Tail-GAN: Learning to Simulate Tail Risk Scenarios
GAN-based approach for generating high-dimensional multi-asset price scenarios that preserve the tail risk (Value-at-Risk and Expected Shortfall) of a set of trading strategies, for estimating portfolio loss distributions.
Verdict
Novel data-driven scenario generator built specifically for tail risk. In the authors' experiments it captures VaR and Expected Shortfall better than earlier data-driven generators, including a Wasserstein GAN baseline, by exploiting the joint elicitability of VaR and ES. Training on PCA eigenportfolios lets it scale to more assets (shown on 20 simulated assets). Strong for tail-risk scenario generation and risk-management research.
Practical Use
Built for risk teams that need realistic multi-asset scenarios which get tail risk right, for example when estimating Value-at-Risk and Expected Shortfall, the measures the Basel FRTB rules emphasize. Unlike generic GANs, Tail-GAN is trained to match VaR and ES for a set of benchmark static and dynamic trading strategies, so the generated scenarios stay relevant for risk measurement. Training on PCA eigenportfolios makes it scale to larger asset sets (the paper shows 20 assets). The official Python code (PyTorch, GPL-3.0) maps each script to the paper's tables and figures and can generate synthetic data; the paper's intraday market data (LOBSTER Nasdaq ITCH for AAPL, AMZN, GOOG, JPM and QQQ) must be bought from LOBSTER. The authors ran it on an Nvidia A100 GPU.
Paper
https://arxiv.org/abs/2203.01664Rama Cont, Mihai Cucuringu, Renyuan Xu, Chao Zhang. Management Science, 2026
Code Repository
https://github.com/chaozhang-ox/Tail-GANLicense
Custom. Code GPL-3.0 (github.com/chaozhang-ox/Tail-GAN); paper CC BY-NC-ND 4.0 (arXiv)
Installation
git clone https://github.com/chaozhang-ox/Tail-GAN