Slow Learning

This paper provides an analytic characterization of the speed of convergence under learning to a rational expectations equilibrium (REE) for a large class of multivariate models. We show that learning is slower when people's beliefs about model outcomes are more self-fulfilling. The paper also investigates which features of a model economy make beliefs more self-fulfilling, using variants of the simple new-Keynesian model and a medium-scale DSGE model. For empirically plausible specifications of these models, convergence of a learning equilibrium to the REE is so slow that analysis based on rational expectations can be misleading.