In this paper we provide new tests for the difference in predictive accuracy of two prognostic factors X1 and X2 on a common output Y. Given a set of independent replicates of (X1,X2,Y), we split this sample into a learning part for estimating the Unknown regression functions, and a validation part for which the residuals need to be computed. We show that the null distributions of our test statistics may be approximated by a normal. In simulations, the power is promising already for small to moderate sample sizes. Extensions to the time series context are briefly outlined.