The dissertation examines several policy-related implications of relaxing the assumption that economic agents are guided by rational expectations. A first, introductory chapter presents the main technical issues related to adaptive learning. The second chapter studies the implications for monetary policy of positing that both the private sector and the central bank form their expectations through adaptive learning and that the central bank has private information on shocks to the economy but cannot credibly commit. The main finding of this chapter is that when agents learn adaptively a bias against activist policy arises. The following chapter focuses on large, non-linear models, where no unambiguous linear approximation eligible as perceived law of motion exists. Accordingly, there are heterogeneous expectations and the system converges to a misspecification equilibrium, affected by the communication strategies of the central bank. The main results are: (1) the heterogeneity of expectations persists even when a large number of observations are available; (2) the monetary policymaker has no incentive to be an inflation hawk; (3) partial transparency enhances welfare somewhat but full transparency does not. The final chapter adopts a model in which agents are fully informed and use Bayesian techniques to estimate the hidden states of the economy. The monetary policy stance is unobservable and state-independent, generating uncertainty among agents, who try to gauge it from inflation: a change in consumer prices that confirms beliefs reduces stock risk premia, while a change that contradicts beliefs drives the risk premia upward. This may generate a negative correlation between returns and inflation that explains the Fisher puzzle. The model is tested on US data. The econometric evidence suggests: (1) that a mimickingportfolio proxying for monetary policy uncertainty is a risk factor priced by financial markets; and (2) that conditioning on monetary uncertainty and fundamentals eliminates the Fisher puzzle.