Preparing Scope 3 Carbon Emission Disclosure: A Machine Learning Approach

This study employs machine learning models to facilitate relatively low-cost disclosure of Scope 3 carbon emissions, utilizing novel Korean data. Scope 3 CO2 emissions are a significant concern for companies, particularly after the ISSB and EFRAG have implemented mandatory reporting standards. Currently, accurately determining these values is difficult, with many estimates being necessary, leading to substantial costs. Consequently, this presents a substantial challenge for individual companies to manage independently. In line with this, in Korea, the Financial Services Commission is planning to introduce climate change disclosure required by the ISSB, focusing on companies with assets of more than $2 trillion, but the most important part is how to accurately and efficiently measure and disclose carbon dioxide emissions. Through evaluating five models - Random Forest, Gradient Boosting, Adaboost, XGBoost, and LightGBM, we identify LightGBM as the most accurate for Korean companies in Scope 3 carbon emissions, with a 77.01% accuracy based on R-square. Furthermore, based on our research model, the estimation results for Scope 1 and Scope 2 showed prediction accuracies of 84% and 88%, respectively. The result of this paper offers empirical insights for future regulatory ESG disclosures, showcasing the study’s academic and practical contributions to enhancing Scope 3 emissions estimation and expanding the ESG research domain. Specifically, the contribution of this paper is to demonstrate that machine learning methodologies can be employed by stakeholders—such as accounting firms to verify companies’ disclosed Scope 3 emissions, supervisory authorities to enhance regulatory oversight, and investors to identify companies with lower Scope 3 emissions.