Climate and ESG Signal Extraction for Asset Pricing Using Large-Scale Text and Graph Models

Abstract

The integration of Environmental, Social, and Governance (ESG) criteria into quantitative asset pricing has transitioned from a niche investment style to a central pillar of modern financial engineering. However, the extraction of reliable ESG signals remains plagued by the unstructured nature of corporate disclosures, the prevalence of greenwashing, and the complex, non-linear dependencies inherent in global supply chains. This paper presents a novel, dual-stream neural architecture that synergizes large-scale Natural Language Processing (NLP) with Graph Neural Networks (GNNs) to distill robust pricing signals. We propose a methodology that first employs a domain-specific transformer model to extract semantic sentiment and latent risk embeddings from diverse textual corpora, including regulatory filings and news media. Concurrently, we construct a dynamic industry-relation graph that propagates these idiosyncratic signals across supply chain and ownership linkages using a graph attention mechanism. This approach addresses the limitation of independent and identically distributed (i.i.d.) assumptions in traditional factor models by explicitly modeling the spillover effects of climate risks. Our empirical analysis, conducted on a universe of global equities over a ten-year period, demonstrates that the proposed Text-Graph fusion model significantly outperforms both traditional linear factor models and standalone deep learning baselines. The resulting alpha signals exhibit low correlation with standard risk factors, offering diversification benefits and enhanced Sharpe ratios for ESG-integrated portfolios.