Understanding trading interactions and behavior in over-the-counter markets

Abstract

This research applies machine learning methods, in particular probabilistic topic modeling, to understand patterns of interactions for Over-the-Counter (OTC) trading in corporate bonds. The interactions are between broker-dealers (dealers) and clients, or between dealers. From reports of dealer transactions, we create documents representing the daily activity of each dealer. This includes four types of dealer activities: Buy from / Sell to a client, and Buy from / Sell to another dealer. We use Latent Dirichlet Allocation (LDA) based topic models to identify communities of bonds that are bought or sold (co-traded) on the same day. Some communities reflect an industry sector, while others have a concentration of specific bonds. Several topics temporally align to notable financial events. We group dealers around topics to understand their interactions with clients and other dealers. We observe a range of interaction patterns that merit further study, including the centrality of some dealer(s) to some topics. This research illustrates that topic modeling / community detection can indeed provide insight into dealer behavior for OTC trades.