Benchmarking Machine Learning Models for ESG Prediction in South Korea Using News-Derived Time Series

Existing ESG ratings have limitations like disclosure delays, inconsistencies, and uneven coverage, particularly in non-English markets. This paper addresses these issues by establishing the first machine learning benchmark for ESG prediction in the Korean market using news-derived time-series features. A standardized dataset of 278 Korean firms was constructed, and monthly sentiment and ESG-relevance features were generated from news using Korean-specific language models. A mask-aware CNN explicitly handles missing data by distinguishing observed months from imputed ones. The model achieved a Mean Absolute Error (MAE) of 17.9, a Root Mean Squared Error (RMSE) of 22.0, an 𝑅2 of 0.12, and a Spearman’s 𝜌 of 0.38, demonstrating that temporal modeling and explicit handling of missing data are crucial for improving predictive accuracy.