Abstract Background RNA-seq data are increasingly used to derive prognostic signatures for cancer outcome prediction. A limitation of current predictors is their reliance on reference gene annotations, which amounts to ignoring large numbers of non-canonical RNAs produced in disease tissues. A recently introduced kind of transcriptome classifier operates entirely in a reference-free manner, relying on k-mers extracted from patient RNA-seq data. Methods In this paper, we set out to compare conventional and reference-free signatures in risk and relapse prediction of prostate cancer. To compare the two approaches as fairly as possible, we set up a common procedure that takes as input either a k-mer count matrix or a gene expression matrix, extracts a signature and evaluates this signature in an independent dataset. Results We find that both gene-based and k-mer based classifiers had similarly high performances for risk prediction and a markedly lower performance for relapse prediction. Interestingly, the reference-free signatures included a set of sequences mapping to novel lncRNAs or variable regions of cancer driver genes that were not part of gene-based signatures. Conclusions Reference-free classifiers are thus a promising strategy for the identification of novel prognostic RNA biomarkers.