View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

10:30 A.M. CST (11:30 A.M. EST)

Remarks by
Henry C. Wallich
Member, Board of Governors of the Federal Reserve System
at the Session
Application of Optimal Control to Problems of Economic Stabilization

at the Annual Meeting of the
American Economic Association
Dallas, Texas
Sunday, December 28, 1975

Remarks by
Henry C. Wallich
Member, Board of Governors of the Federal Reserve System
at the Session
Application of Optimal Control to Problems of Economic Stabilization
at the Annual Meeting of the
American Economic Association
Dallas, Texas
Sunday, December 28, 1975

The use of optimal control techniques in planning for
economic stabilization is approaching the policy stage.

At the

present time, as the papers before us show, the principal applica­
tion of these techniques has been the examination of models and of
past policies.

Its use for effective policy advice still seems some

distance away.

But initial efforts to build an optimal control

approach into Federal Reserve policymaking are underway.

I believe

that there is potential for progress at both the technical and the
policy levels.

It is important, therefore, for the producers and the

Note: The views expressed herein are my own and do not
necessarily reflect those of the Board of Governors or the B o a r d fs

-2 -

potential users of this technique to become better acquainted.
Model builders and policymakers must explore one anotherfs needs
and capabilities.
I appear here, of course, as a potential user, with no
pretense to technical expertise.

In this capacity, I would like to

comment on a number of points raised by the papers of Ando-Palash,
Chow, and Kalchbrenner-Tinsley. My remarks will be addressed mainly
to the loss function, the departures from present practices implied,
some features of the models employed, and the relation between
uncertainty and the scale of policy action.
Policymakers, I believe, regard their role as somewhat more
modest than that with which the terminology of the loss function some­
times endows them.

The Federal Reserve, to be specific, is responsible

for only one phase of the n a t i o n fs economic policy —
monetary policy.
Employment Act.

the handling of

The overall objectives, moreover, are given by the
Most of the economic policies that influence the

rate of growth, employment, and the degree of price stability, are
handled elsewhere in the government.

Particularly when several

objectives are involved, which obviously cannot all be attained with
one instrument, it seems somewhat presumptuous to state o n e fs preferences
in the form of "targets."
The monetary policy official naturally has ideas also about
desirable fiscal policy, and about many other policies that influence
economic development.

Only in the very short run can he make fairly

-3firm assumptions as to what these policies will be.

For the longer

run, a not unreasonable attitude for him may be to think of monetary
policy as helping to create the environment in which other public
policies, as well as decisions made in the private sector, will
become effective.

Any given monetary policy may be consistent with

alternative combinations of growth, unemployment, and inflation.
The monetary policymaker will adjust his action to what he sees
happening in these other spheres.

But he should not overestimate

his ability to influence the outcome.
The time horizon over which target values are to be set
likewise presents difficult problems.

One may believe that a lower

rate of economic expansion in the immediate future will lead to more
sustainable growth and lower ultimate unemployment and inflation than
would a more aggressive policy.

But unless such preferences are built

into a loss function, and a long time horizon is allowed for, rather
extreme policy proposals may follow from optimal control techniques
applied to econometric models with long lag structures, as some of
the papers at this meeting indicate.
The policymaker may also be troubled by an appearance of
misplaced concreteness.
terms -

up or down -

faster or slower.

He may be accustomed to thinking in directional
or in terms of rates of change --

He may want to reserve judgment as to precise

targets for unemployment and inflation until the economy is a little

-4closer to what he might consider optimal.

And if in addition he were

asked whether he has a quadratic loss function, i.e., whether he is
indifferent to an equal degree of over- and undershooting of his
targets, he might be tempted to think the whole thing a spherical
nuisance, i.e., a nuisance from every angle.
Monetary policy in the United States, moreover, is made by
a group, the Federal Open Market Committee.
its members are unlikely to be identical.

