基于加性提升模型的可解释ESG评级研究

Interpretable ESG Rating Research Based on Additive Boosting Models

摘要: 在全球可持续发展趋势和负责任投资理念日益受到重视的背景下，环境(Environmental)、社会(Social)与治理(Governance)，即ESG，已成为衡量企业长期价值与综合风险的关键维度。但ESG评级体系中固有的分歧性与不透明性，极大地限制了其在可持续投资和企业战略中的实用价值。为解决ESG评级中长期存在的模型性能与可解释性的平衡难题，本研究提出并应用加性提升机(Additive Boosting Machine, ABM)构建一个内在可解释的ESG评级模型。ABM作为一种先进的广义加性模型，其核心优势在于能够在保持模型完全透明度的同时，通过Boosting算法学习复杂的非线性模式。该模型将复杂的ESG评分预测任务分解为一系列透明、可视化的组成部分——即单一特征的非线性影响函数与关键的成对特征交互效应函数。这使得模型的每一个决策路径和影响因素都可以被精确追溯和直观理解。结果表明，ABM在保持可解释性的前提下，在MSE、RMSE、MAE、R²、Gini系数和斯皮尔曼系数六大指标上均优于其他模型。该模型不仅成功识别出关键的非线性规律，还揭示了具有实践指导意义的特征交互作用。因此，本研究不仅为ESG评分领域提供了新的解决方案，也为企业管理者、投资者和监管机构提供清晰、可靠决策依据，有助于推动财务回报与可持续发展目标的统一。

Abstract: Against the backdrop of the global sustainable development and the growing emphasis on responsible investment, Environmental, Social, and Governance (ESG) factors have become critical dimensions for assessing corporate long-term value and comprehensive risk. However, the inherent divergence and opacity within existing ESG rating systems significantly limit their practical utility in sustainable investing and corporate strategy. To address the long-standing challenge of balancing model performance and interpretability in ESG ratings, this study proposes and applies an Additive Boosting Machine (ABM) to construct an intrinsically interpretable ESG rating model. As an advanced type of Generalized Additive Model, the ABM’s core strength lies in its ability to learn complex non-linear patterns through boosting algorithms while maintaining complete model transparency. The model decomposes the complex task of ESG score prediction into a series of transparent, visualizable components—namely, non-linear shape functions for individual features and key pairwise feature interaction effects. This allows every decision path and contributing factor within the model to be precisely traced and intuitively understood. Our results demonstrate that the ABM outperforms other benchmark models across six key metrics—MSE, RMSE, MAE, R², Gini coefficient, and Spearman’s rank correlation coefficient—while retaining interpretability. Furthermore, the model not only identifies crucial non-linear relationships but also reveals feature interactions with significant practical implications. Consequently, this research provides a novel solution for the field of ESG scoring and offers clear, reliable decision-support for corporate managers, investors, and regulators, thereby facilitating the alignment of financial returns with sustainable development goals.