Predictive Models for Environmental Disclosure Quality and Market Reactions

Abstract

This research introduces a novel, cross-disciplinary framework that applies computational
linguistics and machine learning techniques, traditionally used in information retrieval and
sentiment analysis, to the domain of corporate environmental disclosure. We depart from
conventional econometric models by proposing a hybrid methodology that integrates Latent Dirichlet Allocation (LDA) for thematic decomposition of sustainability reports with
a bio-inspired optimization algorithm—specifically, a modified Ant Colony Optimization
(ACO) technique—to select predictive features for market reaction modeling. The core innovation lies in reformulating the problem of assessing disclosure quality not as a static
classification task but as a dynamic, multi-dimensional signal extraction challenge, where
the ’quality’ is inferred from the semantic coherence, specificity, and forward-looking content of disclosures relative to industry-specific environmental materiality thresholds. Our
model processes unstructured textual data from corporate environmental reports and regulatory filings to generate a continuous Disclosure Quality Index (DQI). We then examine
the market’s reaction to this index, hypothesizing that investors process nuanced qualitative information differently than quantitative metrics. Using a unique dataset compiled from
2000 to 2004, our results demonstrate that the DQI, derived from our hybrid LDA-ACO
model, has superior predictive power for abnormal stock returns around disclosure events
compared to traditional metrics based on word counts or binary compliance checks. Furthermore, we identify a non-linear, threshold-based market reaction, suggesting that investors
discount disclosures until a certain level of specificity and coherence is achieved, a finding
with significant implications for both corporate reporting strategies and regulatory policy.
This work establishes a new paradigm for computational analysis in environmental, social,
and governance (ESG) research, moving beyond keyword spotting towards a sophisticated
understanding of informational substance.