This paper studies whether the evident statistical predictability of bond risk premia translates into economic gains for bond investors. We show that ane term structure models (ATSMs) estimated by jointly tting yields and bond excess returns capture this predictive information otherwise hidden to standard ATSM estimations. The model's excess return predictions are unbiased, produce regression R2s beyond those reported in the literature, exhibit high forecast accuracy, and allow to generate positive bond portfolio excess returns in- and out-of-sample. Nevertheless, these models cannot beat the expectations hypothesis (EH) out-ofsample: the forecasts do not add economic value compared to using the average historical excess return as an EH-consistent estimate of constant risk premia. We show that in general statistical signicance does not necessarily translate into economic signicance because EH deviations mainly matter at short horizons and standard predictability metrics are not compatible with common measures of economic value. Overall, the EH remains the benchmark for investment decisions and should be considered an economic prior in models of bond risk premia.
Term structure of interest rates; Expectations hypothesis; Affine models; Risk premia