ESG Alpha Through Generative AI: A New Paradigm for Sustainable Trading Strategies

Abstract

Institutional investors increasingly seek to reconcile return objectives with measurable sustainability outcomes, yet ESG information remains noisy, heterogeneous, and slow to update. This paper proposes a unified framework—ESG Alpha Through Generative AI—that operationalizes large-language models (LLMs) and generative modeling to construct transparent, regulation-ready trading strategies. First, we develop an LLM-powered ESG sentiment engine that ingests multi-lingual, unstructured disclosures and news, applies retrieval-augmented extraction with source attribution, and synthesizes security-level signals calibrated against established ESG taxonomies. Second, we introduce an automated scenario generator that translates narrative ESG risks (e.g., transition policy shocks, supply-chain violations, climate physical hazards) into factor-consistent return and fundamentals shocks, enabling robust portfolio construction via mean–CVaR optimization with explicit sustainability and concentration constraints. Third, we embed a “compliance-by-design” layer that maps exposures to the EU Taxonomy and related regimes (e.g., SFDR), supports pre- and post-trade checks, and preserves auditability through prompt logging, data lineage, and human-in-the-loop review. Back-tested experiments against benchmarks reliant on static ESG scores indicate improved timeliness of signals, enhanced risk-adjusted performance, and reduced drawdowns while maintaining taxonomy alignment. The paper’s contributions include: (i) an end-to-end architecture for LLM-derived ESG signal generation, (ii) a generative ESG risk-scenario apparatus for portfolio optimization, and (iii) a governance blueprint that aligns model risk management and regulatory compliance. We discuss implementation considerations, limitations, and avenues for future research on bias mitigation, domain adaptation, and supervisory evaluation.