Synthetic data

June 2014: Ian Goodfellow and seven coauthors, including Yoshua Bengio, published "Generative Adversarial Nets." Two networks compete. One produces examples, the other tries to tell them from real ones, and the first ends up mimicking the distribution of the real data. Synthetic data comes out of algorithms, typically generative models such as GANs, VAEs or, more recently, diffusion models and LLMs. It mimics the statistical distributions of a real dataset without copying individual records. Actuaries use it for two things. It extends datasets too small to train models that generalize, particularly for rare sub-populations such as severe claims. And it offers a path to GDPR compliance, since the link to identifiable individuals is gone. Insurers also use it to simulate catastrophic claim scenarios that historical data does not cover, notably systemic cyber events. The main difficulty is validation. Synthetic data can perfectly preserve the marginal distribution of each variable and still miss extreme correlations, and tail risk then gets underestimated. Fidelity, diversity and privacy rank high in emerging standards for evaluating synthetic data. The link with AI red teaming is direct: adversarial synthetic data can test how a model holds up on distributions outside its training range.