Multimodal multi-speaker merger & acquisition (M3A) financial forecasting: a new task, dataset, and neural baselines

Abstract

Financial risk prediction is an essential task in today’s financial markets, especially with the disclosure of important information such as Merger and Acquisition (M&A) calls nowadays. M&A calls often provide key insights into the confidence levels and agreeability between the different high ranking members of an organization concerning a merger or acquisition in the form of subtle vocal cues and verbal messages. Variations in vocal features such as pitch can suggest doubt and a lack of confidence that may discredit the message spoken. Additionally, information about the speaker of the message can assist the system in making better predictions. To aid the analysis of M&A calls, we curate a dataset of conference call transcripts and audios from the past 5 years. We propose strong baseline architectures to accompany the model that takes advantage of the multimodal multi-speaker input to perform financial forecasting. Empirical results show the improvement that our architecture gives over existing models.