This article extends Chinn and Coibion (2014)'s work-Journal of Futures Markets 34 on predictive content of commodity futures by considering a more comprehensive database and a longer time span, ranging from 25 to 65 years, and by presenting two extensions: multi-equation estimation of risk premiums and testing for the theory of storage. The empirical results show that futures-based forecasts for animal and agricultural products and industrial metals tend to be more efficient, in terms of mean absolute error, than random walk based-forecasts at a one-year horizon. On the other hand, based on robust rolling estimates, there is evidence of constant and time-varying risk premiums in agricultural and precious metals, but their statistical significance vary considerably along the sample period. In particular, gold and silver show evidence of a negative time-varying risk premium, as opposed to platinum. Multi-equation estimation brings efficiency gains in premium gauging, which leads to reject that the futures price is an unbiased estimate of the spot price for all commodity classes. On the other hand, the sampled commodities lend only partial support to the theory of storage, and for the specific case of industrial metals, inventories seem to matter more than interest rates to explain the basis. Altogether, this article finds mixed support for the premium-based model and for the theory of storage.