Combinations of survival regression trees called survival forests (SF) applied to micro-array data provide both prediction of individual survival functions and the corresponding ranking of variable importance although without assessment for the latter. A basic question is whether the structures of SF and resulting statistics can be attributed to chance alone. For small data sets we propose the use of permutation tests as a way to determine significance of the SF performance, both for assessment of the fit and to test the significance of the genes identified as prognostic markers.