Tail-GAN: Learning to Simulate Tail Risk Scenarios.

The article focuses on Tail-GAN, a novel data-driven generative adversarial network (GAN) framework designed to simulate realistic, high-dimensional multiasset financial scenarios that accurately capture tail risk features for a broad class of static and dynamic trading strategies. Leveraging the joint elicitability property of Value-at-Risk (VaR) and Expected Shortfall (ES), Tail-GAN trains a discriminator to evaluate simulated scenarios based on these tail risk measures, guiding the generator to produce scenarios that preserve critical tail risk characteristics. Extensive numerical experiments on synthetic and real market data demonstrate that Tail-GAN outperforms existing methods—including variants trained only on raw returns or static portfolios and Wasserstein GANs—in accurately estimating tail risks, capturing temporal and cross-asset dependence, and generalizing to unseen data. The framework also scales effectively to large-dimensional portfolios by incorporating principal component analysis through eigenportfolios, making it suitable for practical financial risk management applications such as intraday market scenario simulation.