks (GANs) are the most popular. GANs learn by pitting a generator (synthesizes data from random noise) and a discriminator (classifies samples as real or fake) against each other, resulting in two highly skilled networks (Goodfellow etal., 2014). This architecture is highly adaptable, as discriminator and generator can be easily adjusted to new tasks (e.g., time series or graph generation) while being frequently the best-performing synthetic data generation method (Bauer etal., 2024). Another commonly employed synthetic data generation method is autoencoder-based architectures, especially variational autoencoder (VAE) (Kingma & Welling, 2013). VAEs are trained by mapping an input sample to a hidden representation, which is then mapped back to the original vector, thus creating a model that synthesizes valid data from a lower dimensional representation. This decoder model is then used to generate data from random noise which makes it especially useful for learning from data with disentangled features (Bauer etal., 2024). Third, recurrent neural networks, feedforward neural networks which include recurrent edges, are able to generate sequential data of arbitrary length. This makes them ideal for sequence generation tasks such as speech synthesis, music, and time series generation (Lipton etal., 2015). Finally, virtual environments are computer simulations in which algorithms interact with each other based on predefined rules, generating synthetic data in the process (Bonabeau, 2002). In the context of machine learning, synthetic data is primarily utilized in three key areas: (i) private data release, (ii) data de-biasing and fairness, and (iii) data augmentation for robustness (Jordon etal., 2022). As the focus of this paper is employing synthetic data for private data release, it will be investigated in more detail. Hereby, private data release describes the case where synthetic data is used to mitigate disclosure risk, allowing privacy concerns and regulatory issues to be circumvented by substituting real data with synthetic data (Esteban etal., 2017; Jordon etal., 2018). However, this comes with certain risks of disclosure, which users need to be aware of. While multiple risks exist, the most relevant is membership inference which seeks to determine if an individual was part of the original dataset (Bun etal., 2021; Jordon etal., 2022). This risk is particularly critical in the context of financial transaction data, as revealing a user’s membership in a specific bank’s dataset could enable malicious actors to carry out more targeted fraudulent activities, making fraud prevention more difficult. Research on dealing with membership inference risks in synthetic data, primarily drawn from the healthcare domain, can be divided into two major streams. The first stream focuses on achieving guaranteed privacy by modifying models to conform to differential privacy principles, ensuring both the data and the model are protected. Algorithms implementing this are the PATE-GAN (Jordon etal., 2018) or DP2-VAE (Jiang etal., 2022) architectures. The second research stream focuses on evaluating and managing privacy risks within acceptable limits for a given volume of published synthetic data, providing various metrics and thresholds for guidance (H. Chen etal., 2023; Yan etal., 2022). Popular measures are the nearest neighbor adversarial accuracy risk (Yale etal., 2020), the membership inference risk (Choi etal., 2018), and the meaningful identity disclosure risk (Emam etal., 2020). Furthermore, these measures have also been adopted by regulators such as the European Medicines Agency and Health Canada which both provide thresholds for identifying disclosure risk (Yan etal., 2022). As the complexity of models continues to grow, necessitating larger datasets, synthetic data has been applied in a variety of fields, where it is used to facilitate more efficient and effective development of AI solutions (Lu etal., 2023). In financial services, these have been mainly use cases that inhibit a strong class imbalance such as anti-money laundering and financial fraud detection. Here, synthetic data generation is used to increase the amount of data within the minority class, thereby increasing training efficiency (E. Altman etal., 2024; Hilal etal., 2022). The current landscape is largely dominated by GAN-based architectures especially Wasserstein GANs due to their superior training stability (Hilal etal., 2022; Sethia etal., 2018; Strelcenia & Prakoonwit, 2023). However, recent advancements have seen transformer-based architectures (Nickerson etal., 2023) and diffusion-based models (Sattarov etal., 2023) emerging as competitive alternatives to GANs. Due to the internal usage of this synthetic data, data privacy has not been a main consideration when building these models. Data privacy considerations have mostly been explored in academic studies that aim to make their synthetic data publicly available. These studies typically employ virtual environment-based systems, such as multi-agent simulations, which simulate financial transaction data by modeling interactions between known actors and behaviors (E. Altman etal., 2024; Jensen etal., 2023; Lopez-Rojas etal., 2016). While these approaches are very secure from a privacy perspective as real data is only used during model evaluation of the synthetic data, they require significant manual work to identify patterns and changing behaviors need to be detected first, before they can Electronic Markets (2025) 35:7 Page 5 of 28 7 be integrated into the simulation (Bauer etal., 2024). However, the automatic generation and sharing of synthetic data derived from real data have not been extensively explored. As privacy concerns intensify due to regulatory pressure and customer expectations, as well as a growing necessity for extensive datasets to support cutting-edge machine learning models (Hittmeir etal., 2019), employing synthetic data has the potential to address privacy challenges in data ecosystems. Recent studies by Sattarov etal. (2023) and Langevin etal. (2022) have begun to investigate this potential for financial services. However, these studies primarily focus on comparing different data generation methods and present synthetic data sharing as merely one potential application. This leaves significant research gaps regarding the mechanisms for data exchange, the optimal strategies for learning from cross-institutional synthetic data, and the incentives for participating institutions, reaching beyond financial services and tackling current challenges in data ecosystems in general. Moreover, these studies offer little guidance on the design of such an ecosystem, highlighting a clear need for establishing design principles and best practices. Research approach A design science research project was initiated to address a research gap in approaches to enhance privacy protection within data ecosystems while preserving data utility for machine learning applications. This need, combined with the financial services industry’s demand for solutions to address the limitations of inter-organizational collaboration in tackling financial fraud and anti-money laundering detection, prompted the research effort. This project is aimed at designing an innovative artifact that provides financial institutions with a tool to easily exchange high-quality data with each other enabling them to increase their fraud and anti-money laundering detection performance, creating guidance on how to implement such a system, as well as to evaluate its benefits and the associated privacy risks (Gregor & Hevner, 2013; Peffers etal., 2007). To achieve these objectives, we adopted design science research (DSR), a framework particularly suited for the iterative development of novel artifacts addressing solution spaces with broad implications for both theoretical and practical problem domains (Peffers etal., 2007) and providing theoretically justified prescriptive knowledge (Gregor etal., 2020). Following this paradigm, we focus on creating artifacts that serve organizational purpose, in our case enabling data sharing despite privacy restrictions, through a structured research process that rigorously builds and evaluates viable solutions (A. R. Hevner etal., 2004; March & Smith, 1995). Following Scheider etal. (2023), our artifact is a “model” (March & Smith, 1995), a type of DSR artifact that serves as a simplified representation of reality and accumulates specific design knowledge (March & Smith, 1995); thus, DSR provides a suitable framework for our study (A. R. Hevner, 2007; Iivari, 2007). Our model presents a structured approach to designing a data ecosystem under privacy and data complexity constraints, exemplifying a solution to the problem discussed in the earlier sections. Our methodological approach to DSR—the design science research methodology (DSRM) by Peffers etal. (2007) has six