This paper introduces a novel methodology for the extraction and modelling of the unobserved term structure of interest rates which incorporates in a single inferential framework both cross-sectional and time-series information from observed bond prices. In doing so, the paper introduces both a parametric and a semi-parametric dynamic model for the term structure which can be extended to capture also features such as heteroscedasticity in both the time and cross-section dimension, as well as the zero-lowerbound constraint. The models provide a coherent description of the term structure and outperform current term structure extraction methods in fitting the observed bond prices surface. The models also outperform other existing dynamic term structure models in forecasting observed bond prices at one, six and twelve months horizons. Moreover, the study highlights a strong sensitivity of the forecasting errors of the existing dynamic models to the choice of the term structure extraction method used to construct the samples for the parameters' estimation.