Semantic drift in long-form financial disclosures in Portuguese

Abstract

Financial disclosures contain critical information that is not always immediately reflected in market prices. Tracking semantic change in these communications can surface narrative shifts, but it is difficult because financial documents are long, templated, and evolve gradually, which blurs meaningful drift with routine variation. Moreover, Portuguese-language financial disclosures remain underrepresented in the financial natural language processing literature. We propose a framework for measuring semantic drift in long-form Portuguese disclosures using document embeddings, chunking, and aggregation, with drift defined as consecutive cosine distance. Article-scope inference uses 3 models and 5 aggregation techniques (15 configurations per dataset) across five datasets spanning corporate and public-sector reporting. We validate drift against volatility signals using circular-shift tests, filing-date time windows, and Granger causality, complemented by descriptive event alignment and event-study diagnostics. We find that Vale exhibits a positive drift–volatility association under within-model correction, EDP shows robust drift-to-volatility Granger predictability across configurations, and SLC shows a robust filing-date window association. Conab exhibits a similar Granger pattern only when mapped to SLC stock as market proxy, making that result proxy-sensitive. Other model–dataset combinations show weaker or non-significant links, highlighting sensitivity to document type and template structure.