Much has been made of
the failure of modern macroeconomics to predict or understand the Great
Recession of 2007–2009. In this MACRO FOCUS, our resident time-series econometrician,
James Morley*, explains what is currently meant by
“modern” macroeconomics, what is behind its failure, and what can be done to
rehabilitate its reputation.
In a recent essay, Narayana Kocherlakota, President of the
Federal Reserve Bank of Minneapolis, acknowledged that modern macroeconomics
failed during the recent financial crisis.[1]
However, his essay misses the point of why it failed.
Like many in academia, Kocherlakota associates modern
macroeconomics with a particular school of thought that takes something called
the “Lucas critique” as its guiding principle. The Lucas critique refers to an
argument put forth by the Nobel Prize-winning macroeconomist Robert Lucas about
how the changing expectations of economic agents will confound forecasting and
policy analysis based on macroeconomic data.[2]
Its main implication is that an economic model with “deep structural
parameters” related to preferences and technology for households and firms
should provide more reliable forecasts, especially when predicting the effects
of policy, than a model based more on the apparent historical correlations
between macroeconomic variables. This is sometimes referred to as the
“microfoundations” approach to macroeconomics because it presumes that a
microeconomic structure — in particular, the metaphor of optimizing economic
agents— is more robust to changes in the policy environment than macroeconomic
correlations.[3]
Rather than question the relevance of the Lucas critique,
Kocherlakota explains the recent failure of modern macroeconomics as due to
much narrower issues. In his view, the micro-founded models failed because they
lack sufficient complexity, especially in terms of their treatment of financial
markets. Also, he points out, rightly, that the models are driven by “patently
unrealistic shocks”.
However, the rehabilitation of modern macroeconomics requires a
different tack than suggested by Kocherlakota. In particular, macroeconomists
need to do more than simply add complexity to their models. They should also
remember that it is empirically testable whether models that put most of their
weight on “deep structural parameters” produce more accurate predictions than
models that put more weight on historical correlations. In doing so, it may be
found that some macroeconomic relationships are useful even if they cannot be
easily motivated as the literal outcome of a micro-founded model. This is not
to deny the important role that economic theory plays in the formulation of
models used for explanation, prediction, and policy analysis. However, there is
no reason for models to take theory quite so literally as is typically done in
modern macroeconomics. Instead, the data should be taken more seriously.
History
Repeats Itself
The idea of the Lucas critique arose out of the 1970s. It was a time
when large-scale macroeconometric models that relied heavily on historical
correlations — especially the traditional Phillips curve tradeoff between
unemployment and inflation — failed to predict or even explain “stagflation” in
the form of simultaneously high rates of unemployment and inflation.
After their failure in the 1970s, the large-scale macroeconometric
models were modified to include supply shocks and changes in inflation
expectations. With these additions, the models could explain stagflation ex post. But their ex ante failure represented a serious blow to their reputation,
especially within the ivory towers of academia.
Ironically, this historical episode should remind us somewhat of the
present. Now it is “dynamic stochastic general equilibrium” (DSGE) models
inspired by the Lucas critique that have failed to predict or even explain the
Great Recession of 2007–2009. More precisely, the implicit “explanations” based
on these models are that the recession, including the millions of net jobs
lost, was primarily due to large negative shocks to both technology and
willingness to work. In his essay on modern macroeconomics, Kocherlakota admits
the inadequacy of these explanations:
… most models in macroeconomics
rely on some form of large quarterly movements in the technological frontier
(usually advances, but sometimes not). Some models have collective shocks to
workers’ willingness to work. Other models have large quarterly shocks to the
depreciation rate in the capital stock (in order to generate high asset price
volatilities). To my mind, these collective shocks to preferences and
technology are problematic. Why should everyone want to work less in the fourth
quarter of 2009?…
So can the reputation of modern macroeconomics be rehabilitated by
simply modifying DSGE models to include a few more realistic shocks? As
discussed below, the problems for DSGE models run deeper than a lack of
complexity, while the large-scale macroeconometric models have improved
considerably since the 1970s.
