Diffusion-Based Generative Modeling for Tail Risk Simulation and Systemic Stress Testing: A Deep Learning Framework for Financial Crisis Scenarios

This study develops and validates a conditional diffusion-based generative framework for simulating extreme tail events and conducting systemic stress tests in financial markets. Traditional risk models, including value at risk, conditional value at risk, and GARCH-type specifications, as well as more recent deep generative models such as generative adversarial networks and variational autoencoders, systematically fail to capture the heavy-tailed, high-dimensional, and regime-switching nature of financial crises. To address this gap, we propose a denoising diffusion probabilistic model conditioned on market state variables (volatility regimes, correlation structures, and macroeconomic proxies) and implemented using a time?conditioned U?Net architecture. The model is trained on daily returns of 50 S&P 500 equities and 5 major cryptocurrencies over the period 2018–2025, with out-of-sample testing in July-December 2025 and historical crisis benchmarking against the March 2020 COVID-19 selloff and the March 2023 banking turmoil. Extensive comparisons against historical bootstrap, GJR-GARCH, a recurrent Wasserstein GAN, and an LSTM variational autoencoder are performed using three tiers of metrics: statistical fidelity (maximum mean discrepancy, tail Kolmogorov?Smirnov statistic, tail concentration ratio), financial risk accuracy (absolute percentage errors for VaR and CVaR at 95%, 99%, and 99.9% confidence levels), and systemic risk relevance (ΔCoVaR, SRISK, plausibility, and novelty). The conditional diffusion model outperforms the evaluated baselines across the reported metrics in this experimental setting. At the 99.9% VaR level, the absolute percentage error is 5.4% compared to 18.3% for the best baseline. The tail concentration ratio is 1.02, indicating negligible bias, while the GAN yields 0.67 (severe underestimation). Under a combined stress condition, the diffusion model generates counterfactual scenarios with ΔCoVaR (8.7%) and SRISK (23.4%) that closely match real crisis values (9.0% and 24.0%), with an implausibility rate of only 0.3% and a novelty rate of 92%. Sensitivity analysis confirms robustness across hyperparameter choices. These results indicate that conditional diffusion models can provide improved statistical fidelity and practical value for forward-looking systemic risk assessment, offering a powerful alternative to existing generative methods in financial risk management.