Scholarlink

<p>Volatility modelling is a crucial aspect of financial risk management and asset pricing. Traditional approaches rely on historical data, which can be limited in capturing extreme events and future market dynamics. This paper presents a novel approach for generating synthetic stock data using Conditional Generative Adversarial Networks (CGANs) to enhance volatility modelling and risk assessment, comparing its effectiveness against established Extreme Value Theory (EVT) methods. This is the first comprehensive study examining synthetic data generation across 39 Global Systemically Important Financial Institutions (G-SIFIs) spanning US, UK, and European markets from 2018 to 2023. Our empirical analysis demonstrates that while CGANs successfully replicate broad market trends with 95% accuracy in normal conditions, they show varying performance in capturing extreme events, with Peaks over Threshold (POT) models outperforming in tail risk estimation. The CGAN model achieved correlation coefficients ranging from 0.59 to 0.62 across markets, with the UK market showing the strongest alignment between synthetic and original data. Value at Risk (VaR) and Conditional Value at Risk (CVaR) analyses reveal the model's tendency to overestimate risk metrics by 15–20%, suggesting its potential utility for conservative risk assessment. The synthetic data maintained temporal consistency whilst successfully replicating major global events, including the COVID-19 market crash, with a maximum deviation of 8% from actual price movements. These findings have significant implications for regulatory stress testing, risk management, and portfolio optimisation, particularly in markets where historical data is limited or scenarios need to be augmented for robust risk assessment.</p>

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

Advances in Volatility Modelling; The Costs and Benefits of Synthetic Data for Global Banks

  • 期刊论文

  • 发表日期：

    2025年5月29日

查看全文
  • Mohamad Hassan

摘要

Volatility modelling is a crucial aspect of financial risk management and asset pricing. Traditional approaches rely on historical data, which can be limited in capturing extreme events and future market dynamics. This paper presents a novel approach for generating synthetic stock data using Conditional Generative Adversarial Networks (CGANs) to enhance volatility modelling and risk assessment, comparing its effectiveness against established Extreme Value Theory (EVT) methods. This is the first comprehensive study examining synthetic data generation across 39 Global Systemically Important Financial Institutions (G-SIFIs) spanning US, UK, and European markets from 2018 to 2023. Our empirical analysis demonstrates that while CGANs successfully replicate broad market trends with 95% accuracy in normal conditions, they show varying performance in capturing extreme events, with Peaks over Threshold (POT) models outperforming in tail risk estimation. The CGAN model achieved correlation coefficients ranging from 0.59 to 0.62 across markets, with the UK market showing the strongest alignment between synthetic and original data. Value at Risk (VaR) and Conditional Value at Risk (CVaR) analyses reveal the model's tendency to overestimate risk metrics by 15–20%, suggesting its potential utility for conservative risk assessment. The synthetic data maintained temporal consistency whilst successfully replicating major global events, including the COVID-19 market crash, with a maximum deviation of 8% from actual price movements. These findings have significant implications for regulatory stress testing, risk management, and portfolio optimisation, particularly in markets where historical data is limited or scenarios need to be augmented for robust risk assessment.