In this paper we document the results of a forecast evaluation exercise for the real world price of crude oil using VAR models estimated by sparse (regularization) estimators. These methods have the property to constrain single parameters to zero. We find that estimating VARs with three core variables (real price of oil, index of global real economic activity, change in global crude oil production) by the sparse methods is associated with substantial reductions of forecast errors. The transformation of the variables (taking logs or differences) is also crucial. Extending the VARs by further variables is not associated with additonal gains in forecast performance as is the application of impulse indicator saturation before the estimation.