Learning from Earnings Calls: Graph-Based Conversational Modeling for Financial Prediction

Earnings conference calls are valuable venues for business communication. Empirical research has shown that the content of earnings calls contains predictive signals about future market risks, which has motivated a line of computational studies that utilize earnings transcripts for financial forecasting tasks. However, earnings call transcripts are typically very long, and the predictive signals within them are often sparsely distributed across different sections of the transcript. As a result, existing computational methods often fail to capture the essential information within the transcript that is relevant to market risks. In this work, we design a novel method to model earnings transcripts as a conversational graph where graph nodes represent discussed topics and graph edges indicate the similarity between topics. By doing so, we aim to explicitly model what is discussed (i.e., topical content), how it is discussed (e.g., cross-referencing or newly introduced topics), and in what manner it is discussed (e.g., sentiment and other linguistic features) within the transcript. We then leverage a graph neural network to learn transcript-level representations for financial risk forecasting. Experimental results show that the proposed method significantly reduces risk forecasting errors, demonstrating its capability of modeling earnings call transcripts. Moreover, this predictive power holds even after considering the firm’s fundamentals, and the performance gain over baseline models continues to grow as transcript length increases. The interpretability analysis shows that the proposed method identifies cross-referencing and newly introduced topics as influential for risk prediction. Moreover, the model tends to associate transcripts with a higher number of new topics in the Q&A section, more cross-referencing discussions, and more positive sentiment with lower predicted financial risks. This work contributes methodologically by proposing a novel predictive approach specifically tailored to the domain-specific challenge of transcript-based risk forecasting. We also discuss key design insights and implications.

Yi Yang (HKUST), Yixuan Tang (HKUST), Yangyang Fan (HK PolyU), and Kunpeng Zhang (UMD)