The Rabbit and the Hat
A simple example helps illustrate for the uninitiated just how DSGE
models work and why it should come as little surprise that they are largely
inadequate for the task of explaining the Great Recession.
For this simple DSGE model, consider the following technical
assumptions: i) an infinitely-lived representative agent with rational
expectations and additive utility in current and discounted future log
consumption and leisure; ii) a Cobb-Douglas aggregate production function with
labor-augmenting technology; iii) capital accumulation with a fixed
depreciation rate; and iv) a stochastic process for exogenous technology
shocks.
Before discussing the particular implications of this model, it is
worth making two basic points about the setup. First, by construction,
technology shocks are the only underlying source of fluctuations in this simple
model. Thus, if we were to assume that U.S. real GDP was the literal outcome of
this model, we would be assuming a priori
that fluctuations in real GDP were ultimately due to technology. When faced
with the Great Recession, this model would have no choice but to imply that
technology shocks were somehow to blame.[4]
Second, despite the underlying role of technology, the observed fluctuations in
real GDP can be divided into those that directly reflect the behavior of the
exogenous shocks and those that reflect the endogenous capital accumulation in
response to these shocks.
To be more precise about these two points, it is necessary to assume
a particular process for the exogenous technology shocks. In this case, let’s
assume technology follows a random walk with drift:
So, with this simple DSGE model and for typical measures of the
capital share, we have the implication that output growth follows an AR(1)
process with an AR coefficient of about one third. This is notable given that
such a time-series model does reasonably well as a parsimonious description of
quarterly real GDP dynamics for the U.S. economy. In particular, the estimated
AR coefficient for different sample periods of postwar quarterly U.S. real GDP
growth is fairly robust around 0.33 and such a model does well in terms of
various model selection criteria.
At this point, it is tempting to recall Milton Friedman’s classic
essay on methodology in which he argued that we should not judge a model based
on its assumptions, but instead we should focus on its predictions.[6]
However, the rather absurd assumption of a 100% depreciation rate at the
quarterly horizon would surely still have prompted a sharp question or two in a
University of Chicago seminar back in the days. So, with this in mind, what
happens if we consider the more general case?
Unfortunately, for more realistic depreciation rates, we cannot
solve the model analytically. Instead, taking a log-linearization around steady
state, we can use standard methods to solve for output growth as
However, let’s take moment to reflect on this result. This simple
DSGE model is able to mimic the apparent AR(1) dynamics in real GDP growth. But
it does so by assuming the exogenous technology shocks also follow an AR(1) process
with an AR coefficient that happens to be the same as the estimated AR
coefficient for output growth. Thus, the magic trick has been revealed: a
rabbit was stuffed into the hat and then a rabbit jumped out of the hat.[8]
Ever-Increasing
Sophistication?
The simple DSGE model in the previous section may be familiar to
readers under its alternative label of a “real business cycle” (RBC) model.
Modern DSGE models typically have more complicated preferences (e.g., habit
formation), production technologies, and market structures than the original
RBC models. Frictions such as sticky prices even open up the possibility of
using these models for monetary policy analysis, something that RBC models
assume a priori doesn’t really matter
for real economic activity.
However, despite their increasing sophistication, DSGE models share
one key thing in common with their RBC predecessors. After more than two
decades of earnest promises to do better in the “future directions” sections of
academic papers, they still have those serially-correlated shocks.[9]
Thus, the models now “explain” variables like real GDP, inflation, and interest
rates as the outcome of more than just serially-correlated technology shocks.
They also consider serially-correlated preference shocks and serially-correlated
policy shocks.
Revisiting Friedman’s essay on methodology, it should be recalled
that explaining economic phenomena as being due to changing preferences is
basically the nightmare scenario for economists, as it implies an eventual
ascendancy of sociology over economics. Meanwhile, in terms of policy, it is
important to recognize the following fundamental issue with the DSGE framework.
