The expectations hypothesis (EH) remains the standard benchmark for Treasury bond-return forecasts despite its well-established empirical failure. I propose a yield-curve slope measure as an alternative. Under the belief that yields follow a random walk, this measure provides directly observable expected excess returns. Out of sample, this random-walk (RW) benchmark outperforms EH in forecast accuracy and economic value. An affine model and neural networks improve through refinements motivated by RW, yet none outperforms RW in statistical and economic terms simultaneously. Portfolio constraints act as an economic regularizer, letting independent variance forecasts discipline the return forecasts used for investment. […]