The performance of hybrids can be predicted with gene expression data from their parental inbred lines. Implementing such prediction approaches in breeding programs promises to increase the efficiency of hybrid breeding. The objectives of our study were to compare the accuracy of prediction models employing multiple linear regression (MLR), partial least squares regression (PLS), support vector machine regression (SVM), and transcriptome-based distances (D(B)). For a factorial of 7 flint and 14 dent maize lines, the grain yield of the hybrids was assessed and the gene expression of the parental lines was profiled with a 56k microarray. The accuracy of the prediction models was measured by the correlation between predicted and observed yield employing two cross-validation schemes. The first modeled the prediction of hybrids when testcross data are available for both parental lines (type 2 hybrids), and the second modeled the prediction of hybrids when no testcross data for the parental lines were available (type 0 hybrids). MLR, SVM, and PLS resulted in a high correlation between predicted and observed yield for type 2 hybrids, whereas for type 0 hybrids D(B) had greater prediction accuracy. The regression methods were robust to the choice of the set of profiled genes and required only a few hundred genes. In contrast, for an accurate hybrid prediction with D(B), 1,000-1,500 genes were required, and the prediction accuracy depended strongly on the set of profiled genes. We conclude that for prediction within one set of genetic material MLR is a promising approach, and for transferring prediction models from one set of genetic material to a related one, the transcriptome-based distance D(B) is most promising.