Given the assumption that the data are the literal outcome of the model, it
will be the case, by construction, that policy is to blame for the severity of
the welfare loss relative to a frictionless equilibrium implied by technology
and preference shocks. For example, if policy in the model is monetary policy,
then any bad outcome that is not directly due to a negative technology shock or
a preference shock — which is harder to label as being either “good” or “bad” —
must be due to suboptimal monetary policy failing to mitigate the effects of
frictions, at least according to the model.
The example of a state-of-the-art DSGE model mentioned in
Kocherlakota’s essay on modern macroeconomics is a “New Keynesian” model of the
European economy developed by Frank Smets and Raf Wouters.[10]
On one level, their model is much more impressive than the RBC model in the preceding
section. It has seven observable variables and ten different types of
structural shocks, not all of which are serially correlated. The model
incorporates sticky prices, sticky wages, habit formation, costs of adjustment
in capital accumulation, and variable capacity utilization. Perhaps most
impressively, Smets and Wouters estimate rather than calibrate their model,
which had been the standard practice in the RBC literature.
Much has been made of the fact that Smets and Wouters’ estimated
DSGE model appears to forecast macroeconomic data almost as well as a “vector
autoregressive” (VAR) model. But before getting too carried away, there are a
few issues that should be considered.
First, like the RBC model in the preceding section, Smets and
Wouters’ model is solved by linearizing around steady state. This might sound
like a minor technical detail, but it has a direct practical consequence that
the model needs to be estimated using deviations from steady-state levels of
variables such as real GDP. Also, the predictions of the model are for the
deviations from steady state rather than the levels of the variables
themselves.
Smets and Wouters “solve” the problem of measuring deviations from
steady state by considering deviations of variables from their sample means or,
in some cases, from estimated linear time trends. Other papers in this
literature consider Hodrick-Prescott (HP) or Bandpass (BP) filtered data. From
an econometric point of view, the possible presence of unit roots (a.k.a.
“stochastic trends”) in macroeconomic variables raises strong concerns that
these procedures will lead to inconsistent parameter estimates and inaccurate
out-of-sample forecasts, even in the unlikely event that the theoretical model
is correctly specified.[11]
Second, even ignoring the statistical issues involved in measuring
deviations from steady state, there is still the underlying economic problem
that Smets and Wouters’ model relies on serially-correlated shocks for
productivity, the inflation target, consumer preferences, government spending,
labor supply, and investment. All of these shocks are assumed to follow AR(1)
processes, with posterior mean estimates of the AR coefficients ranging from
0.81 to 0.94 for the demeaned and linearly-detrended quarterly data.
The DSGE model places restrictions on the mapping of the
serially-correlated shocks to the observable variables. However, it is not
clear what role the shocks play in “explaining” the observable variables
relative to the endogenous economic mechanisms. Indeed, at one point in their
paper, Smets and Wouters ponder as follows: “[I]t would be interesting to see
which of the various frictions are crucial for capturing the persistence and
covariances in the data”. However, what they neglect to mention is that the answer
could very well be “none”. In reality, it is quite possible the persistence and
covariances in the data are being captured by what amounts to a statistical
factor model. This could be tested by comparing the forecasting ability of the
DSGE model to a parsimonious “atheoretical” dynamic factor model. Instead, the
focus is on a comparison to VAR models. But VAR models have many parameters
(especially when there are large number of variables) and, as acknowledged by
Smets and Wouters, may be suffering for their heavy-parameterization given the
short sample period for European data.[12]
Meanwhile, Smets and Wouters readily acknowledge that some
of their shocks may be serving as proxies for important, omitted macroeconomic
phenomena: “Of course, these shocks could capture a whole range of
shocks that are not accounted for in the stylised model such as changes in oil
prices, terms-of-trade shocks, changes in taxes, etc.” This simply begs the
question of why not consider models, such as large-scale macroeconometric
models, that directly incorporate such variables.
