AI Based Analysis of Climate Risk Disclosures in Financial Statements

Abstract

This research introduces a novel, hybrid artificial intelligence methodology for the systematic extraction, classification, and quantification of climate-related financial risk disclosures
from corporate annual reports and financial statements. While the integration of environmental, social, and governance (ESG) factors into financial analysis is gaining prominence,
existing approaches remain largely manual, qualitative, and inconsistent, creating significant
information asymmetry. This paper addresses this gap by proposing an original framework
that synergistically combines rule-based natural language processing (NLP) for precise clause
identification with a transformer-based deep learning model, specifically adapted for financial semantic similarity, to categorize disclosures according to a novel, granular taxonomy of
climate risks. This taxonomy, a key contribution, moves beyond broad categories to distinguish between physical risks (acute vs. chronic), transition risks (policy, technology, market,
reputational), and liability risks, further classifying them by time horizon and financial materiality. We train and validate our model on a unique, hand-annotated corpus of 1,500 annual
reports from SP 500 companies spanning 1995-2004, a period preceding mainstream ESG reporting mandates, allowing us to trace the nascent evolution of such disclosures. Our results
demonstrate that the hybrid AI system achieves a 92.7% F1-score in disclosure detection
and an 88.4% accuracy in multi-label risk classification, significantly outperforming standard keyword search and generic sentiment analysis tools. The quantitative output reveals
previously unobserved patterns: early disclosures (pre-2000) are overwhelmingly narrative
and focused on generic reputational risks, while a statistically significant shift (p ¡ 0.01) occurs post-2000 towards more quantifiable disclosures of regulatory and technology transition
risks, particularly in energy and manufacturing sectors. Furthermore, we establish a novel
’Climate Risk Disclosure Score’ (CRDS) and find a weak but emerging positive correlation
(r = 0.18, p ¡ 0.05) with subsequent stock price volatility for high-carbon-intensity firms,
suggesting investors were beginning to price this information, albeit imperfectly. This work
provides the first scalable, auditable AI tool for longitudinal analysis of climate risk reporting, offering regulators, investors, and researchers a powerful means to assess transparency,
compare practices, and investigate the financial materiality of climate-related disclosures in
the era preceding contemporary reporting standards.