W e know a considerable amount about the performance of tactical style allocation models in equity markets, but very little evidence is available on the performance of systematic dynamic allocation decisions on various bond benchmarks with different maturities. Most of the literature on predictability in bond returns focuses on timing bonds versus stocks or bonds versus cash, with no emphasis on the timing of bonds with different maturities.Research on tactical asset allocation decisions involving bond markets includes Shiller [1979], Fama [1981], Shiller, Campbell, andSchoenholtz [1983], Keim and Stambaugh [1986], Campbell [1987], Fama and Bliss [1987], Fama and French [1989], Campbell andShiller [1991], Ilmanen [1995, 1997], Bekaert, Hodrick, and Marshall [1997], Lekkos and Milas [2001], Ilmanen and Sayood [2002], and Baker, Greenwood, and Wurgler [2003]. These authors focus on exploiting predictability in a global bond portfolio and hence in the level of interest rates, but they do not try to exploit predictability on other dimensions of the shape of the yield curve such as slope and curvature. 1 It is only recently that some articles have recognized the benefits of exploiting predictability in the shape of the yield curve, although to the best of our knowledge, there are only two. Dolan [1999] argues that the curvature parameter of the yield curve, estimated using the Nelson-Siegel [1987] model, can be predicted using simple parsimonious models, and shows these forecasts have investment significance in the selection of bullet over barbell portfolios. Diebold and Li [2002] estimate autoregressive models for predicting Nelson-Siegel level, slope, and curvature factors.We extend this research on several dimensions. First, we test for statistical significance in the predictive power of a series of economically meaningful variables. This approach stands in sharp contrast with Dolan [1999] and Diebold and Li [2002], who use only information about past values of the term structure parameters in their predictive experiments. We thus bridge the gap in the literature on predictability of asset returns on the basis of variables such as dividend yields or term spread. Our work is also related to research based on joint macrofinance modeling strategy of the term structure of interest rates (e.g., Ang and Piazzesi [2003], Diebold, Rudebusch, andAruoba [2005], or Rudebusch and Wu [2004]).Like Pesaran and Timmermann [1995], we investigate the predictability of bond portfolio returns using a robust recursive modeling approach based on multifactor models. This allows us to alleviate concerns over spurious results driven by data-mining biases. In the interest of robustness and in an attempt to account for model uncertainty, we use a Bayesian econometric approach, known as thick modeling, which selects at each date a "council" of models to make predictions rather than a single model.