steps, arranged in sequential order, and incorporates an iterative research procedure by design. The process typically starts with the identification of a research problem with practical relevance, in our case, the challenge of data scarcity within financial fraud detection. Next, the solution objectives are designed to address the stated challenges and to create a meaningful artifact. In line with DSR, the insights gained from the build-and-evaluate process must be generalizable and therefore applicable in more generic settings (Jones & Gregor, 2007). Also, the design artifacts should result in profound disruptions to traditional ways of doing business (A. Hevner & Gregor, 2022). Based on these objectives and on theory, the artifact is designed and developed in the next research process step. Phase 5 comprises evaluation, which is necessary to test whether an artifact achieves the purpose of its creation and to prove this achievement using rigorous methods (Venable etal., 2016). The evaluation phase also helps one to better understand the problem at hand and thus to realize improved outcomes (A. R. Hevner etal., 2004). Due to the iterative nature of this process, it can be repeated until a suitable artifact is derived. The design knowledge in the form of DPs with their DRs and MRs generated during this process can be seen as a nascent design theory, capturing a general solution in a class of artifacts (Baskerville etal., 2018). While MRs are high-level, generalized goals that an artifact must satisfy to address a class of problems, providing the foundational objectives for artifact design (Walls etal., 1992), DRs are specific, actionable specifications that detail the necessary features and characteristics an artifact must have to fulfill the meta-requirements (Gregor & Hevner, 2013). Lastly, DPs are prescriptive, actionable guidelines derived from design requirements and grounded in both theoretical foundations and empirical evidence, providing clear instructions for creating artifacts that meet the specified requirements and address the underlying problem space (Gregor etal., 2020). Thus, especially the DPs can be used to guide actions in a wider range of problems, in particular, data ecosystems where data with a complex structure needs to be shared under privacy restrictions (A. R. Hevner etal., 2004). They contribute to the theoretical advancement of the information systems (IS) community and provide valuable guidance for practitioners in designing similar artifacts (Baskerville etal., 2018; Sein etal., 2011). Since the DSR approach requires integration into an organizational Electronic Markets (2025) 35:77 Page 6 of 28 context, the project was conducted in collaboration with the UnionBank of the Philippines, a rapidly growing digital bank, as well as a European neo bank with a focus on wholesale transaction banking. Both banks rapidly scaled their digital transaction infrastructure in recent years and are now looking for new ways to tackle transaction fraud and money laundering. While the banks granted us deep insights into the problem of limited transaction data and provided invaluable feedback through all cycles, it was decided that prototyping and evaluation would be conducted on publicly available datasets instead of real bank data to reduce risks and allow fast iterations to create a solid understanding of potential pitfalls. Within this DSRM framework, four iterative design cycles were conducted, thus allowing for continuous refinement of the artifact’s design based on feedback and derive insights (Mullarkey & Hevner, 2019; Sein etal., 2011). In the next paragraph, the activities in each cycle are introduced which are outlined in the following graphic (Fig.1). First, the DSRM project starts with problem identification and motivation, focusing on stakeholder problems and challenges. This was done by conducting a systematic literature review on data ecosystems, synthetic data, and financial fraud detection as well as semi-structured interviews with employees at different levels at our partner banks, who are engaged in data sharing initiatives, fraud detection or data analytics, and machine learning projects. Furthermore, these interviews were used to identify the objectives of our solution by deriving DRs and MRs. Next, we iterated the first “Design—Demonstrate—Evaluate” cycle. In the design phase, we formulated the initial set of DPs. These principles were then translated into a system architecture during the demonstration phase, specifying its material properties like algorithms and interaction layers. Subsequently, an evaluation was conducted, involving feedback from academics and industry experts through four semi-structured interviews. The outcomes helped evaluate the feasibility of the initial design and led to the refinement of selected DPs in the second iteration. In cycle 2, we conducted a literature review identifying suitable algorithms for synthetic financial transaction data generation and based on them, instantiated a prototype which was subsequently evaluated on a publicly available real-world credit card transaction dataset to identify the most suitable synthetic data generation algorithm, establish the feasibility of the solution, and demonstrate the privacy-preserving properties of synthetic data. Based on additional expert feedback as well as two large simulated financial transaction data sets, cycles 3 and 4 refine the existing DPs and introduce new ones where needed. While cycle 3 explores the local level of the ecosystem in more detail, cycle 4 focuses on the global level and cooperative challenges within the ecosystem. Throughout the DSRM cycles, Fig. 1 Steps and design cycles within our design science research study based on Peffers etal. (2007) Electronic Markets (2025) 35:7 Page 7 of 28 7 we iteratively abstracted the requirements, DPs, and system features. Thus, our main theoretical contributions lie in the abstracted artifacts, particularly the DPs, which are first derived in “Design of initial DPs” and continuously refined throughout the paper. Problem identification andmotivation The diagnosis phase consists of two tasks: understanding the problem and solution domain and defining the ecosystem’s requirements. First, we positioned our DSRM project within the domain of inter-institutional collaboration within financial services. With a major focus of such collaboration being financial fraud detection, a first literature review on data ecosystems, synthetic data, and financial fraud detection was conducted. Following the methodology by Webster and Watson (2002), four search strings were established (Table1) and the following databases: ScienceDirect, EBSCOhost, SpringerLink, IEEE Xplore, and AISeL, were queried for articles containing the previously defined search string in title, abstract, or the author keywords. Furthermore, only papers written in the English language and published within the past 5years were included. This initial query resulted in a total of 3794 papers, which were then filtered based on a screening of titles and abstracts. While for papers identified by the “Fraud Detection” query strings only papers were included that deal with financial transaction fraud and either focus on privacy or a multi-organizational context, for papers selected by the “Data Ecosystem” string the only inclusion criteria were a focus on data ecosystems. After adding more relevant papers through a forward and backward search a total of 61 papers were selected for inclusion in the literature review. The analysis of the first part of our literature review focusing on fraud detection revealed that the limited availability of data is a significant challenge, especially for smaller organizations (Kulatilleke, 2022; Pranto etal., 2022). Especially with increasingly sophisticated adversaries (Qiao etal., 2024) and thus, more complex fraud detection models, frequently built based on deep learning architectures, more data is needed for model training (Aurna etal., 2023; Hilal etal., 2022). This need for increasing amounts of training data is further aggregated by the extreme class imbalance of datasets (large datasets are needed for a sufficient number of samples in the minority class) as well as the fastchanging nature of fraudulent patterns (Abdul Salam etal., 2024; Ryman-Tubb etal., 2018). Tackling this, frequently, the proliferation of cross-institutional data is presented as a potential solution, to increase the amount of available data and train better and more robust models (Kong etal., 2024; Myalil etal., 2021; Qiao etal., 2024). However, due to the high sensitivity of financial transactions and the connected risk of privacy leakage, this exchange is usually prohibited by external regulation or internal guidelines (Bian & Zheng, 2023; Pranto etal., 2022; Ryman-Tubb etal., 2018). To overcome this problem, frequently federated-learning-based