Page 480 - ISC PROCEEDINGS 21.4

foundational elements that support repeated use, coordination, and scalable innovation
across an entire system. Synthetic financial data can be interpreted as part of this
infrastructure because it provides reusable datasets that support experimentation, model
development, regulatory testing, and cross-institutional collaboration without exposing
raw confidential information. In this sense, Financial GANs do not only produce artificial
records; they help build a digital layer that enables safer and broader access to financial
data for multiple stakeholders. This infrastructure role also implies that synthetic data
should be evaluated not only by statistical similarity, but also by its reproducibility,
interoperability, documentation, and ability to integrate into institutional workflows. Thus,
GAN-based synthetic financial data contributes to the development of a more flexible and
scalable digital financial infrastructure.
3. Research methodology
This paper adopts a structured literature review and conceptual synthesis approach
to examine the role of Generative Adversarial Networks in financial data generation and
to assess their relevance as a component of digital infrastructure. Unlike empirical studies
that train and test a single model on a fixed dataset, the present work surveys the main
Financial GAN families, compares their technical characteristics, and links their modeling
strengths to economic and institutional needs such as privacy protection, supervised data
sharing, and infrastructure readiness. The review covers representative studies on
financial time-series generation, transaction synthesis, conditional generation, and
privacy-aware synthetic data, including work on recurrent and autoregressive GANs,
conditional GANs, transformer-based approaches, and differentially private variants such
as those discussed by Takahashi et al. (2019), Wiese et al. (2020), Jin et al. (2024), Labiad
et al. (2024), Li et al. (2022), Mehri et al. (2024), Ramachandra and Vaithiyanathan (2025),
Xia et al. (2024), and Zhang et al. (2025).
To organize the literature, the reviewed studies are classified according to the type
of financial data they address, the architecture of the generator, the stylized facts they
reproduce, and the practical constraints they consider. In particular, the paper
distinguishes between recurrent and autoregressive models, which are typically effective
for short-range dependence and volatility clustering; conditional and hybrid models,
which better handle regime-dependent behavior and heterogeneous financial contexts;
transformer-based models, which are more suitable for long-range dependence and
cross-series interactions; and privacy-focused variants, which aim to balance utility with
disclosure control. This classification is based on the observation that different Financial
GAN architectures serve different financial tasks, from realistic stock market simulation
and transaction synthesis to conditional default-risk modeling and privacy-preserving data
release.
The analytical framework used in this paper further evaluates the reviewed
methods from the perspective of infrastructure readiness. Rather than relying only on
visual realism or marginal statistical similarity, the paper emphasizes a broader set of
criteria that includes temporal fidelity, downstream utility, tail-risk preservation, privacy
safeguards, reproducibility, and auditability. Accordingly, the evaluation and
benchmarking discussion is built around statistical diagnostics such as moments and
distributional tests, time-series diagnostics such as autocorrelation and pathwise
similarity, task-oriented utility measures such as train-on-synthetic/test-on-real
performance, and risk-sensitive indicators such as VaR and CVaR. This methodology is
aligned with the paper’s central argument that synthetic financial data should be assessed

479