Modeling Firm-Level ESG-Sentiment Interactions in Stock Returns: Evidence from 16 Companies Using Retrofitted Word Embeddings
This study investigates how emotion-specific sentiment embedded in financial news headlines interacts with firm-level Environmental, Social, and Governance (ESG) ratings to influence stock return behavior. Addressing key methodological gaps in existing literature, the analysis leverages Retrofitted Word Embeddings to encode discrete emotional cues tailored to ESG-relevant narratives. Unlike prior studies that rely on lexicon-based or transformer-based models, this approach explicitly incorporates domain-specific emotional semantics while accounting for firm-level heterogeneity and temporal sentiment fluctuations. Using a dataset of 16 multinational firms and sentiment data extracted from Seeking Alpha headlines, the study tests three hypotheses: (1) emotion-specific sentiment independently predicts stock returns; (2) the moderating effect of sentiment varies across ESG dimensions; and (3) positive (negative) sentiment amplifies (dampens) ESG performance effects. The analysis implements a dual sentiment aggregation strategy and introduces a triple-significance filtering criterion to identify robust interactions. Results support Hypotheses 1 and 2, with emotions such as anticipation and trust showing consistent associations with return variation across firms and ESG categories. However, findings for Hypothesis 3 are mixed: while some sentiment-ESG combinations align with theoretical expectations, many contradictory interactions exhibit stronger effects. In this study, retrofitted embeddings outperform the NRC Emotion Lexicon in explaining stock return variation within ESG-sentiment interaction models, underscoring the value of emotional nuance in ESG-finance modeling. These results underscore the importance of emotion-sensitive sentiment modeling in understanding investor behavior and ESG-related stock price movements.