solutions are proposed, allowing the raw data to remain local, while a joined model is trained (Kong etal., 2024; Lei etal., 2023; Pranto etal., 2022). While these approaches show some promise, they retain significant drawbacks such as the computational overhead, scalability issues, and the necessity to agree on a single model architecture, which is difficult to modify once established (Baabdullah etal., 2024; Chatterjee etal., 2024). This leads us to the conclusion that there is a need for a data ecosystem that allows financial institutions to exchange data with one another while staying compliant with laws and internal regulations on data privacy and giving them the freedom to use this data to fulfill their specific needs. Definition ofsolution objectives Looking for potential solutions, we drew on the second part of our literature review focusing on data ecosystems providing relevant insights on how such challenges can be navigated and potentially overcome in the context of financial data. Particularly papers from the healthcare domain (H. Chen etal., 2023; Morley-Fletcher, 2022), investigations into the emergence (Gelhaar & Otto, 2020) and organization (Langer & Mukherjee, 2023) of data ecosystems as well as Table 1 Results of systematic literature search a Detailed filter criteria can be found at https:// anony mous. 4open. scien ce/r/ Synth eticD ataEc osyst ems801C/ Cycle1_ Initi alDes ignPr incip les/ README. MD ID Search string Hits Filter: titleaRemove duplicates Filter: abstractaFwd and Bwd search Total I “Financial” AND “Fraud Detection” 2471 336 449 30 5 35 “Transaction” AND “Fraud Detection” 990 139 II “Financial” AND “Data Ecosystem” 164 13 19 18 8 26 “Synthetic Data” AND “Data Ecosystem” 169 6 Electronic Markets (2025) 35:77 Page 8 of 28 the preconditions for data sharing (Fassnacht etal., 2023), were detrimental in deriving the design requirements presented in the following section. To extend our insights into the domain beyond academic literature next, nine semi-structured interviews with employees at various levels at our project partners, with a focus on fraud detection or data science, were conducted (for details, see Table2). Querying them for challenges as well as potential solutions for tackling data scarcity within their domain. Based on this, we formulated two meta-requirements (MR) that any solution must adhere to. MR1 emphasizes the ease of data sharing between financial institutions, encompassing both technical, legal, and collaboration aspects. The need for technical ease of use was informed by insights drawn from the medical field, where challenges related to tool availability and varying data standards were identified as hindrances to data sharing (van Panhuis etal., 2014). The legal dimension in ecosystem usability was motivated by diverse regulatory requirements across jurisdictions, as observed in existing approaches to sharing financial transaction data (Blake etal., 2019). Lastly, ease of collaboration was drawn from the ecosystem literature, where cooperative challenges were outlined as a major hurdle to data ecosystem development (Gelhaar & Otto, 2020). MR2 highlights the necessity of increased utility as a result of sharing data. This requirement emanated from discussions with our partners regarding their goal of establishing a data-sharing ecosystem and from the literature describing incentives for participation in data ecosystems (Gelhaar etal., 2021). Next, we refined the MRs into more specific DRs, drawing from literature as well as the knowledge of our project partners.1 To incentivize users to participate in data-sharing, setup as well as reoccurring costs need to be as low as possible, which is reflected in MR1 and propagates into DR1 and DR2. This is important because while a data standard for financial transaction data exists, different banks diverge from it (Major & Mangano, 2020), which was also confirmed during our interviews (“Different data providers have different schemas and transaction languages.”—Interviewee 2); thus, a data ecosystem needs to be flexible enough to accommodate various input data structures (DR1). This is particularly important as data needs to be regularly updated and the cost for these updates should be as low as possible. Furthermore, data privacy standards imposed by regulators and internal policies must be upheld (“In terms of data sharing we do not engage in anything, because this is the pain with financial institutions, we are really protective of our data”—Interviewee 8–1). Our interviews revealed that in the context of our partner institutions, this means that all real data must be processed locally within the financial institution (DR2). From a data-centric perspective, the performance of machine learning methods can be enhanced by increasing the volume of training data available (Sun etal., 2017). Thus, MR2 can be achieved by enabling the combination of data from multiple sources through the data ecosystem and making it accessible as a unified data source (DR3). Given the goal of creating an ecosystem that is applicable to multiple tasks, the absence of a dominant algorithm in many fields (e.g., fraud detection), and the insight from our interviews that banks prefer to build and exclusively own their solutions (“One model will not be enough, it will be a collection of models which answer different questions …”—Interviewee 8–1), the data ecosystem must support diverse types of algorithms (DR4). Additionally, the imbalanced nature of fraud data necessitates tools on the ecosystem to address data imbalances through filtering, oversampling, and undersampling (DR7), as most machine learning algorithms perform better on balanced datasets (Longadge & Dongre, 2013). As fraud patterns change quickly when discovered, the timely integration of recent fraud patterns into fraud detection algorithms is crucial (Benchaji etal., 2021; Zhu etal., 2021). As this is utterly important, two DRs were dedicated to achieving this. First, institutions should have the capability to automatically Table 2 Overview interviewees for solution requirements *Interviewee from UnionBank of the Philippines ID Job title Expertise Years of experience Length of interview Interviewee 1 Chief data scientist Data science 10years 00:51:10 Interviewee 2 Senior data scientist Data science 5years 00:37:34 Interviewee 3 Data scientist Data science 5years 00:36:30 Interviewee 4 Chief financial officer Fraud detection > 20years 00:38:36 Interviewee 5 Senior compliance officer Fraud detection > 20years 00:59:08 Interviewee 6 Junior compliance officer Fraud detection 4years 00:44:27 Interviewee 7*Head of the AI center of excellence Data science > 20years 00:19:25 Interviewee 8*Head of data science ventures Data science 10years 00:31:03 1 A detailed mapping from interview quotes to DRs can be found on: https:// github. com/ Far um an/ Synth eticD ataEc osyst ems/ blob/ master/ Cycle1_ Initi alDes ignPr incip les/ README. MD Electronic Markets (2025) 35:7 Page 15 of 28 7 an examination of the literature regarding the enhancement of machine learning performance through the incorporation of synthetic data was conducted, aiming to determine an optimal ratio of real to synthetic data (mix-in percentage). While some researchers only oversample the minority class using synthetic data (Charitou etal., 2021; Strelcenia & Prakoonwit, 2023), others train models exclusively on synthetic data (Sattarov etal., 2023) or combine real with synthetic data (Dahmen & Cook, 2019). Thus, it remains unclear if there is an optimal mix-in percentage that individual institutions should incorporate into their design. To find the optimal way to generate synthetic data for our ecosystem, this section investigates data generation configurations that utilize the entire dataset as well as those trained on distinct data subsets and further analyzes the benefit of different pre-processing steps during the synthesizing process. Due to the challenge of dataset imbalance, models tend to be biased towards the majority class, decreasing the quality of data in the minority class; mitigating this issue, oversampling can be applied during the generation process to enhance generator robustness, albeit at the risk of distorting dataset composition (too many positive samples) (Kiran & Kumar, 2024). Second, the construction of distinct synthetic data generators for each class has been proposed as an alternative solution. Enabling the generator to better capture the characteristics of each individual class. However, this results in the problem that the minority class generator is only trained with a small dataset, which might harm its generalizability (Eilertsen etal., 2021). To remedy this, Fan etal. (2022) have suggested a novel methodology where the generator for the minority class is pre-trained using samples from the majority class, thus circumventing the problem. Demonstration