The loss functions of all
Perhaps one could think of

a consensus loss function, or at least of one that would be Pareto
optimal as between two disagreeing groups.

But as a practical matter,

it is probably easier for such a group to arrive at agreement on some­
thing on which they are compelled to take immediate action such as
bank reserves, or the money supply, or the Federal funds rate, than
about desirable conditions in the economy over which they have no
immediate control.

This leads me to my next topic, the role of inter­

mediate targets.
Intermediate targets and other recent developments
in monetary policy
In an optimal control framework, it is argued, there is no
need for intermediate targets.
the loss function.

It is the ultimate goals that go into

The pursuit of these ultimate targets by means of

intermediate targets such as money supply or interest rates, it is
argued, is in theory suboptimal.

Their function is to serve as

-5information variables from which insights
real sector can be gathered.

into developments in the

The central bank should look at them

as it should look at other readily observable data —


should be looked at as a source of information about real develop­
ments that are not directly or not frequently observable.
This line of argument is in conflict, of course, with some
of the main developments in monetary policymaking in the United States
during the last 10 or 15 years.

There was a time when the Federal

Reserve indeed did "look at everything."

It was not a technique

lending itself to much precision, and the development of intermediate
targets was generally regarded as a step forward.
The advantages of intermediate targets have often been described
as those of better observability and better control.
larly the latter as significant.

I regard particu­

Moving directly from money market

conditions or bank reserves, which the central bank controls, to the
ultimate targets of growth, employment and price stability conveys
very little of a sense of the quantitative impact of monetary policy

At best, one can have a sense of the direction of policy,

and even there, as we have learned, mistakes are possible if the
central bank does not distinguish between endogenous and exogenous
movements in its policy variables.
long lags of policy action.

This uncertainty is enhanced by

By focusing on the money supply or on

interest rates, a better intuitive sense of the thrust of policy is

-6 -

likely to be achieved.

The sense of having some protection against

the extreme kind of error that might manifest it by extreme values
of these variables, can be helpful.
Policymakers may be reluctant to surrender this means of
obtaining some direct feel for the economic meaning of their actions
and to replace it by a system that tells them to move some variable
like unborrowed reserves, or the Federal funds rate, perhaps drastically,
in order to achieve some particular results in the real sector.
might come close to flying with an automatic pilot.


Manual control,

I believe, would instill greater confidence.
Concern over possibly extreme values of policy variables is
likely to be another obstacle to the greater use of optimal control
that will have to be dealt with.

The experience of the postwar period

seems to show that extreme settings of policy variables, even for
relatively short periods, can be destabilizing.

Frequent variation in

instrument settings likewise may add to instability.

Given the lags

and the uncertainties, moderate instrument settings and a degree of
steadiness seems preferable most of the time, quite aside from
possible side effects of wide policy gyrations on the functioning of
financial markets#

Policymakers who do have such preferences for

moderation and steadiness can, of course, put them into their loss
function, as the paper by Kalchbrenner and Tinsley does.

But that

is only partial protection, unless the penalties assigned to wide
deviations in the paths of instrument variables are very high and

-7hence perhaps unreasonable.

Moreover, one may remain suspicious

about the properties of a feedback process that needs to be
disciplined in this artificial way.

The policy evaluation - - o r advice —

derived from an

optimal control system presumably is no better than the model through
which the feedback flows.

Policymakers are likely to be interested

in several aspects*
First, while models have reached a certain degree of
proficiency in short-run predictions under ordinary circumstances,
and in that sense agree with each other, there nevertheless seem to
be important differences.
among models.

Policy multipliers seem to vary importantly

Even within particular models, these multipliers seem

to be sensitive to small changes in the specification of particular
equations, or to the choice among alternative equations of seemingly
equal theoretical and empirical plausibility.

Chow*s paper suggests

a minimax strategy, choosing among the models on the basis of which
minimizes the worst case.

