Page 484 - ISC PROCEEDINGS 21.4

Metric or Description References
metric family
and path (signatures) important for time series realism al., 2024)
properties
Downstream Utility of synthetic data measured by task metrics (Xia, H. et al.,
task (forecast error, classifier accuracy, trading 2024)
performance strategy returns) when models are trained or
tested on generated data
GAN-specific Training and sample quality assessed via (Jin, Z. C. et al.,
divergences adversarial losses or Wasserstein/GAN distances; 2024)
and WGAN variants also used for stability and as
discriminator diagnostic criteria
signals
Tail risk and Evaluates whether generated scenarios
regulatory reproduce extreme losses and tail behavior used (Jin, Z. C. et al.,
measures (VaR, in risk management (e.g., Value at Risk) 2024)
tail loss)
Classification Standard supervised metrics (AUROC) are used (Labiad, B. et al.,
and ranking when GANs generate or augment labeled 2024)
metrics for anomaly/fraud datasets, and performance is
anomaly evaluated by classifiers
detection
Similarity and Global similarity or scenario-matching metrics (Allen, D. E. et
scenario-level (including signature-based similarity) for high- al., 2024)
scores dimensional market scenario comparison
Source: Authors' compilation based on the studies reviewed in this paper.
5.3. Case illustration: Market-GAN benchmark
To make the proposed benchmarking framework concrete, we present a brief case
illustration based on Market-GAN, introduced by Xia et al. (2024), which adds control to
financial market data generation through semantic context. In this example, the model is
used as a representative conditional generator for financial time-series synthesis and is
evaluated along several dimensions. First, statistical fidelity is examined by comparing key
distributional properties of the synthetic and real data, such as central tendency,
dispersion, and tail behavior. Second, temporal realism is assessed by checking whether
the generated series preserve important market dynamics, including dependence
structure and realistic sequential patterns. Third, control and conditional consistency are
evaluated by determining whether the generated samples respond appropriately to the
intended semantic context, which is especially important in market simulation tasks.
Finally, practical utility is measured by testing whether the synthetic data can support
downstream financial analysis, risk assessment, and model development. This illustration
shows that Financial GANs should be judged not only by visual or marginal similarity, but
also by their ability to generate controllable, realistic, and useful market data for real
financial applications.
6. Discussion: trade-offs & policy recommendations
6.1. For regulators and supervisory authorities
Regulators should move beyond general encouragement of synthetic data and
define concrete compliance requirements for its use in financial workflows. In particular,

483