Beyond Cosine Similarity: Predicting Stock Declines from Financial Disclosures with LLM Sentiment and Deep Learning

ACL ARR 2025 May Submission229 Authors

Abstract: This paper analyzes quarterly SEC corporate disclosures for S&P 500 companies from Jan- uary 2000 to December 2019 demonstrating how large language models (LLMs) and Con- catenated Deep Learning are able to detect which companies under perform. This research finds that by comparing two quarterly corporate disclosures combined with the reasoning capa- bilities of the Claude2 large language model, negative excess returns of -11% over a 180 day period (-22% annualized) can be avoided. The paper introduces two novel approaches: (A) Concatenating Deep Learning architectures comparing quarterly filings, and (B) Summa- rization methods using Claude2 to extract sen- timent signals related to major business risks, profitability, legal, market pressures, etc. To- gether, these techniques demonstrate new ways of expanding beyond rudimentary natural language processing approaches, such as lexicons and cosine similarity, to answer fundamental questions related to firm performance.

Paper Type: Short

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Large Language Models, Deep Learning, Sentiment Analysis, Corporate Disclosures, Financial Forecasting, Natural Language Processing, SEC Filings, Firm Performance, Text Mining, Business Risk Analysis

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 229