SwiGAN model uses GANs to generate climate scenarios for insurance risk

As natural catastrophe costs have surged from $70-80B annually (1970-2000) to $180-200B (2001-2020), insurers face pressure to extend their planning horizons beyond traditional one-year regulatory frameworks like Solvency II. In response, researchers from BioSP and CREST have introduced SwiGAN, a Conditional Generative Adversarial Network (GAN) built on a Wasserstein loss function. The model generates realistic spatio-temporal trajectories of climatic indices—specifically the Soil Wetness Index (SWI), a key drought indicator in France. Drought accounts for roughly 30% of all indemnities paid under the French natural catastrophe insurance scheme. By learning from historical data and conditioning on future climate scenarios, SwiGAN produces plausible sequences of SWI maps up to 2050 for a particularly exposed region in France.

Beyond the specific drought application, the methodology is designed to be generalizable to other climate-related perils (e.g., floods, wildfires) and other actuarial tasks such as economic scenario generation. The use of a Wasserstein GAN improves training stability and generates more realistic spatial patterns compared to traditional GANs. This allows insurers to simulate a wide range of possible futures, stress-test their portfolios, and develop adaptive risk management strategies that go beyond short-term regulatory requirements. The paper, available on arXiv, positions SwiGAN as a practical tool for the insurance industry to better quantify and price the accelerating risks of climate change.