Algorithmically generated data that reproduces the statistical properties of real data without containing information on existing individuals, used to train models or test systems.
Synthetic data is produced by algorithms, typically generative models such as GANs, VAEs or, more recently, diffusion models and LLMs, to mimic the statistical distributions of a real dataset without copying individual records. In actuarial science, their value is twofold: they allow insufficiently large datasets to be extended for training robust models, particularly for rare sub-populations such as severe claims, and they offer a path to GDPR compliance by removing the link to identifiable individuals. In insurance, they are also used to simulate catastrophic claim scenarios that historical data does not cover, notably for the modeling of systemic cyber events. The main difficulty is validation: synthetic data that perfectly preserves the marginal distribution of each variable but fails to capture extreme correlations can lead to underestimation of tail risk. The questions of fidelity, diversity and privacy are central to emerging standards for evaluating synthetic data. The link with AI red teaming is direct: adversarial synthetic data can serve to test the robustness of a model against distributions outside its training range.
A reinsurer wishes to train a cyber pricing model on historical loss data, but its portfolio contains only 300 major losses over ten years. It generates 50,000 synthetic losses using a diffusion model conditioned on its real data, after validating that marginal distributions and tail correlations are correctly reproduced. The resulting model improves calibration on large losses by 22%.
synthetic data, données synthétiques, données générées, données simulées, data augmentation