Omni-FinAI: Unlocking Financial Disclosure Insights

Abstract

This study introduces Omni-FinAI, a finance-specific large language model based on the LLaMA 3.1 8B and 70B architectures, pre-trained on 143 billion tokens of financial text. To demonstrate its utility, the model is applied to ChatGPT-condensed earnings conference call summaries, designed to extract essential disclosure insights. Omni-FinAI achieves high predictive accuracy for financial statement variables, such as asset growth, revenue growth, and Tobin’s Q growth. It also demonstrates strong sentiment analysis capabilities, identifying embedded sentiment signals to reveal dynamics in financial variables and event-based market reactions. Furthermore, our experiments validate Omni-FinAI’s effectiveness in firm-specific stock performance prediction. The trading strategies formed by AI signals achieve an average monthly Sharpe ratio of 0.548, significantly outperforming the market benchmark of 0.171. These findings highlight the transformative potential of domain-specific pre-training for advancing textual analysis with generative AI models.

Keywords: Generative AI, Large Language Model, Textual Analysis, Earnings Conference Call, Disclosure, Sentiment Analysis C45

JEL Classification: C45, D81, G12, G30, G32, M41