This paper investigates the problems of interference prediction and sensing for efficient spectrum access and link adaptation. The considered approach for interference prediction relies on a parametric model. However, we assume that the number of observations available to learn theses parameters is limited. This implies that they should be treated as random variables rather than fixed values. We show how this can impact the spectrum access and link adaptation strategies. We also introduce the notion of "interferer-coherence time" to establish the number of independent interferer state realizations experienced by a codeword. We explain how it can be computed taking into account the model uncertainty and how this impacts the link adaptation.