Multivariate Network Modeling: A Novel Approach to Corporate Bond Pricing

In this paper, we present a novel, nonlinear network model approach for multivariate complex systems based on sparse network representations. The proposed model combines a sparse information filtering network representation of the dependency structure with a multivariate probabilistic modeling. This allows us estimate the high-dimensional joint probability distribution of a large number of variables from low-dimensional local estimates of small subgroups of variables. We demonstrate the effectiveness of this methodology by applying it to price discovery for a large system of US Corporate Bonds. The Corporate Bond Market is good testing ground because it is characterized by a high degree of complexity due to many non-standard securities, making price discovery particularly challenging. Our network model approach yields the probability distribution of a given bond return conditioned on the returns of the other bonds in the market. The network structure also provides information on the subset of bonds most closely related to the return of a given bond, which can be used for risk assessment and recommender systems. We show that the proposed multivariate network model outperforms traditional factor-based models. This is achieved with a purely data-driven approach without using any information about the securities besides their returns.