That procedure would provide some insurance,

but otherwise seems to adopt a rather pessimistic slant.


procedures might be to examine models for robustness of their policy
advice under varying assumptions, or perhaps to look for a policy
that is rpbust with respect to switches among models.

Clearly one

of the precautions policymakers would want to apply is to use a variety
of models.

But when there are significant differences, it is not easy

to work up much confidence.

-8Second, there is likely to be concern about the possibility
of bias in particular models or specifications of loss functions*
Ando-Palash point out that in a quadratic loss function, if the
target values for unemployment and inflation are set low, such as
at zero, the unemployment variable will obtain an unintendedly high

My own concern is that on the contrary an inflation bias

may enter the process, via the structure and the typical use made
of most models.

They seem to underestimate inflation because of a

questionable process of forming price expectations, which relies on
distributed lags of past experience instead of on rational expectations
based on observed government policy.

The paper by Kalchbrenner-Tinsley

makes reference to the severe underestimation of inflation.
Additionally, an inflation bias may appear if model simulations
are kept too short, since price effects typically lag volume effects.
Long model simulations into the future are not popular, owing to the
difficulty of estimating exogenous variables and perhaps also because
of the longer run instability of some models.

Nevertheless, by limiting

a simulation or forecast to only a few quarters ahead, the long-run
price effects may be cut off.

These effects then will carry less weight

in the optimal control simulation while volume effects, which occur with
less of a lag, dominate.
Third, still another question about the performance of presentday models relates to their abiJ^ty to deal with severe exogenous shocks
such as the devaluation

or the rise in oil prices.


-9effects, as Kalchbrenner-Tins ley put it, simply "had no place to
go in traditional econometric models,"

Now that the high rates of

unemployment and inflation that resulted have become part of the data,
users of the m o d e l s 1 output are likely to be concerned about possible
distortions from these outliers.
Policymakers may be concerned about the kind of advice they
are likely to get when the outlook appears more uncertain than usual.
At such times, anyone with a firm opinion is likely to carry dispro­
portionate weight, but in the case of advice from a model that is part
of the risk to be guarded against.

The natural tendency of policymakers,

under such conditions, will be in the direction of greater conservatism,
i,e,, to do "less" than they otherwise would.

The meaning of "less"

may not be the same for everyone, although technically it would seem
to imply that policy action should then be so designed as not to add
to the variance of the loss.

In practice, it may just mean to keep

doing whatever was being done before.
Among technicians, views do not seem to be unified concerning
the implication of varying uncertainty,

A well known theorem by

William Brainard states that, under specified conditions, uncertainty
reduces the scale of action.

Kalchbrenner-Tinsley, in an earlier paper,

seem to be of the same opinion.

I have heard others quoted to the

effect that uncertainty probably but not necessarily implies greater


In any event, the users of optimal control probably

would like to know whether they run the risk, under particularly
ticklish conditions, of being confronted with extreme advice from
this source as they often are from other sources as well.
The outcome of the policy simulations
Ando-Palash and Kalchbrenner-Tinsley have very properly
indicated to what extent their findings point to alternative ex ante
policies that in the light of contemporaneous information could have
been adopted to produce better results.

Ex post simulations, employing

information that policymakers did not have at the time, may provide
valuable lessons for the future but do not constitute a valid criticism
of past policies.

If I understand the two pairs of authors correctly,

they both claim that, with the benefit of hindsight, policy could have
been significantly improved.

Kalchbrenner-Tinsley also seem to find that,

on an jex ante basis, optimal control would not have done better.


backing, in other words, is not enough to produce better policies,
according to their findings.

I might add that in a set of papers by

Hyman-Shapiro and Hirsch that will be discussed this afternoon,
evaluating recent policies with the aid of eight alternative models
in an optimal control framework, the conclusion was reached that even
with the benefit of hindsight the inflation and recession of the last
few years could not have been avoided, although policy could have been
improved upon.