Exploring the Impact of ESG Ratings on Corporate Carbon Emissions in Korean Firms: Evidence from Machine Learning and Deep Learning Models

This study examines corporate carbon emissions of Korean firms from an ESG perspective and develops an AI-based screening framework to improve the identification of firms likely to exceed regulatory emission thresholds. As global climate policies and carbon pricing mechanisms expand, understanding the emission profiles of listed companies has become increasingly important for regulators, investors, and policymakers. Despite growing ESG disclosure, reliable firm-level screening tools for carbon emissions remain limited. Using a pooled annual panel of KOSPI-listed non-financial firms from 2019 to 2024, the study constructs a dataset of 552 firm-year observations. Firms are classified as high-emission when annual emissions exceed the Korean Emissions Trading Scheme (K-ETS) regulatory threshold of 125,000 tCO 2 e. To evaluate predictive performance, the analysis compares multiple machine learning models (RF, SVM, XGBoost, LightGBM, and CatBoost) and deep learning models (CNN, RNN, GAN, LSTM, and Transformer). In addition, a hybrid ensemble combining CatBoost, GAN, and Transformer is proposed to enhance predictive reliability. The empirical results show that ESG-augmented models consistently outperform financial-only baselines across AUC and F1 metrics. Among individual models, the ESG-enhanced Transformer achieves the strongest discriminatory power, while the proposed hybrid ensemble delivers the best overall predictive performance. The findings contribute to the literature by demonstrating the incremental value of ESG information in predicting corporate carbon emissions and by presenting a practical AI-based framework for compliance-oriented screening under carbon regulation. From a policy and investment perspective, the model provides a useful decision support tool for anticipating potential inclusion in emissions trading schemes, assessing transition exposure, and supporting data-driven decarbonization strategies.