throughimplementation ofdifferent training anddata‑fusing schemes In this section, we operationalized the derived DPs into a prototype system in Python using a modified version of the synthetic data vault library (Patki etal., 2016). Building upon the architecture from design cycle two, the local layer was modified to accommodate for different generation schemes with and without oversampling as well as pre-training on the local level. Furthermore, the training scheme of the prediction model was modified so that the system was able to accommodate training with different mix-in percentages.5 Evaluation ofdifferent training anddata‑fusing schemes One challenge in evaluating the broader feasibility of our synthetic data sharing ecosystem is the lack of publicly available financial transaction data (Jensen etal., 2023). However, multiple researchers have shown that simulated financial transactions can be suitable for validating new models or even evaluating interventions (Langevin etal., 2022; Sattarov etal., 2023). Therefore, in this as well as the next cycle, we will use two datasets, one for anti-money laundering (IBM-AML6) and one for fraudulent transactions (IBM-CCF7), which were generated by using a multiagent-based approach, simulating actors that act according to predefined rules, thus creating a stream of transactions (E. Altman etal., 2024; E. R. Altman, 2019). The resulting datasets have the advantage of being magnitudes larger in size (IBM-AML: 31898238/ IBM-CCF: 24386900) than the data used in the previous cycle (IEEE-CIS: 1097231) and have a network structure more similar to the one in real data. However, due to its simulation-based nature, it might not Fig. 6 Updated system architecture (version 2) 5 The full implementation of Cycle 3 can be found here: https:// github. com/ Far um an/ Synt h eticD ataEc osyst ems/ blob/ master/ Cycle34_ Ecosy stemE valua tion/ README. MD 6 https:// www. kaggle. com/ datas ets/ ealtm an2019/ ibmtrans actio nsforantimoneylaund eringaml 7 https:// www. kaggle. com/ datas ets/ ealtm an2019/ creditcardtrans actio ns Electronic Markets (2025) 35:77 Page 16 of 28 inhibit all characteristics found in real data. As the selected datasets do not include a financial institution (IBM-CFF) or the number of financial institutions present in the data is too big (IBM-AML, 122333 different banks), the data was artificially grouped. This was done by segmenting the data based on the location of the individual (IBM-CCF)/ bank (IBM-AML) connected to a transaction, creating clusters that simulate the transactional networks of hypothetical financial institutions. As a result, the IBM-CCF dataset included four financial institutions with a relatively even data distribution, while the IBM-AML dataset emerged with seven banks of which two banks held over 75% of the data. This contrast in dataset composition affords a unique chance to explore the synthetic data sharing ecosystem’s functionality under a broad array of conditions. Moreover, an analysis of client distribution post-split for each provider highlighted significant disparities, aligning with the anticipated diversity suspected within multi-institutional datasets. Details on the specific distributions are outlined in Table7. To limit the variables of this investigation, the synthetic data generation model and the fraud prediction model were kept constant. For the synthetic data generation model, the previously superior TVAE-based generator8 with hyper parameters tuned to the individual institution was used. Similar to Cycle 2, a XGBoost classifier with hyperparameter selection using threefold-cross validation was chosen as the prediction model and the performance comparisons were done using the ROC AUC score on a holdout dataset (30% of the data). Before, investigating the approaches modifying the local layer, the transferability of synthetic transaction data sharing beyond transaction fraud detection was evaluated (Fig.7). To do this, we compared the average ROC AUC score between the dataset constructed for financial fraud detection (IBMCFF) and the one constructed for anti-money laundering detection (IBM-AML). Figure7 demonstrates that across both datasets, models trained with synthetic shared data surpassed those trained without it, enhancing the ROC AUC score by 3.6% in the transaction fraud dataset (IBM-CCF) and 6.6% in the antimoney laundering dataset (IBM-AML). This effect can be considered substantial within this context as even recently introduced fraud detection algorithms often only increase the ROC-AUC score by a few percentage points (Hashemi etal., 2023; Lebichot etal., 2021). This performance gain suggests that the data ecosystem’s effectiveness extends beyond merely detecting financial fraud but is also suitable for other use cases utilizing financial transaction data such as money laundering detection. Thus, confirming the versatility and potential of the synthetic data ecosystem in addressing a broad range of data challenges in financial services. Subsequently, we explore whether a specific mix-in percentage of real and synthetic data yields optimal results for machine learning performance. To accomplish this, we systematically assess the impact on model performance by varying the proportion of real and synthetic data used in training the models, exploring a spectrum from 0% (no synthetic data) to 300% (3 times as much synthetic as real data). Figure8visualizes this experiment. Observing the modest upward trajectory of the aggregated performance line (black), we can conclude that there is a positive effect of adding synthetic data. However, in contrast to the more volatile performance trends of individual banks (grey), it appears there is not a universally optimal mix-in percentage. Instead, distinct peaks in performance suggest that the most effective mix-in ratios vary by bank. Consequently, we infer that allowing banks to adjust the mix-in percentage independently is most beneficial. This insight has been integrated into DP5, which mandates that banks have the autonomy to determine their mix-in ratios, leading to the updated principle: DP5—Provides the capability to combine synthetic data to find optimal composition for the training of machine learning models given scenarios with data from multiple institutions. Finally, we explored various configurations and preprocessing methods for synthetic data generation to offer optimal guidance for setting up these processes at the local level. Essentially, there are two primary setups. The first, referred to as “full,” involves training the synthetic data generation model on the entire dataset. To mitigate the risk of the model predominantly generating samples from the majority class, versions that randomly oversample the minority class to a specified percentage of the data (“_OS{X}”) while training the synthetic data generator have been implemented. The second setup, “sep” entails training distinct generation models for each class. An extension of this approach, “sepPre” utilizes separate generators for each class but pre-trains the minority class generator with majority class data. The outcomes of these varied approaches are detailed in Table8. The analysis of the data presented in Table8 yields several key findings. Initially, the “full” model demonstrates its ability to surpass the baseline performance, yet models built on the same training scheme but utilizing oversampled data Table 7 Distribution of data across the different banks IBM-CCF IBM-AML Bank Pct of data Bank Pct of data Bank Pct of data 0 21.54% 0 5.93% 4 29.94% 1 18.58% 1 11.90% 5 45.35% 2 39.16% 2 2.87% 6 2.12% 3 20.72% 3 1.90% 8 A detailed description of the hyperparameter tuning procedure can be found here: https:// github. com/ Farum an/ Synth eticD ataEc osyst ems/ blob/ master/ Cycle34_ Ecosy stemE valua tion/ 02_ param Search/ README. MD Electronic Markets (2025) 35:7 Page 17 of 28 7 exhibit a notable decline in performance. Thus, leading us to the conclusion, that for financial transaction data, oversampling the data before training the synthetic data generation model is not suitable. Moreover, the “full” setup outperforms configurations where synthetic data generators are trained separately for each class (“sep”). This subpar performance stems from the “sep” model’s poor-quality synthetic data for the minority class, which fails to capture training data patterns due to limited training dataset size. However, when the minority class model is pre-trained using data from the majority class (“sepPre”), a significant performance improvement is observed, surpassing all other methods. This enhancement is primarily due to the model’s capacity to generate higher-quality samples of the minority class with greater variability. Further discussions with partner institution experts emphasized the advantage of creating class data separately as it enhances privacy by preventing leaks