| This dissertation provides a new comprehensive study of a class of theory-based yield curve models that are due to Ang and Piazzesi (2003). The most general model that we study is related to Gaussian term structure models (Vasicek, 1977 and Dai and Singleton, 2000) and the model of Ang and Piazzesi (2003). In this model, yields are explained by three latent and three macro factors. In addition, the parameters of the model are restricted to satisfy a certain theoretically relevant no-arbitrage condition. A second goal of this dissertation is to provide a detailed study (and comparative assessment in relation to the first class of models) of a popular class of nontheory based yield curve models that stem from the contributions of Nelson and Siegel (1987) and Diebold, Rudebusch and Aruoba (2005). Yields in this model also depend on both latent and observed macro factors. Both classes of models are completed by the assumptions about the evolution of the factors.; Although each model has a state space form, the fitting of the no-arbitrage model is substantially more complex partly on account of the no-arbitrage restrictions and partly on account of the large number of parameters. We resolve these difficulties from a Bayesian approach that we implement via tuned Markov chain Monte Carlo (MCMC) methods (Chib, 2001). Centerpieces of our approach include the use of square-root filtering techniques for calculating the likelihood function and appropriate blocking of the parameters and tuned Metropolis steps to efficiently sample the posterior distribution. We provide simulation evidence to support the validity of our approach. Each model is fit to a data set on monthly yields on 8 US Treasuries with maturities ranging from 3 to 120 months. The macro factors are capacity utilization, Federal funds rate and the CPI. In contrast to findings reported elsewhere, we find quite different factor risk estimates in the case of the theory-based models and a stronger link between macro and latent factors in the case of the non-theory based models. We also compare the two classes of models in terms of the in-sample fit, as measured by the respective model marginal likelihoods (Chib, 1995 and Chib and Jeliazkov, 2001), and multi-period ahead out-of-sample Bayesian point forecast and predictive intervals. Our conclusion is that the no-arbitrage models provide better forecasts of both yields and macro factors, while the non-theory based models provide a better fit to the data. |
