Multimodal Data Fusion for Corporate Credit Assessment: Integrating Sentiment Signals, Textual Information, and Financial Indicators

Abstract

Corporate credit assessment traditionally relies on financial indicators, although accounting data may not capture qualitative risk signals embedded in corporate disclosures. This study develops a non-experimental, multi-source evidence-integration framework combining structured financial indicators, annual-report text, and financial sentiment. Three public data sources are used: a Financial Distress dataset containing 3,672 company-period observations from 422 anonymous firms, Financial PhraseBank containing 4,846 labelled sentences, and Enron's 2000 annual report containing 960 valid extracted sentences. Because these datasets do not share firm-year identifiers, they are not artificially merged. Instead, the first two datasets provide financial-risk and sentiment references, while Enron serves as the only aligned case. The results show that Enron's revenue and assets expanded rapidly between 1999 and 2000, but its debt ratio increased from 0.713 to 0.825 and return on assets declined from 2.68% to 1.49%. Textual analysis identified strong emphasis on growth, expansion, innovation, and market leadership alongside risk-, exposure-, and contingency-related language. Neutral and positive sentiment dominated, whereas direct negative language remained limited. The findings indicate that financial indicators remain fundamental, annual-report text reveals disclosure priorities, and sentiment should be interpreted jointly with financial conditions. Cross-modal inconsistency can provide additional evidence for corporate credit review and counterparty monitoring.