of sensitive information like fraud rates by independently Fig. 7 Comparison between models trained with and without synthetic data for both datasets Fig. 8 Effect of synthetic data mix-in percentage on performance Table 8 Comparison between different synthetic data generation models Dataset Method ROC AUC score Dataset Method ROC AUC score IBM-AML Without shared data 0.7168 IBM-CCF Without shared data 0.6817 Full 0.7371 Full 0.7042 fullOS_10 0.6435 fullOS_10 0.6618 fullOS_20 0.6199 fullOS_20 0.6360 sep 0.7209 sep 0.6817 sepPre 0.7473 sepPre 0.7323 Electronic Markets (2025) 35:77 Page 18 of 28 producing the samples for each class. Consequently, we have refined DP4 to encapsulate these insights: DP2—Provide the system with the ability to identify, validate, and apply context-specific synthetic data generation techniques with mutually agreed on over-sampling in order to remove private data, given guidelines or regulations on data sharing. Cycle 4: Network effects offinancial data sharing In cycle four of our DSRM project, we delve into the global data layer, guided by the literature and expert insights to address cooperative challenges within the proposed synthetic financial data ecosystems. Aiming to refine our DPs to enhance the ecosystem’s capability to effectively manage these challenges. Design ofmechanisms attheglobal data level Addressing the second aspect of expert feedback and informed by the literature on data ecosystems, this cycle focuses on the global data layer and its DPs to ensure that the created ecosystem is able to handle the challenges of data ecosystems described by Gelhaar and Otto (2020). Because cooperative challenges play a dominant role in the early stage of an ecosystem, the following cycle will focus on these (Autio & Thomas, 2014). In their paper, Gelhaar and Otto (2020) describe four major cooperative challenges that need to be addressed for a data ecosystem to emerge successfully. First, it is necessary to build trust between the participants. Second, it needs to be shown that all actors benefit from participating in the ecosystem. Third, it is important to identify the right number of participants. Fourth, interoperability needs to be enabled through the agreement on standards. Thus, the focus of this section is to evaluate existing DPs through this lens and analyze whether refinements or additional principles are necessary for the development of an ecosystem capable of effectively addressing these challenges. First, trust between ecosystem partners can be built in multiple ways. On the one hand, trust can be increased by adequate control mechanisms (Geisler etal., 2021), which is already reflected in DP3—Provide the system with a back-testing mechanism in order to ensure newly generated synthetic data matches in composition and fraud detection training performance with real data given that data quality cannot be independently verified which ensures sufficient data quality in the synthetic data ecosystem. On the other hand, Majava etal. (2016) show that intermediaries play a significant role in increasing participants’ trust in an ecosystem. In the financial services ecosystem, this role is typically held by public regulators. To incentivize them to participate in the ecosystem and allow them to ensure data quality and thus increase trust, we propose DP6—Provide access for external collaborators, such as regulators, to leverage the synthetic data within the ecosystem given a diverse set of synthetic data available. This gives regulators access to the ecosystem while adhering to the existing privacy measures. However, it remains unclear if access to purely synthetic data can provide enough value and thus incentivize their participation in the ecosystem. The next challenge data ecosystems face is that all actors need to benefit from participating in the ecosystem. While we already demonstrated in previous iterations that our data ecosystem is able to increase the overall performance, it remains unclear how this performance gain is distributed between institutions. To address this, further investigation is needed to check if adjustments to our design need to be made to create sufficient incentives for all institutions. Connected to this problem is identifying the right number of participants. While our previous cycles show that the ecosystem is beneficial if all institutions participate, it remains unclear if a similar effect exists, if only part of the institutions is included in the ecosystem. To incorporate this into our DPs, DP3 was extended to not only describe the monitoring of outgoing synthetic data but also cover the evaluation of performance gained by using the shared synthetic data from the data ecosystem. This results in DP3—Provide the system with a back-testing mechanism in order to ensure newly generated synthetic data matches in composition and fraud detection training performance with real data given that data quality cannot be independently verified. The last cooperative challenge that needs to be overcome is interoperability through the agreement on standards. At the moment, that is already incorporated in DP1—Provide the system with modular systems design in order to ensure independence of local data and cross-institutional proliferation of synthetic data given that the raw data is sensitive, from a data perspective where the local layer of the ecosystem is used to align the data so that it can be easily shared with the system later. Furthermore, we argue that creating DPs for the financial data ecosystem contributes to the standardization of the ecosystem from an infrastructure and ecosystem perspective and thus by creating these DPs we contribute to overcoming this challenge. Demonstration throughtheintroduction ofnon‑sharing entities andindividual performance benchmarking In this part, we further improved the prototype developed in Python, by altering the global data layer to allow for the participation of entities that do not contribute data. Additionally, we updated the system to track and report the Electronic Markets (2025) 35:7 Page 19 of 28 7 performance to each participating institution, thus allowing institutions on an individual level to see the performance gain from engaging in the data ecosystem.9 Evaluation ofecosystems withnon‑sharing entities andindividual performance benchmarking Similar to the previous cycle, this evaluation again utilizes the two synthetic datasets (IBM-AML and IBM-CFF), due to their high data quality, size, and diversity. Moreover, the ecosystem setup and evaluation scheme are adopted from the previous cycle, utilizing the “sepPre” training scheme. We start with evaluating DP6, which allows regulators to access purely synthetic data within the data ecosystem. To validate this DP, the synthetic data that is provided to the regulators need to be of sufficient quality for them to derive meaningful insights and effectively improve their models. However, as this cannot be easily evaluated, we use the performance of a prediction model trained on the data available to the regulator (only synthetic data) as a proxy for the quality of the data. As the architecture chosen in Cycle 3 generates separate models for different classes, more data of a specific class can easily be generated. This is especially relevant for cases where only synthetic data is used as no additional positive samples from the real data exist. Thus, the experiment conducted had two steps. In the first step, regulator models were trained on synthetic data with different amounts of minority class samples (indicated by OS_{percentage of minority class cases}). From this selection, the over-sampling ratio with the best performance was chosen and compared to the performance of the models trained at the different banks, once trained on a combination of real and synthetic data, and once trained with only real data. The results of this experiment can be seen in the following diagram(Fig.9). The regulator model, trained exclusively on synthetic data, exhibits performance that, while not matching that of the bank’s internal models (trained on a mix of real and synthetic data), remains significant. The model approaches the performance of the bank’s baseline models (trained on real data only), as illustrated in Fig.9. This capability offers considerable advantages to collaborators who would otherwise lack access to such data. Consequently, allowing regulators to access synthetic data emerges as an effective strategy to foster collaboration and enhance trust in the ecosystem. Therefore, DP6—Provide access for external collaborators, such as regulators, to leverage the synthetic data within the ecosystem given a diverse set of synthetic data available is validated and was added to our DPs for synthetic data ecosystems. Subsequently, the adjustment to DP3 is validated, checking if all banks profit from the synthetic data ecosystem and evaluating