Kohn, Donald

Of course, gradualism and model averaging may not be appropriate in all circumstances.  For example, it may be necessary for monetary policy to respond to what might be called "tail events," along the lines suggested by recent work on "robust control."  To simplify greatly, this approach often amounts to choosing policy settings to minimize the maximum possible loss across different models of the economy, in contrast to the standard Bayesian approach, which (loosely speaking) seeks to minimize the average loss across models.  Much of the research on robust control has been a bit technical and esoteric.  But the notion that policymakers may at times base policy settings on especially pernicious risks has an important ring of truth. 

For example, in 2003 the FOMC noted that a continued fall in inflation would be unwelcome largely because such an eventuality might potentially lead to persistently weak real activity with interest rates stuck at zero.  Partly in response, the FOMC reduced the federal funds rate to an unusually low level and kept it there for an extended period, in a manner that perhaps would not have occurred in the absence of concerns about the "worst case" effects of deflation.  This type of risk management--in which the central bank takes out some insurance against a bad but improbable event--has been an aspect of policymaking for some time and does seem to respond to extreme risks in a way reminiscent of the literature on robust control.