The financial instability hypothesis advanced by Minsky (1975, 1982, 1986) is not compatible with the rational expectations hypothesis in that firms persist in adopting liability structures which give rise to outcomes which in turn violate the assumptions on the basis of which those liability structures were chosen. This occurs both in the case of excessive caution following a period of instability, and excessive boldness following a long expansion. However, within the context of a formal model, it is shown that such behavior need not result from irrationality or myopia, but may arise instead from the fact that agents do not have sufficient information to compute the RE magnitudes, and must therefore rely on a learning process that makes use of publicly observable data. The dynamics of two commonly used learning processes are examined, one of them Bayesian. Both generate trajectories which differ sharply in their qualitative properties from the RE trajectory, and are in fact quite consistent with the predictions of the FIH.