if the synthetic ecosystem including fewer institutions is still able to profit from the network effects of the ecosystem. To investigate this, we plot the performance of each institution against its baseline (score without any artificial data), which can be seen below(Fig.10). As evidenced in Fig.10 for each single bank in both datasets, the performance increases by combining real and synthetic data. Furthermore, looking at the rightmost panel of Fig.10, it can clearly be seen that there is a negative correlation (− 0.09) between the performance gained by participating in the ecosystem and the size of the bank. Thus, showing that small banks over proportionally profit from participation, providing a clear incentive for them to engage in the ecosystem. However, even if absolute performance gained by bigger banks is lower, we argue that they still have a sufficient incentive to participate due to their large volume of transactions, where even small changes in the Fig. 9 Regulator models using different resampled data and performance of the regulator model (only synthetic data) vs. the bank models 9 The full implementation of Cycle 4 can be found here: https:// github. com/ Far um an/ Synt h eticD ataEc osyst ems/ blob/ master/ Cycle34_ Ecosy stemE valua tion/ README. MD Electronic Markets (2025) 35:77 Page 20 of 28 fraud detection percentage result in a high absolute sum of prevented losses. These results lead us to the conclusion that all banks contributing to the ecosystem profit from their involvement and thus the designed ecosystem is able to overcome another one of the previously outlined challenges. Next, we investigate our synthetic data ecosystem for cases where not all institutions engage in synthetic data sharing. To achieve this, we simulated environments, where none, 50%, 75%, or 100% of all banks were part of the ecosystem. The results can be seen in Fig.11. Despite the significant difference between the two data sets regarding their data distribution (with IBM-CCF having an equal distribution between banks, while IBM-AML has a highly skewed one), we can clearly see that in both cases, even with only half of the banks being part of the ecosystem (IBMCCF: 2 banks/IBM-AML: 3 banks), a significant performance gain is achieved. Thus, it seems the benefits of the synthetic data ecosystem can be realized from an early stage onwards, making it easy to overcome the hurdle of a minimum number of members needing to participate in the ecosystem, thus tackling another of the challenges outlined previously. Summarizing these results, we were able to demonstrate that the proposed data ecosystem is able to deliver excess performance for all participants in the network on an individual level and it can be seen that even for data ecosystems with only a fraction of the institutions participating in synthetic data sharing, still a significant performance gain can be achieved. Furthermore, there seem to be network effects to some extent where more partners in the ecosystem increase its overall utility. As these results validate the incentives for partners to participate in an ecosystem, we confirm our DP3—Provide the system with a back-testing mechanism in order to ensure newly generated synthetic data matches in composition and fraud detection training performance with real data given that data quality cannot be independently verified. Discussion This research paper is aimed at extending the research on privacy in data ecosystems as well as machine learning of multi-organizational datasets by investigating these challenges in the field of financial fraud detection. This was done by deriving DPs for an innovative synthetic data-sharing ecosystem that allows financial institutions to exchange financial transaction data while protecting client privacy and learning effectively from this multi-institutional data. To create this artifact, we followed the process of DSRM (Peffers etal., 2007), with this paper covering four “designimplement-evaluate” cycles. Starting with the problem identification our study contributes to descriptive knowledge concerning the problem space by identifying data scarcity in combination with the inability to share data due to privacy protection as a major hurdle for financial institutions, validating the existing research on cross-organizational fraud detection collaboration within financial services (Abdul Salam etal., 2024; Kong etal., 2024). During the exploration of the solution, space synthetic data sharing Fig. 10 Performance gain per individual bank and performance gain by institution size Fig. 11 Performance (avg per bank) by percentage of participating institutions Electronic Markets (2025) 35:7 Page 21 of 28 7 was identified as an underexplored solution to tackle data scarcity in financial fraud detection extending the literature on cross-organizational collaboration in the field (Chatterjee etal., 2024). Furthermore, the exploration of synthetic data to allow privacy-compliant data sharing as well as our experimentation on multi-organizational synthetic data during multiple “design-implement-evaluate” cycles reaches beyond financial services and addresses significant challenges in the realm of data ecosystems (Brée etal., 2024). Moreover, our research extends beyond studies that simply outline the requirements of such a data ecosystem (Immonen etal., 2014). We validate these requirements and the derived DPs through rigorous experimentation on publicly available datasets and through close collaboration with industry partners and experts, ensuring the practical applicability and robustness of our findings. Furthermore, by extending data ecosystem research into a less frequently explored domain (Cappiello etal., 2020), we are able to validate the applicability of existing knowledge and uncover new insights with potential for generalization. We achieve this by developing prescriptive knowledge and nascent theory concerning the solution space, offering a set of DPs for designing a synthetic data sharing ecosystem and providing a first instantiation in the form of a platform architecture. To provide more detailed insights into this solution space, additional key findings are encapsulated in Table9, clustered by key areas which we deductively derived a posteriori from our study. As shown in Table9 under the generation dimension, we contribute to the literature on synthetic data generation in multiple ways. First, we identified the necessity for a strictly separated local layer (where real data is transformed) and a global layer (where data is shared). Second, we transfer existing algorithms to a new setup including cross-organizational data with a complex data structure and compare their performance on a prediction task (Pathare etal., 2023) identifying TVAE as the most performant algorithm for synthetic financial data generation while still showing sufficient privacy. Third, we extend the research on the generation setup by consolidating different training schemes from multiple sources (Eilertsen etal., 2021; Fan etal., 2022; Kiran & Kumar, 2024) and comparing them to each other, identifying training on data sub-clusters as the most beneficial setup. Moving forward to training models based on synthetic data, as shown in Table9 under the prediction dimension, we extend the literature which often looks at synthetic data generation performance separately but provides little guidance on how the generated data is best used in a data ecosystem (Dankar etal., 2022). Our research further shows that a mixture of synthetic and real data is most useful when combined; however, the exact mix-in percentage is highly organization and context-specific. Moreover, we demonstrated that using purely synthetic data can still be beneficial for players with no access to real data; however, adjustments need to be made to the composition of the data by artificially rebalancing it. As can be seen in Table9’s ecosystem dimension, our research investigates the complexities of data ecosystems, analyzing how the incentives for participation affect performance outcomes across various sizes of institutions. This analysis also places our findings in the context of the research by Gelhaar and Otto (2020) about the initial challenges encountered within data ecosystems. By implementing design interventions that clearly articulate performance benefits and facilitate the integration of external collaborators, our research substantiates the ecosystem’s capacity to overcome these early hurdles. Further, our empirical evidence suggests that even partial participation in the data ecosystem can lead to substantial performance improvements, thereby affirming the ecosystem’s operational feasibility and enhancing its attractiveness to potential participants. In the last dimension in Table9, we demonstrate the generalizability of our derived design knowledge beyond a single use case and application