The
Lucas Critique and Large-Scale Macroeconometric Models
Before discussing the details of large-scale
macroeconometric models, it is perhaps useful to revisit Kocherlakota’s essay
on modern macroeconomics. He writes:
In terms of fiscal policy
(especially short-term fiscal policy), modern macro modeling seems to have had
little impact. The discussion about the fiscal stimulus in January 2009 is
highly revealing along these lines. An argument certainly could be made for the
stimulus plan using the logic of New Keynesian or heterogeneous agent models.
However, most, if not all, of the motivation for the fiscal stimulus was based
largely on the long-discarded models of the 1960s and 1970s.
The “long-discarded models” that Kocherlakota has in mind are
large-scale macroeconometric models and the natural question that arises when
reading his last sentence is “discarded by whom?” Certainly, it is not by
policymakers trying to predict the quantitative effects of fiscal stimulus.
The answer to “discarded by whom?” that Kocherlakota clearly has in mind
is “modern” macroeconomists, by which he means academics who take the Lucas
critique as the central tenet of macroeconomics. Here again Kocherlakota
writes:
The macro models used in
the 1960s and 1970s were based on large numbers of interlocking demand and
supply relationships estimated using various kinds of data. In his powerful
critique, Lucas demonstrated that the demand and supply relationships estimated
using data generated from one macroeconomic policy regime would necessarily
change when the policy regime changed. Hence, such estimated relationships,
while useful for forecasting when the macro policy regime was kept fixed, could
not be of use in evaluating the impact of policy regime changes.
An immediately noticeable thing about this quote is the deliberate use
of the words “policy regime” rather than just “policy”. For example, consider
monetary policy. This distinction would be between a change in the Federal
Funds Rate (which is a change in “policy”) versus a change in the coefficients
for a Taylor-type rule (which is a change in “policy regime”). The choice of
words is relevant because the previous quote about short-term fiscal policy
would seem to be about a change in policy, not necessarily a change in policy
regime. Thus, at least according to Kocherlakota’s qualification about
forecasting in a fixed regime, the estimated relationships considered in
large-scale macroeconometric models may well be useful for predicting the
effects of the stimulus plan.
Another noticeable thing about Kocherlakota’s discussion of the Lucas
critique is that he presents it as some sort of universal truth that estimated
demand and supply relationships will be unstable and of limited use for policy
analysis. This might be valid if a DSGE model were reality. But DSGE models are
models, not reality. Thus, the relevance of the Lucas critique is testable and the
tests have not been favorable.[13]
Meanwhile, according to a meta-critique of the Lucas critique by Christopher
Sims, the lack of practical relevance should come as no surprise.[14] He writes:
The only
coherent interpretation of the Lucas critique is that it states that if one
uses a model which incorrectly describes the reaction of expectations formation
to policy choice, it will produce incorrect evaluations of policy. The
implication is not that econometric evaluation of policy using models fitted to
history is impossible, but that it requires correct specification of the reaction
of the economy to policy… There may be some policy issues where the simple
rational expectations policy analysis paradigm – treating policy as given by a
rule with deterministic parameters, which are to be changed once and for all,
with no one knowing beforehand that the change may occur and no one doubting
afterward that the change is permanent – is a useful approximate simplifying
assumption. To the extent that the rational expectations literature has led us
to suppose that all “real” policy change must fit into this internally
inconsistent mold, it is has led us onto sterile ground.
Thus, contrary to the precepts of “modern” macroeconomics, the Lucas
critique in no way proves that DSGE models will predict the effects of policy
better than large-scale macroeconometric models based more on historical
correlations.
However, to avoid getting lost in a long, fruitless debate over the
semantics of the Lucas critique, it is perhaps more constructive to simply
review the features of modern macro models that Kocherlakota argues have been
inspired by it. In his words, modern macro models have the five following
properties:
- They specify budget constraints for households, technologies for firms, and resource constraints for the overall economy.
- They specify household preferences and firm objectives.
- They assume forward-looking behavior for firms and households.
- They include the shocks that firms and households face.
- They are models of the entire macroeconomy.
