Accounting Information Usefulness in Credit Risk Evaluation by Financial Institutions

Abstract

This research presents a novel, cross-disciplinary framework that re-conceptualizes the utility of accounting information in credit risk evaluation by integrating principles from computational linguistics,
network theory, and behavioral finance. Moving beyond traditional ratio analysis and conventional financial statement evaluation, we propose a methodology that treats accounting disclosures as a complex,
multi-layered semantic network. This network is analyzed not only for its explicit numerical content but
also for its implicit narrative structures, temporal consistency patterns, and inter-statement relational
dynamics. Our approach employs a hybrid technique combining transformer-based natural language
processing models, specifically fine-tuned for financial discourse, with graph neural networks to map
the latent connections between accounting line items, management discussion narratives, and footnote
disclosures. This creates a holistic ’Accounting Information Coherence Score’ (AICS). We formulate and
address the unconventional research question: To what extent does the semantic and structural coherence
of accounting information, as a system, predict credit risk more accurately than the information’s individual quantitative components? Using a unique dataset of corporate loan applications and subsequent
default events from a consortium of mid-sized U.S. financial institutions, we demonstrate that the AICS
provides a statistically significant improvement in default prediction accuracy (AUC-ROC improvement
of 0.12) over models relying solely on traditional financial ratios and credit scores. Furthermore, the
model reveals that specific patterns of narrative-quantitative dissonance and footnote network fragmentation are strong, early indicators of financial distress often missed by human analysts and standard
models. The findings challenge the prevailing reductionist view of accounting data in credit analysis
and advocate for a systemic, integrative evaluation paradigm. This research contributes original insights
to accounting information systems, risk management, and fintech, proposing a shift from information
extraction to system coherence assessment in financial decision-making.