Learning with Bounded Memory in Stochastic Models

There exists by now a sizeable literature that studies the dynamics of adaptive learning in stochastic macroeconomic models. A common starting point is to postulate that economic agents use standard econometric techniques to estimate the unknown parameters of the stochastic process of the relevant variables and forecast the future values using these estimated parameter values. A feature of learning is that, in the limit, agents are assumed to have access to an infinite amount of data. Our goal here, by contrast, is to analyze finite memory rules in stochastic economic models. We consider a wide variety of macroeconomic models, both linear and nonlinear, where agents are learning steady states. We study some basic issues here. Does the state of the economy have some invariant distribution in the long run? Is there convergence of the moments of the forecast? What is the influence of memory length on the residual variance of these forecasts? What can one say about these moments in nonlinear models? We provide answers to these questions for the models we analyze.