This paper investigates the time-series predictability of commodity futures excess returns from factor models that exploit two risk factors – the equally weighted average excess return on long positions in a universe of futures contracts and the return difference between the high- and low-basis portfolios. Adopting a standard set of statistical evaluation metrics, we find weak evidence that the factor models provide out-of-sample forecasts of monthly excess returns significantly better than the benchmark of random walk with drift model. We also show, in a dynamic asset allocation environment, that the information contained in the commodity-based risk factors does not generate systematic economic value to risk-averse investors pursuing a commodity stand-alone strategy or a diversification strategy.

We compare major factor models and find that the Stambaugh and Yuan (2016) 4-factor model is the overall winner in the time-series domain. The Hou, Xue, and Zhang (2015) q -factor model takes second place and the Fama and French (2015) 5-factor model and the Barillas and Shanken (2018) 6-factor model jointly take third place. The pairwise cross-sectional R 2 and the multiple model comparison tests show that the Hou et al. (2015) q -factor model, the Fama and French (2015) 5-factor and 4-factor models, and the Barillas and Shanken (2018) 6-factor model take equal first place in the horse race.