Synthetic Data in Financial Services: Governance Is the Key to Trust

Synthetic data has long been touted as a breakthrough for innovation in financial services — offering the promise of privacy-preserving experimentation, better model development, and richer testing environments. But as the FCA’s Synthetic Data Expert Group (SDEG) highlights in its second and final report (August 2025), the real challenge is not just how synthetic data is generated, but how it is governed.

Why Synthetic Data Matters

Financial services thrive on data, yet real-world data often comes with barriers: privacy constraints, limited access, or inherent biases. Synthetic data — designed to replicate the statistical properties of real data without exposing sensitive information — can unlock innovation while mitigating these risks.

We’ve already seen it applied to areas such as:

• Fraud detection, by generating realistic rare-event data;

• Credit fairness, by reducing bias in training sets;

• AML systems, through projects like the FCA–Alan Turing Institute collaboration.

The potential is vast — but so are the risks if governance is overlooked.

Governance Foundations

The SDEG stresses that synthetic data is not a “free pass” around compliance. Instead, it must sit within strong governance frameworks. Key foundations include:

• Clear accountability across the lifecycle, especially when third-party vendors are involved;

• Documentation and auditability, maintaining an unbroken chain of evidence on decisions, assumptions, and trade-offs;

• Continuous monitoring, recognising that risks (privacy leakage, bias, model drift) evolve over time.

Managing the Trade-Offs

Synthetic data projects are never neutral. Each choice — from generation method (GANs, agent-based models, distribution-based algorithms) to downstream use — involves balancing fidelity, utility, privacy, and fairness.

• Optimising for fidelity may increase privacy risks.

• Optimising for privacy may reduce model performance.

• Addressing bias may inadvertently introduce new distortions.

The SDEG’s recommendation? Treat these as explicit, documented trade-offs, supported by cross-disciplinary governance forums.

Building Confidence

Ultimately, synthetic data adoption depends on confidence — from regulators, firms, and consumers. That confidence isn’t achieved through technical perfection but through:

• Transparency (why synthetic data is used, how it was generated, safeguards in place),

• Consistency (alignment with existing model risk management and AI ethics frameworks),

• Collaboration (bringing together regulators, industry, academia, and civil society).

As the report concludes: synthetic data quality is contextual. Its value lies not in abstract measures but in whether it is fit for purpose — accurate enough, private enough, and fair enough for the use case it serves.

The FCA’s work through the Synthetic Data Expert Group signals an important shift: synthetic data is moving from concept to practice. But its future in financial services will be shaped not by the sophistication of algorithms alone, but by the robustness of governance frameworks that surround them.