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

Abstract

Practice- and policy-oriented abstract Earnings conference calls are a critical channel through which public companies communicate with investors, analysts, and regulators. These conversations contain timely signals about firms’ future risk, yet their length and unstructured nature make systematic analysis difficult in practice. This study develops an artificial intelligence (AI)–based approach, motivated by a body of theoretical and empirical work from finance and accounting, that transforms earnings call transcripts into structured representations, allowing key aspects of managerial communication, such as topic flow, cross-referencing, sentiment, and other semantics, to be explicitly modeled. Empirical results show that the proposed model significantly improves the prediction of future financial risk compared with existing deep learning and large language model approaches, particularly for firms with complex and lengthy disclosures. The findings highlight that nuanced manager–analyst interactions within earnings calls contain value-relevant information for market participants. In particular, cross-referencing, the introduction of new topics, and a more positive tone are associated with lower subsequent risk as identified by the proposed approach. For researchers and practitioners, this work demonstrates how theoretical and empirical evidence on managerial communication can be incorporated into predictive model design, supporting more accurate, interpretable, and responsible use of AI in financial markets.