Scholarlink

<p>Accurate forecasting of Environmental, Social, and Governance (ESG) performance has emerged as a critical imperative for quantifying long-term investment risks, yet traditional rating methodologies remain constrained by subjective biases and signal indeterminacy. While machine learning offers a data-driven paradigm for extracting predictive sustainability signals, existing approaches struggle to reconcile the trade-off between forecasting accuracy and model interpretability, particularly when processing the high-dimensional, heterogeneous feature spaces inherent to ESG datasets. To overcome these methodological barriers, this study proposes the Feature-selective Adaptive Cascade Forest (FACForest), a novel interpretable ensemble architecture designed to enhance the predictive reliability of objective ESG assessments. By integrating a model-driven feature selection mechanism within a deep cascade structure, the framework systematically eliminates redundant noise variables through iterative importance ranking, thereby effectively mitigating the curse of dimensionality that plagues conventional forecasting models. Furthermore, the architecture employs progressive residual learning across successive layers to capture complex nonlinear interactions among sustainability indicators, generating transparent decision paths that explicitly elucidate the propagation of feature influence.</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

Feature-Selective Adaptive Cascade Forest for Interpretable High-Dimensional ESG Rating Forecasting

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Yunduo Han,
  • Wanan Liu,
  • Xingyu Lan,
  • Renbang Shan,
  • Baoli Shi,
  • Yunchen Wang,
  • Congyuan Pang,
  • Naixi Chen

摘要

Accurate forecasting of Environmental, Social, and Governance (ESG) performance has emerged as a critical imperative for quantifying long-term investment risks, yet traditional rating methodologies remain constrained by subjective biases and signal indeterminacy. While machine learning offers a data-driven paradigm for extracting predictive sustainability signals, existing approaches struggle to reconcile the trade-off between forecasting accuracy and model interpretability, particularly when processing the high-dimensional, heterogeneous feature spaces inherent to ESG datasets. To overcome these methodological barriers, this study proposes the Feature-selective Adaptive Cascade Forest (FACForest), a novel interpretable ensemble architecture designed to enhance the predictive reliability of objective ESG assessments. By integrating a model-driven feature selection mechanism within a deep cascade structure, the framework systematically eliminates redundant noise variables through iterative importance ranking, thereby effectively mitigating the curse of dimensionality that plagues conventional forecasting models. Furthermore, the architecture employs progressive residual learning across successive layers to capture complex nonlinear interactions among sustainability indicators, generating transparent decision paths that explicitly elucidate the propagation of feature influence.