Scholarlink

<p>As the frequency and intensity of catastrophic events escalate, insurance companies face mounting risks of substantial claims, which can potentially lead to financial liquidity challenges. The Extreme Value Index (EVI) emerges as a critical indicator, offering essential foundations for rational catastrophe risk management. This study proposes a novel attention-stacking ensemble model that adaptively integrates traditional statistical models and machine learning algorithms as base learners through an attention mechanism. This innovative approach combines tornado intensity and spatiotemporal data to achieve precise EVI estimation, leveraging the predictive strengths of different base models and adaptively adjusting model weights to accommodate diverse catastrophic risk scenarios. Gamma deviance is selected as the objective function, making it particularly suitable for heavy-tailed catastrophic losses. Regarding the tornado disaster in the U.S., the attention-stacking framework demonstrates a 10.24% reduction in EVI prediction deviance compared to XGBoost, showing substantial optimization in extreme risk assessment with error percentages of 2.22% for <i>VaR</i><sub>0.90</sub> and 8.60% for <i>ES</i><sub>0.90</sub>. Furthermore, the predicted EVI accurately reflects tornado loss conditions across various states and enables the assessment and comparison of their extreme catastrophe risks. Overall, this stacking model achieves optimal EVI prediction for both national and state-level applications, underscoring its stability in evaluating tornado catastrophe risks.</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

Extreme Value Index Estimation for Catastrophe Risk Evaluation Using Stacking Ensemble Model

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Xin Yang

摘要

As the frequency and intensity of catastrophic events escalate, insurance companies face mounting risks of substantial claims, which can potentially lead to financial liquidity challenges. The Extreme Value Index (EVI) emerges as a critical indicator, offering essential foundations for rational catastrophe risk management. This study proposes a novel attention-stacking ensemble model that adaptively integrates traditional statistical models and machine learning algorithms as base learners through an attention mechanism. This innovative approach combines tornado intensity and spatiotemporal data to achieve precise EVI estimation, leveraging the predictive strengths of different base models and adaptively adjusting model weights to accommodate diverse catastrophic risk scenarios. Gamma deviance is selected as the objective function, making it particularly suitable for heavy-tailed catastrophic losses. Regarding the tornado disaster in the U.S., the attention-stacking framework demonstrates a 10.24% reduction in EVI prediction deviance compared to XGBoost, showing substantial optimization in extreme risk assessment with error percentages of 2.22% for VaR0.90 and 8.60% for ES0.90. Furthermore, the predicted EVI accurately reflects tornado loss conditions across various states and enables the assessment and comparison of their extreme catastrophe risks. Overall, this stacking model achieves optimal EVI prediction for both national and state-level applications, underscoring its stability in evaluating tornado catastrophe risks.