Analyzing Textual Information from Financial Statements for Default Prediction

Abstract

Financial statements provide a view of company’s financial status at a specific point in time including the quantitative as well as qualitative view. Besides the quantitative information, the paper asserts that the qualitative information present in the form of textual disclosures have high discriminating power to predict the financial default. Towards this, the paper presents a technique to capture comprehensive 360- ∘ features from qualitative textual data at multiple granularities. The paper proposes a new sentence embedding (SE) from large language models specifically built for financial domain to encode the textual data and presents three deep learning models built on SE for financial default prediction. To accommodate unstructured and non-standard financial statements from small and unlisted companies, the paper also presents a document processing pipeline to be inclusive of such companies in the financial text modelling. Finally, the paper presents comprehensive experimental results on two datasets demonstrating the discriminating power of textual features to predict financial defaults.