The thing that is notable about this list is that these features are far
from the sole domain of DSGE models. They are also present in contemporary
versions of the large-scale macroeconometric models of the U.S. economy such as
those developed by Macroeconomic Advisers (WUMM/MAUS), the Federal Reserve Board
(FRB/US), and the Bank of Canada (MUSE). In this sense, it might well be
reasonable to acknowledge the impact of some interpretations of the Lucas
critique on macroeconomics. However, it is quite a leap to go from this
acknowledgment to discarding all but DSGE models, as Kocherlakota would seem to
have us do.
So what about contemporary large-scale macroeconometric
models? They have a number of advantages over DSGE models. First, there are
many more variables in the models. This allows for more useful details that are
typically missing from DSGE models, from “small” things like the consideration
of different types of consumption (e.g., durables vs. nondurables and services)
and different forms of fiscal policy (i.e., more than just lump-sum transfers),
to larger things like foreign trade.[15]
Second, the models consider levels data rather than deviations from steady
state. This is helpful for statistical and economic identification, as well as
for forecasting. Third, the models are grounded in macroeconomic theory, but
they are not intended to be a literal description of reality.
An example might help illustrate the nature of large-scale
macroeconometric models. The model developed by Macroeconomic Advisers is based
in part on the life-cycle hypothesis in which households are forward-looking
and smooth their consumption across their lifetime income profiles. This
theoretical setting implies a consumption function in which the marginal
propensities to consume can be thought of as complicated functions of “deep structural
parameters”.[16] A key point
is that the estimates for the marginal propensities to consume are remarkably
stable over the postwar period, implying that the deep structural parameters
for this model are also fairly stable.
The Macroeconomic Advisers model is estimated using the
directly observable data rather than deviations from steady state. One possibly
surprising benefit of this is that estimation of the model, which takes unit
roots and cointegration into account, makes use of permanent variation to help
identify parameter values. Also, exogeneity restrictions which might be
debatable for short-run fluctuations are arguably more reasonable in terms of
long-run variation. Thus, long-run variation helps with both statistical and
economic identification.
A more obvious benefit of using the levels data is that the
Macroeconomic Advisers model provides direct forecasts of the data rather than
deviations from steady state. Meanwhile, the model allows for relatively
flexible short-run dynamics. Thus, the macroeconomic theory pinning down
long-run relationships and some of the adjustment dynamics provides a guide to
explaining the data, but theory is not taken quite so literally as it is in
estimated DSGE models.[17]
As a key concession to reality, the Macroeconomic Advisers
model allows for residuals. Consequently, when faced with the Great Recession,
the model did not explain 100% of the movements in real GDP. Although the
consumption function in the model implied a decline in economic activity in
response to the dramatic declines in household wealth associated with the
collapse of the housing market and then financial markets, the model — along
with pretty much every other pre-existing quantitative model — did not quite
predict the severity of the recession. But this is surely better than
pretending that the model explains 100% of real economic activity by relying on
serially-correlated technology and preference shocks.
Indeed, the Macroeconomic Advisers model can capture the
effects of wealth on consumption more accurately because it does not assume that the model must explain 100% of real
economic activity. The residuals act as a kind of “safety pressure relief
valve” to address the fact that models are not reality. The notion that
estimated DSGE models are too literal corresponds to the idea that they have no
such safety valve.[18] A
consequence is that, as discussed in more detail below, estimates of the
supposed deep structural parameters for DSGE models can actually be fairly
sensitive over time, including to changes in policy regimes.
So, returning to Kocherlakota’s requirements for a “modern
macro model”, the treatment of households in the Macroeconomic Advisers model
i) includes intertemporal budget constraints, ii) specifies preferences for
households that appear to be stable, iii) assumes households are forward
looking, iv) includes realistic shocks that households face such as different
types of fiscal interventions, and v) is part of a model of the entire economy.
Evidently, even though the model differs from DSGE models in terms of certain
assumptions and general implementation, it has a claim to being “modern” by
Kocherlakota’s criteria.[19]
Can’t We All Just Get
Along?
