Open‐Source Data‐Driven Prediction of Environmental, Social, and Governance (ESG) Ratings Using Deep Learning Techniques

The evaluation of ESG ratings by ESG rating agencies is time‐consuming and requires the participation of numerous human specialists. In this paper, we propose a method for creating proxies of ESG scores by collecting corporate ESG news and publicly available ESG‐related data using data crawling techniques and deep learning‐based classification technology while minimizing human involvement. To validate the effectiveness of the proposed approach, we suggest three hypotheses. Two of them are related to the connection between open‐source information and ESG ratings, while one concerns the link between proxy ESG rating and firm performance. To validate the effectiveness of the proposed approach, we conduct an empirical analysis based on 976 unique companies listed by the Korean Corporate Governance Agency (KCGS) from 2016 to 2019. Initially, we gather ESG indicators from open sources including disclosures and firms' news articles from a news portal site. We utilize Bidirectional Encoder Representations from Transformers (BERT) to classify news articles into environment, social, and governance categories and determine their sentiments. We confirm that ESG news sentiment and variables extracted from open‐source data are related to ESG ratings. Furthermore, we find a significantly positive relationship between E, S, and G ratings predicted based on open‐source data and Tobin's Q.