Natural Language Processing of ESG Disclosures with FinBERT and AraBERT: Insights into Retail Investor Flows in the Abu Dhabi Securities Exchange (ADX)

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a novel NLP methodology for quantifying ESG disclosure credibility and testing signaling theory in emerging markets, replicable in other contexts.

🏢実務担当者:Corporate sustainability teams can learn to differentiate credible ESG signals from marketing and adjust disclosure strategies to attract retail investors.

🏛政策担当者:Regulators can use these findings to understand the market impact of ESG disclosure quality and emphasize verifiable content over symbolic reporting.

Abstract This study introduces a novel computational framework for understanding how the credibility of environmental, social, and governance (ESG) disclosures shapes retail investor behavior in emerging markets. Focusing on 125 firms listed on the Abu Dhabi Securities Exchange (ADX) from 2021 to 2025, the research develops two proprietary indices ESG_MKT (marketing intensity) and ESG_AI (AI-based disclosure credibility) derived through a multilingual natural language processing (NLP) pipeline integrating FinBERT for English and AraBERT for Arabic texts. This bilingual design represents one of the first large-scale applications of transformer models to sustainability reporting in the Gulf region. By fusing computational linguistics with behavioral finance, the study bridges the gap between symbolic communication and substantive disclosure, offering a new lens through which to examine signaling credibility in capital markets. A two-way fixed effects (TWFE) model and event-study design are employed on a balanced panel of over 80,000 firm–day observations, revealing that high-credibility ESG signals—those supported by quantifiable evidence and external verification—generate statistically significant positive abnormal retail flows. Conversely, symbolic ESG marketing campaigns lacking verifiable content produce muted or even negative investor reactions. Robustness tests using alternative sentiment frameworks (RoBERTa, VADER), ownership stratification, and dynamic panel estimation confirm the persistence of these effects. The findings provide theoretical evidence that costly, verifiable signals enhance market trust, while inexpensive, reputational messages erode it—empirically validating signaling theory (Spence, 1973) within a computational finance context. This research contributes to the intersection of NLP, ESG analytics, and market microstructure, establishing a reproducible methodological foundation for measuring credibility in sustainability communication. Beyond empirical insight, it advances the discourse on algorithmic transparency, linguistic asymmetry, and investor cognition, positioning the UAE as a testbed for the future of AI-driven sustainable finance in emerging economies.

• openaire https://doi.org/10.21203/rs.3.rs-8073898/v1first seen 2026-05-05 19:08:43