Scholarlink

This study explores the application of generative AI, particularly Denoising Diffusion Probabilistic Model (DDPM), for financial scenario simulation. To address the limitations of existing models, this study introduces Tail-Diffusion, a novel framework that incorporates tail-risk metrics such as Value at Risk (VaR) and Expected Shortfall (ES) to capture extreme market conditions. Tail-Diffusion outperforms standard diffusion models and other generative approaches, like GANs, by generating more realistic and diverse financial data. The evaluation, conducted on both synthetic (GBM-generated) and real-world financial data, shows that Tail-Diffusion effectively replicates statistical characteristics while better aligning with real-world behaviours, such as return distributions and market fluctuations. Despite these advancements, the study highlights areas for improvement, including the need for quantitative evaluation metrics and validation in downstream tasks. This research offers a robust foundation for using generative AI in financial modelling, with significant potential for applications in risk management and strategy development.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Tail-Diffusion: Simulation Scenarios in Finance with Denoising Diffusion Models and Tail Risk Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nguyen Minh Thu,
  • Huynh Hoang Trung Ngia,
  • Tran Hong Huy,
  • Vu Duc Thinh,
  • Ngo Minh Man

摘要

This study explores the application of generative AI, particularly Denoising Diffusion Probabilistic Model (DDPM), for financial scenario simulation. To address the limitations of existing models, this study introduces Tail-Diffusion, a novel framework that incorporates tail-risk metrics such as Value at Risk (VaR) and Expected Shortfall (ES) to capture extreme market conditions. Tail-Diffusion outperforms standard diffusion models and other generative approaches, like GANs, by generating more realistic and diverse financial data. The evaluation, conducted on both synthetic (GBM-generated) and real-world financial data, shows that Tail-Diffusion effectively replicates statistical characteristics while better aligning with real-world behaviours, such as return distributions and market fluctuations. Despite these advancements, the study highlights areas for improvement, including the need for quantitative evaluation metrics and validation in downstream tasks. This research offers a robust foundation for using generative AI in financial modelling, with significant potential for applications in risk management and strategy development.