area. This was done in two ways. First, through validation with experts from the field in academia and the private sector. Second, through performance evaluation in two financial services domains and three datasets which required data sharing with privacy restrictions. While performance gains might seem insignificant, small changes in fraud detection rate can have major implications on financial institutions (Levi, 1998). Thus, our research not only confirms the relevance of our DPs and system architecture but also sets the stage for their application beyond the immediate context of financial transactions, suggesting a blueprint for extending beyond financial services to other domains where data needs to be shared with privacy restrictions (Susha etal., 2019). For practitioners, our contribution is two-fold: For managers and decision-makers, we demonstrate the value of synthetic data-sharing ecosystems that allow both large and small institutions to securely collaborate on data while ensuring privacy. This approach is particularly relevant in industries with complex, highly sensitive data, such as financial services, where data ecosystems do not emerge organically and require careful planning allowing for shared value propositions and services (Adner, 2017; Immonen etal., 2014). Furthermore, our framework addresses regulatory requirements on data privacy and our results suggest a robust foundation for scaling and sustaining privacy-focused data ecosystems. For system architects, we outline a set of DPs that guide practitioners in structuring the architecture of these ecosystems. These principles assist in selecting suitable synthetic data generation methods, implementing mechanisms for data quality assurance, and integrating data to enhance AI model performance. By focusing on these core areas, our contribution provides architects with actionable guidance toward building secure and resilient synthetic Electronic Markets (2025) 35:77 Page 22 of 28 Table 9 Summary of insights generated in the four design-implement-evaluate cycles * Performance evaluation based on ROC AUC scores Area of insight Focus Activities Reasoning Generation This dimension identifies the most suitable synthetic data generation algorithm and the optimal setup for generating synthetic data to enhance performance - Performance comparison of popular synthetic data generation algorithms on real-world financial transaction data (outperformance of TVAE by 49.8%* compared to next best model) - Evaluation of the optimal training setup when generating synthetic data (outperformance of pre-trained class-separated models by 2.0%* compared to next best training scheme) The analyses lead to the selection of a modular, algorithm-agnostic approach that clearly separates between local and global layers. Further, optimal data integration strategies were identified broadening the ecosystem’s applicability and enhancing its performance potential Prediction This dimension focuses on the specifics of training setups for synthetic data generation, especially in contexts with imbalanced classes - Evaluation of the optimal mix-in percentage between real and synthetic data on the institutional level - Performance evaluation of purely synthetic data using different degrees of oversampling (outperformance of data with 10% oversampling by 7.5%* compared to not oversampled baseline) These investigations facilitate the identification of tailored training approaches that accommodate the specific needs of the ecosystem’s participants, ensuring the generation of high-quality, utility-maximizing prediction models Ecosystem This dimension assesses the ecosystem’s balance in terms of participation incentives across institutions of varying sizes and the effect of partial participation - Analysis of performance gain per financial institution and the relationship between institution size and performance gain (negative correlation of − 0.09) - Evaluation of the ecosystem’s performance with varying numbers of participating institutions (significant performance gain with all ecosystem sizes) The findings indicate a proportional benefit across the ecosystem, highlighting particular advantages for smaller institutions. Furthermore, even partial ecosystem participation yields substantial performance improvements. Thus, ensuring the system's viability and attractiveness Generalizability The dimension of generalizability aims to validate the created design knowledge and system architecture across diverse contexts and experts’ perspectives - Semi-structured interviews with experts from financial institutions and academic experts to validate the created DPs and system architecture - Transfer of the synthetic data sharing setup to new contexts like fraud detection (improved performance by 3.6%*) and money laundering detection (improved performance by 6.6%*) using simulation-based datasets The evaluation confirmed the relevance and applicability of the design knowledge across multiple financial services domains Electronic Markets (2025) 35:7 Page 23 of 28 7 data-sharing ecosystems. This framework, therefore, serves as a blueprint for future system designers working within regulated environments where data privacy and AI performance are essential. Limitations and future research opportunities can be identified across our four key areas of insight. Regarding data generation, the current study was constrained by the available data, which prevented the consideration of advanced graph-based synthetic data generation methods such as TransGAN (X. Wang & Yang, 2024). Additionally, while privacy was tested, it was not fully guaranteed by the models used, highlighting the need for future research on the effectiveness of differentially private synthetic data generation methods such as PATEGAN (Jordon etal., 2018) in a synthetic data ecosystem. From a prediction standpoint, further investigation is required to determine how models can be aligned when data schemas—and thus the synthetic data—differ between institutions. Moreover, the design of an effective back-testing mechanism to ensure the ecosystem’s predictive performance should be explored. On the ecosystem level, additional research is necessary to explore ecosystem usage incentives, building on the work by (Gelhaar etal., 2021), which was beyond the scope of this paper. Finally, while this study was limited to financial services due to resource constraints, future research should explore the applicability of the defined DPs beyond this domain, testing their general applicability. Conclusion Based on the need for increased data availability to foster economic growth, this paper provides the design and evaluation of a synthetic data-sharing ecosystem for financial institutions under privacy constraints. The main contribution lies in providing guidance on how to train models based on shared data. By formulating a set of DPs, practical insights, and prototype testing, iterative design cycles were used to provide a robust framework for constructing a data ecosystem that leverages synthetic data. Each DP, from ensuring data quality and enhancing adaptability through transformation and resampling to fostering trust among ecosystem participants and facilitating regulatory access to synthetic data, extends existing research on synthetic data sharing and generation, particularly in the context of financial transaction data. For practice, our example instantiation and codebase can be used as a reference architecture for future instantiations. We not only address the identified need for an efficient, privacy-preserving financial data ecosystem but also set a foundation for future exploration in broader domains where data sharing under privacy restrictions is paramount. Thus, this contribution offers guidance for overcoming technical, trust-related, and regulatory challenges in data ecosystems, unlocking the potential for data-driven innovation and future economic development. Acknowledgements The authors express their gratitude to the UnionBank of the Philippines for their valuable collaboration and the provision of key insights that contributed to this research.This research project was funded by the St. Gallen Symposium and the German Federal Ministry of Education and Research (BMBF) within the “Innovations for Tomorrow’sProduction, Services, and Work” Program (funding number 02K23A001) which is managed by the Project Management Agency Karlsruhe (PTKA). The authors are responsible for thecontent of this publication. Funding Open access funding provided by University of St. Gallen. Data Availability The datasets used during the current study are available in the Kaggle repositories: https:// www. kaggle. com/c/ ieeefrauddetec tion,https:// www. kaggle. com/ datas ets/ ealtm an2019/ ibmtrans actio nsforantimoneylaund eringaml, https:// www. kaggle. com/ datas ets/ ealtm an2019/ creditcardtrans actio ns. Declarations Competing Interests The authors declare that they have no conflict of interest. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. References Abbas, A. E., Agahari, W., van de Ven, M., Zuiderwijk, A., & de Reuver, M. (2021). Business data sharing through data marketplaces: A systematic literature review. Journal of Theoretical and Applied Electronic Commerce Research, 16(7), 7. https:// doi. org/ 10. 3390/ jtaer 16070 180 Abdul Salam, M., Fouad, K. M., Elbably, D. L., & Elsayed, S. M. (2024). Federated learning model for credit card fraud detection with data balancing techniques. Neural Computing and Applications, 36(11), 6231–6256. https:// doi. org/ 10. 1007/ s0052102309410-2 Adner, R. (2017). Ecosystem as structure: An actionable construct for strategy. Journal of Management, 43(1), 39–58. https:// doi. org/ 10. 1177/ 01492 06316 678451 Al-Hashedi, K. G., & Magalingam, P. (2021). Financial fraud detection applying data mining techniques: A comprehensive review from 2009 to 2019. Computer Science Review, 40, 100402. https:// doi. org/ 10. 1016/j. cosrev. 2021. 100402 Altman, E., Blanuša, J., von Niederhäusern, L., Egressy, B., Anghel, A., & Atasu, K. (2024). Realistic synthetic financial transactions for anti-money laundering models (No. arXiv: 2306. 16424). arXiv. https:// doi. org/ 10. 48550/ arXiv. 2306. 16424 Electronic Markets (2025) 35:77 Page 24 of 28 Altman, E. R. (2019). Synthesizing credit card transactions (No. arXiv: 1910. 03033). arXiv. https:// doi. org/ 10. 48550/ arXiv. 1910. 03033 Asrow, K. (2021). The role of individuals in the data ecosystem: Current debates and considerations for individual data protection and data rights in the U.S. Federal Reserve Bank of San Francisco. https:// priva cysec urity acade my. com/ wpconte nt/ uploa ds/ 2021/ 05/ TheRoleofIndiv idualsintheDat aEcosy stem. pdf. Accessed 9 Mar 2023. Assefa, S. (2020). Generating synthetic data in finance: Opportunities, challenges and pitfalls. SSRN Electronic Journal. https:// doi. org/ 10. 2139/ ssrn. 36342 35 Aurna, N. F., Hossain, M. D., Taenaka, Y., & Kadobayashi, Y. (2023). Federated learning-based credit card fraud detection: Performance analysis with sampling methods and deep learning algorithms. 2023 IEEE International Conference on Cyber Security and Resilience (CSR), 2023, 180–186. https:// doi. org/ 10. 1109/ CSR57 506. 2023. 10224 978 Autio, E., spsampsps Thomas, L. D. W. (2014). Innovation ecosystems: Implications for innovation management? In M. Dodgson, D. M. Gann, spsampsps N. Phillips (Eds.), The Oxford Handbook of Innovation Management (p. 0). Oxford University Press. https:// doi. org/ 10. 1093/ oxfor dhb/ 97801 99694 945. 013. 012 Awosika, T., Shukla, R. M., & Pranggono, B. (2024). Transparency and privacy: The role of explainable AI and federated learning in financial fraud detection. IEEE Access, 12, 64551–64560. https:// doi. org/ 10. 1109/ ACCESS. 2024. 33945 28. IEEE Access. Baabdullah, T., Alzahrani, A., Rawat, D. B., & Liu, C. (2024). Efficiency of federated learning and blockchain in preserving privacy and enhancing the performance of credit card fraud detection (CCFD) systems. Future Internet, 16(6), 6. https:// doi. org/ 10. 3390/ fi160 60196 Bagad, P., Mitra, S., Dhamnani, S., Sinha, A. R., Gautam, R., & Khanna, H. (2021). Data-sharing economy: Value-addition from data meets privacy. Proceedings of the 14th ACM International Conference on Web Search and Data Mining, 1105–1108. https:// doi. org/ 10. 1145/ 34379 63. 34417 12 Baskerville, R., Baiyere, A., Gregor, S., Hevner, A., & Rossi, M. (2018). Design science research contributions: Finding a balance between artifact and theory. Journal of the Association for Information Systems, 19(5). https:// aisel. aisnet. org/ jais/ vol19/ iss5/3. Accessed 13 June 2023. Bauer, A., Trapp, S., Stenger, M., Leppich, R., Kounev, S., Leznik, M., Chard, K., & Foster, I. (2024). Comprehensive exploration of synthetic data generation: Asurvey (No. arXiv: 2401. 02524). arXiv. https:// arxiv. org/ abs/ 2401. 02524. Accessed 15 Aug 2024. Benchaji, I., Douzi, S., & Ouahidi, B. E. (2021). Credit card fraud detection model based on LSTM recurrent neural networks. Journal of Advances in Information Technology, 12(2), 113–118. https:// doi. org/ 10. 12720/ jait. 12.2. 113118 Bian, K., & Zheng, H. (2023). FedAvg-DWA: A novel algorithm for enhanced fraud detection in federated learning environment. 2023 4th International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE), 13–17. https:// doi. org/ 10. 1109/ ICBAI E59714. 2023. 10281 317 Blake, M., McWaters, J., & Galaski, R. (2019). The next generation of data-sharing in financial services (p. 33) [White Paper]. World Economic Forum. https:// www2. deloi tte. com/ conte nt/ dam/ Deloi tte/ lu/ Docum ents/ finan cialser vi ces/ lunextgener ationdatashari ngingfinan cialservi ces. pdf. Accessed 29 Jan 2023. Bonabeau, E. (2002). Agent-based modeling: Methods and techniques for simulating human systems. Proceedings of the National Academy of Sciences, 99(suppl_3), 7280–7287. https:// doi. org/ 10. 1073/ pnas. 08208 0899 Brée, T., Karger, E., & Ahlemann, F. (2024). Shaping the future of data ecosystem research—What is still missing? IEEE Access, 12, 103162–103175. IEEE Access. https:// doi. org/ 10. 1109/ ACCESS. 2024. 34329 69 Brodsky, L., & Oakes, L. (2017). Data sharing and open banking. McKinsey. https:// www. mckin sey. com/ ~/ media/ McKin sey/ Indus tries/ Finan cial% 20Ser vices/ Our% 20Ins ights/ Data% 20sha ring% 20and% 20open% 20ban king/ Datashari ngandopenbanki ng. pdf. Accessed 26 Feb 2024. Bun, M., Desfontaines, D., Dwork, C., Naor, M., Nissim, K., Roth, A., Smith, A., Steinke, T., Ullman, J., & Vadhan, S. (2021). Statistical inference is not a privacy violation. https:// diffe renti alpri v acy. org/ infer enceisnot-apriva cyviola tion/. Accessed 14 Aug 2024. Cappiello, C., Gal, A., Jarke, M., & Rehof, J. (2020). Data ecosystems: Sovereign data exchange among organizations (Dagstuhl Seminar 19391). DROPS-IDN/v2/Document/https:// doi. org/ 10. 4230/ DagRep. 9.9. 66. https:// doi. org/ 10. 4230/ DagRep. 9.9. 66 Chakravarty, S., Demirhan, H., & Baser, F. (2020). Fuzzy regression functions with a noise cluster and the impact of outliers on mainstream machine learning methods in the regression setting. Applied Soft Computing, 96, 106535. https:// doi. org/ 10. 1016/j. asoc. 2020. 106535 Chandra, L., Seidel, S., & Gregor, S. (2015). Prescriptive knowledge in IS research: Conceptualizing design principles in terms of materiality, action, and boundary conditions. 2015 48th Hawaii International Conference on System Sciences, 4039–4048. https:// doi. org/ 10. 1109/ HICSS. 2015. 485 Charitou, C., Dragicevic, S., & Garcez, A. d’Avila. (2021). Synthetic data generation for fraud detection using GANs (No. arXiv: 2109. 12546). arXiv. http:// arxiv. org/ abs/ 2109. 12546. Accessed 11 Mar 2024. Chatterjee, P., Das, D., & Rawat, D. B. (2024). Digital twin for credit card fraud detection: Opportunities, challenges, and fraud detection advancements. Future Generation Computer Systems, 158, 410–426. https:// doi. org/ 10. 1016/j. future. 2024. 04. 057 Chen, Z., Van Khoa, L. D., Teoh, E. N., Nazir, A., Karuppiah, E. K., & Lam, K. S. (2018). Machine learning techniques for anti-money laundering (AML) solutions in suspicious transaction detection: A review. Knowledge and Information Systems, 57(2), 245–285. https:// doi. org/ 10. 1007/ s101150171144-z Chen, H., Grossman, M., Sen, A., & Tsao, S.-F. (2023). Establishing a FAIR, CARE, and efficient synthetic health data sharing ecosystem for canada establishing a FAIR, CARE, and efficient synthetic health data sharing ecosystem for Canada. IARIW-CIGI Conference on the Valuation of Data. https:// www. resea rchga te. net/ publi cation/ 37544 6378_ Estab lishi ng_a_ FAIR_ CARE_ and_ Effic ient_ Synth etic_ Health_ Dat a_ Shari ng_ Ecosy stem_ for_ Canada_ Estab lishi ng_a_ FAIR_ CARE_ and_ Effic ient_ Synth etic_ Health_ Data_ Shari ng_ Ecosy stem_ for_ Canada.Accessed 17 Dec 2024 Choi, E., Biswal, S., Malin, B., Duke, J., Stewart, W. F., & Sun, J. (2018). Generating multi-label discrete patient records using generative adversarial networks (No. arXiv: 1703. 06490). arXiv. https:// doi. org/ 10. 48550/ arXiv. 1703. 06490 Cosma, S., Cosma, S., & Pennetta, D. (2023). The rise of financial services ecosystems: Towards open banking platforms. In T. Walker, E. Nikbakht, & M. Kooli (Eds.), The Fintech Disruption: How Financial Innovation Is Transforming the Banking Industry (pp. 191–213). Springer International Publishing. https:// doi. org/ 10. 1007/ 978-303123069-1_8 Dahmen, J., & Cook, D. (2019). SynSys: A synthetic data generation system for healthcare applications. Sensors, 19(5), 5. https:// doi. org/ 10. 3390/ s1905 1181 Dankar, F. K., Ibrahim, M. K., & Ismail, L. (2022). A multi-dimensional evaluation of synthetic data generators. IEEE Access, 10,