A new fitness model has been developed that predicts evolution of the seasonal human influenza A/H3N2 virus from genomic data. The model is based on the fitness effects of mutations in antibody-binding epitope and in non-epitope regions of haemagglutinin, which is a surface protein of the virus. Beneficial mutations in epitopes and harmful (that is, deleterious) mutations in non-epitope regions are used to infer fitness of individual strains. Predictions of growth and decline of viral groups (that is, clades) in one year on the basis of frequency and fitness of individual strains in a previous year were accurate 93% and 76% of the time, respectively.