Motivated by prior evidence that the price of risk varies across frequencies, we study the predictability of monthly excess bond returns estimating latent factors generating common macroeconomic cycles of different lengths. Our method combines a new band spectrum principal component estimator for frequency-specific factors and supervised learning. Not all macroeconomic cycles are found to predict bond returns in real time, on the contrary, predictability concentrates only at some bands of frequencies. Two factors are powerful out-of-sample predictors: an inflation factor obtained maximizing common macroeconomic cycles of at least 8 years and a term spread factor obtained maximizing common macroeconomic cycles of 1 to 3 years. We find that the inflation factor is unspanned by the current cross-section of yields (and forward rates) and relatively more accurate at shorter maturities and during recessions. The term spread factor is relatively more accurate at longer maturities and during expansions. Unlike all previous works using nonoverlapping returns and data available in real-time, our forecasts and generate sizeable economic value for investors of various kinds. In line with models based on countercyclical risk aversion, our factors generate countercyclical expected returns and term premia, and higher predictability during recessions.