zero. For the first 50 simulated quarters, the cumulated real output cost is 15.68 in the
response of the estimated model (also shown in Figure 9) and is 8.93 in the response of the
faster learning model.
Overall, these simulations illustrate that with faster learning, nominal variables approach
their natural rate much more rapidly, total real effects of shocks are smaller, but the
volatility of responses to transitory policy shocks is larger. However, the results suggest
that excess volatility that occurs in the absence of permanent target shocks may be reduced
by allowing the learning parameter to vary over time. One approach is to assume aggregation
across heterogeneous expectations differentiated by learning speed. Time variation of
the aggregate learning parameter could reflect different weightings across alternative
constant-gain learners. In the current model, target shocks, whether from accommodation
of supply shocks or from exogenous sources, are assumed to be homoskedastic. In a
more general specification, target shocks might be viewed as heteroskedastic, either due
to heteroskedasticity in the underlying structural shocks or to a change in the parameter
governing the accommodation of aggregate supply shocks. With heteroskedastic target
shocks, a larger fraction of learners may choose to use fast learning when the size of recent
policy forecast errors increases and slow learning when the size of policy forecast errors
decreases. Such time-varying aggregation may be accomplished using multinomial logit
techniques as in Kozicki and Tinsley (2001a) or Branch (2003).
6FinalRemarks
This paper proposes an empirical model fit to U.S. data, with permanent and transitory
policy shocks and asymmetric policy information. The model incorporates a possibly
time-varying inflation target and a time-varying perceived inflation target. The perceived
inflation target may differ from the actual target because private agents are assumed
to not observe the true target and must learn whether movements in the federal funds
27