In some sense, the various approaches to macroeconomics are
not really as different as they are sometimes made out to be. Although Finn
Kydland and Edward Prescott, the Nobel Prize-winning gurus of the RBC camp,
once wrote dismissively of a “system-of-equations” approach,[20]
the reality is that VARs, large-scale macroeconometric models, and DSGE models
all imply systems of equations. Policy forecasts for these different approaches
are all based on some assumptions from macroeconomic theory and some
consideration of how economic agents perceive a given change in policy — i.e.,
was it anticipated or unanticipated and will it be permanent or transitory? The
main differences across approaches are in terms of how estimation is carried
out and how the theoretical assumptions are imposed. The VAR places the least
(but not zero) weight on theory, while the DSGE models place the most, even to
the extent of imposing strong restrictions on some parameters across equations.
It is ultimately an empirical question as to whether the imposition of these
cross-equation restrictions really helps with predicting the effects of policy
and forecasting more generally.[21]
An important issue, related to the Lucas critique, is the
stability of estimates over time. For any model, parameter estimates will
sometimes change, some more than others. For example, time-series models of
real GDP growth typically have fairly stable estimates for parameters related
to dynamics, but variance estimates have changed greatly since the onset of
Great Moderation in the mid-1980s. However, the possibility of time-varying
parameters hardly invalidates the use of a given model for forecasting or
policy analysis. It merely begs the question of how adept estimates for that
model are at tracking parameters that can generate accurate predictions in real
time. In macroeconomics, parameter stability is a relative concept, not an
absolute.
As an example, one of the most prominent controversies
throughout the history of macroeconomics has been over the stability of the
Phillips curve. This was certainly the example from Robert Lucas’s original
critique that stuck in the collective consciousness of academic
macroeconomists. One reason is that estimates of Phillips curve parameters have
indeed changed over time. However, the notable thing is that this is also true
about the estimates of the supposed “deep structural parameters” of DSGE models
that generate a Phillips curve. For example, Luca Benati finds that estimates
of a key structural parameter that determines the New Keynesian Phillips curve
is not at all robust across different monetary policy regimes.[22]
Meanwhile, instabilities in Phillips curve parameters do not appear to
translate into instabilities in other macroeconomic correlations. It might be
argued that this is due to low statistical power. But, contrary to this view,
it is notable, for example, just how robust point estimates for consumption
functions are to different sample periods. Thus, macroeconomists should pursue
stable relationships as much as possible, but they should not assume a priori that a DSGE model is where they
will find them. In the words of Robert Lucas and Thomas Sargent, “[T]he
question of whether a particular model is structural is an empirical, not
theoretical, one.”[23]
Doing Better in the Future
It is a safe bet that future versions of DSGE models will
incorporate more complicated financial sectors and allow for different types of
fiscal policies. And guess what? The new-and-improved DSGE models will turn out to
imply (ex post) that the Great
Recession was actually due to serially-correlated financial intermediation
shocks and suboptimal fiscal policy.
Alas, these conclusions will be driven much more by the DSGE
framework than by the data. In general, the implications of DSGE models
for policy are more assumed than estimated. For example, the typical assumption
is that households are “Ricardian” in the sense that they are infinitely-lived
and anticipate that the government’s intertemporal budget constraint will
always hold. Thus, households will perceive a current increase in government
spending as being offset by a future increase in taxes. This clearly places
strong restrictions on estimates of fiscal multipliers — i.e., by construction,
everything is a “balanced-budget” multiplier. In some recent DSGE models, a
large fraction of households are assumed to be “hand-to-mouth” in the sense
that they consume all of their disposable income immediately. Although this
assumption represents somewhat of a departure from the original intentions of
the microfoundations agenda, it appears to be necessary to replicate less restrictive
estimates of fiscal multipliers from VAR models.[24]
However, rather than just sticking to the same basic framework with
relatively small changes in assumptions to help fit the data, modern
macroeconomists should also focus on a more fundamental issue: a “shock” should
be a surprise — i.e., a given shock process should really be serially
uncorrelated within a fully-specified theory. At the very least, the analysis
of DSGE models should be geared much more towards convincing a skeptic that
results are being driven by endogenous economic mechanisms that are consistent
with data, not by assumed exogenous processes or, more subtly, by an absence of
residuals that could help cope with model misspecification of the real world.
