Municipal Bond Pricing: A Data Driven Method

Price evaluations of municipal bonds have traditionally been performed by human experts based on their market knowledge and trading experience. Automated evaluation is an attractive alternative providing the advantage of an objective estimation that is transparent, consistent, and scalable. In this paper, we present a statistical model to automatically estimate U.S municipal bond yields based on trade transactions and study the agreement between human evaluations and machine generated estimates. The model uses piecewise polynomials constructed using basis functions. This provides immense flexibility in capturing the wide dispersion of yields. A novel transfer learning based approach that exploits the latent hierarchical relationship of the bonds is applied to enable robust yield estimation even in the absence of adequate trade data. The Bayesian nature of our model offers a principled framework to account for uncertainty in the estimates. Our inference procedure scales well even for large data sets. We demonstrate the empirical effectiveness of our model by assessing over 100,000 active bonds and find that our estimates are in line with hand priced evaluations for a large number of bonds.