Predictors of SDG Disclosures in Chinese Public Companies: A Machine Learning Approach

Abstract

The United Nations’ call for companies to adopt SDGs presents a challenge to report on SDG performance to external stakeholders. Despite emerging discussion by standard setters and within the literature, there remains ambiguity as to the nature and extent of SDG disclosures; hence, a core challenge for future corporate SDG disclosures is identifying quantifiable and consistent measures for performance. This is particularly prevalent in China, where the government embraced the UN 2030 Agenda for Sustainable Development within the 13th National Five-Year Plan, through which they have encouraged companies to become more proactive in SDG. Therefore, this thesis focuses on the growing use of SDG disclosures in Chinese companies and the factors that may predict their SDG disclosure levels. To address this challenge, this thesis, drawing upon existing reporting standards, academic literature and Chinese government publications, has developed a content analysis tool of 27 indicators to capture Chinese corporate SDG disclosures. A sample of 60 public companies were examined during the period of the 13th National FYP (2016-2020) to provide analysis of Chinese corporate SDG disclosures, particularly on SDG 1, 7 and 13. A review of existing literature suggests that a wide range of factors can affect the level of SDG disclosures; however, these studies rely on traditional statistical methods and overlook the complex non-linear associations among predictors. To address these challenges, this thesis employs the gradient boosting machine to 1. identify the predictive power of 159 predictors from government control, financial performance, corporate governance, and ESG performance aspects; 2. offer a more reliable out-of-sample analysis. This thesis observed that the gradient boosting machine outperforms traditional methods to interpret the prediction results. However, the predictive power and direction of the results widely vary due to variable selection and variable measurements.