We compare the performance of the MSM-POT model with competing self-exciting models and a GARCH-EVT approach in an in- and out-of-sample VaR forecasting exercise based on the extreme returns of six daily commodity futures prices (i.e. Brent and WTI crude oil, cocoa, cotton, copper, and gold). Empirical results suggest that the VaR estimates generated by the MSM-POT model for the returns analyzed produce the most accurate forecasts. (C) 2017 Elsevier B.V. All rights reserved.