of learning by agents often specify that agent expectations are generated by recursive least
squares (RLS) estimates of linear forecast functions (perceived laws of motion). RLS
learning is a natural assumption for models that are known to have fixed coefficients.
However, RLS learning is likely to be less relevant for economies with time-varying structures
(including the case of shifting inflation targets).
Alternative learning methods for time-varying shifts include change point models,
explored by Kozicki and Tinsley (2001a), and variants of bounded memory learning, such
as rolling sample estimation or constant gain estimation, illustrated in the simulations
of Orphanides and Williams (2003). Other examples of constant gain learning in the
context of rare permanent real shocks are Bullard and Duffy (2003) and Bullard and Eusepi
(2003). As noted in these studies, benefits of constant gain learning include both recursive
tractability and sensitivity to recent observations that incorporate the current shifts of the
underlying economic structure. Although bounded memory learning can converge to a
rational expectations equilibrium in a deterministic model, convergence will not occur in
models with stochastic shocks, Evans and Honkapohja (2003). Consequently, an offsetting
cost of constant gain learning, even if shifts are infrequent, is that the economy will not
converge to the new rational expectations equilibrium path but to an ergodic distribution
about the REE. Intuitively, smaller rolling sample spans or larger constant gains provide
a tradeoff between faster recognition of new shifts and higher “excess” volatility in steady
states.
In this paper, implications for the transmission mechanism of possibly shifting policy
goals, imperfect policy credibility (or asymmetric information), and private sector learning
are examined by comparing the properties of two empirical models of the U.S. economy:
a model with time-varying policy goals and learning; and, a model with similar dynamic
structure but with a constant and credible inflation target. Many of the studies cited earlier
address asymmetric information and learning issues using analytical or calibrated models.