Seasonal influenza viruses repeatedly infect humans in part because they rapidly change their antigenic properties and evade host immune responses, necessitating frequent updates of the vaccine composition. A better understanding of this evolution could allow for better predictions and would be important for vaccine design. We investigated the predictability of frequency dynamics and fixation of amino acid substitutions and found that the current frequency was the strongest predictor of eventual fixation, as expected in neutral evolution. Simulations of models of adapting populations, in contrast, show clear signals of predictability. This indicates that the evolution of influenza HA and NA, while driven by strong selection pressure to change, is poorly described by common models of directional selection such as travelling fitness waves. In consequence, we propose an alternative way to view influenza's evolution where due to continuous changes in human immune response, adaptive mutations do not fix but rather reach an equilibrium frequency. This leads to think about the viral population in terms of equilibrium between strains rather than of fixed individual fitness values. We explore the rich phenomenology of this model which qualitatively matches observations about influenza, such as ladder-like genealogies and rapid but unpredictable evolution.