We develop a state space framework for estimating term structure models, where latent Markovian state variables are mapped non-linearly into observable market data. The measurement equation of our framework is explicitly constructed such that it takes raw market prices and rates as direct inputs. We thus avoid entirely, the need for data preprocessing, such as the use of ad hoc interpolation and data smoothing techniques. As our general estimation approach, we demonstrate how Markov chain Monte Carlo techniques are well suited for handling complex functional relations between state vari-ables and data, parameter restrictions and other features of popular term structure mod-els, which have proved hard to handle for alternative econometric techniques. Our estimation framework therefore handles popular multi-factor model specifications such as exponential affine and quadratic models, but facilitates richer Markovian HJM model specifications as well. Efficient Markov chain Monte Carlo implementations are highly model dependent. Therefore, having developed the general estimation principles of our framework, we demonstrate how one could approach sampler specification for a particular model example which we fit to a panel data set of swap and money market rates.
