We introduce a network-based methodology, combining the Triangulated Maximally Filtered Graph and node2vec algorithms, to design hedging portfolios for climate risk. Our analysis shows that carbon risk is consistently linked to the Utility sector, enabling effective hedging strategies. In contrast, ESG risk lacks stable sectoral connections, suggesting that ESG scores are too broad for direct market applications. These findings highlight the need to explicitly account for carbon risk in bank capital requirements.

This study replicates LSEG's ESG scoring methodology using machine learning to shed light on the key drivers behind ESG ratings, with a focus on the balance between forward‐looking promises (aspirational) and past achievements (performance). Our analysis finds that approximately 60% of ESG scores are based on aspirational promises, while only approximately 40% reflect actual performance. This imbalance suggests a potential over‐reliance on future commitments, which could inflate ESG scores and mislead investors about a company's true sustainability efforts, making a call for greater transparency in ESG issuance.

ESG scores are widely used to assess companies' sustainability profile, but their computation lacks transparency due to proprietary rating models with limited disclosure. This study applies machine learning techniques to analyze LSEG's ESG rating process, successfully replicating the assessment model with high accuracy while highlighting unlearnable noise in the data. Interpretability tools further identify key factors driving the ratings, providing valuable insights into the opaque rating methodology.