Synthetic data

Synthetic data is generated by algorithms to reproduce the statistical patterns of real data without copying real records. It includes synthetic tabular data such as transactions or customer records, rendered images and video for computer vision, simulated sensor data and generated text. Banks, insurers, health researchers, software testers and AI teams use it to share data safely, test systems, balance rare classes and train models where real data is scarce or restricted.

Typical fields

• Synthetic record or item ID
• Generated fields mirroring the source schema
• Generator name and version
• Fidelity metrics compared with the source distribution
• Privacy metrics such as distance to the closest real record
• For rendered data: labels produced automatically by the simulator

Sources and methods

Tabular synthetic data is produced by statistical or generative models trained on a real source dataset, sometimes with differential privacy. Image and video data is rendered from 3D scenes, with labels generated automatically. Quality is measured in three ways: fidelity, meaning how closely distributions and correlations match the source; utility, meaning how well models trained on synthetic data perform on real data; and privacy risk.

Delivery and pricing

Synthetic data is delivered as files or generated on demand through software. Pricing may be per dataset, per row or image, or a platform subscription, with custom generation projects quoted separately.

Before you buy

• Ask for fidelity and utility metrics measured on held-out real data.
• Check privacy testing for memorised or near-copied records.
• Confirm rights to the source data used to train the generator.
• Test on the rare cases that matter to you.
• Review how labels were produced for rendered data.