Predicting Corporate Risk by Jointly Modeling Company Networks and Dialogues in Earnings Conference Calls

More and more researchers focus on studying company risk prediction based on earnings conference calls because of their free form and rich information. However, existing research does not take speaker role information into account. Besides, current research does not fully consider the impact of inter-company relationships on company risk. The only study integrating company networks and earnings conference calls constructs companies in an undirected graph, which does not meet the requirement of no temporal information leakage for prediction tasks. To solve the above problems, we propose a new model -- Temporal Virtual Graph Neural Network (TVGNN), to incorporate earnings conference calls and company networks for company risk prediction. Our model incorporates the speaker's role information in the dialogue modeling for the first time. In addition, we design a new method to construct company networks that can ensure no temporal information leakage in the graph. The experimental results show that the proposed model exceeds all baselines. The case study shows that the prediction results of the model are interpretable.