In general, promoters of DSGE models need to convince non-believers
that estimates are robust across policy regimes in the sense of producing
better forecasts than other models in changing policy environments. Although it
is certainly true that estimated models can “over-fit” in sample, out-of-sample
comparisons address this problem. Indeed, even if models happen to fit equally
well in sample, they will, if based on different economic assumptions, produce
different out-of-sample predictions for the effects of policy. The bottom line then
is that DSGE models should be subject to the same (market-based) forms of
evaluation that large-scale macroeconometric and other forecasting models have
been subject to — i.e., they need to forecast well in real time.
Meanwhile, it should be acknowledged that, despite including oil
prices and credit conditions, large-scale macroeconometric models didn’t
exactly predict the severity of the Great Recession of 2007–2009. There were
still sequences of negative forecast errors for various measures of real economic
activity. More consideration of nonlinear time-series dynamics could
potentially help on this front.[25]
However, in terms of really predicting the crisis, the award obviously goes to
theories of endogenous financial crises inspired by the ideas of Hyman Minsky.
Formal evaluation of these more narrative approaches is hard and there may be
an element of the “stopped-clock syndrome” at play. But it would be foolish to
dismiss such theories out of hand. In particular, a ludicrous notion sometimes
expressed in the ivory towers of academia is that, for Minsky to be taken
seriously, his ideas need to be put into a DSGE model.[26]
Instead, the converse is true. For DSGE models to be taken more seriously
outside of academia, they need to explain and predict as well as Minsky. And
serially-correlated preference and technology shocks aren’t going to do it!
To be critical of the Lucas critique is not to say it is completely
irrelevant. An important goal of macroeconomic models is to have stable
parameters given changes in the policy environment. But how we get there is not
necessarily through a particular class of micro-founded models. More broadly, if
macroeconomists want to regain the trust of the public at large, they need to
resist the notion that “macroeconomics” is defined as a method, rather than as
a subject matter. Specifically, macroeconomists need to be more pluralistic.
They should draw from different types of analysis, be it time-series models,
large-scale macroeconometric models, DSGE models, and more narrative
approaches. Ultimately, we will know that the reputation of
macroeconomics has been rehabilitated when “modern macroeconomics” is no longer
used as a label for a particular school of thought, but instead refers to a
body of knowledge of substantive and useful insights into how the macroeconomy
actually works and what will happen to it in the future.
* Without
implicating anyone, the author wishes to thank Joel Prakken, Ken Matheny,
Antulio Bomfim, Alan Auerbach, Tara Sinclair, Marc Law, Jeremy Piger, Steve
Fazzari, and Sharon Kozicki for helpful and critical comments.
[2] Lucas, R.E.,
1976, “Economic Policy Evaluation: A Critique,” Carnegie-Rochester Conference Series on
Public Policy 1, 19–46.
[3] The
optimizing agents in a micro-founded model are a metaphor because of the
insurmountable problems involved in aggregating across actual households and
firms. On this issue, the economist/philosopher Kevin Hoover has written many
compelling critical analyses of the microfoundations approach to
macroeconomics. For an example that is particularly relevant to Kocherlakota’s
essay, see Hoover, K., 2006, “A NeoWicksellian in a New Classical World: The Methodology
of Michael Woodford’s Interest and Prices,” Journal of the History of Economic Thought 28,
143-149, [pdf].
[4] To the
extent that other shocks, such as exogenous changes in financial conditions or
oil prices, were actually to blame, they would be labeled by the model as
technology shocks. In particular, the model would be misspecified and its
estimates for the effects of technology shocks would be distorted.
[5] For more details of this
derivation, see McCallum, B., 1989, “Real Business
Cycle Models,” in R.J. Barro, ed., Modern Business Cycle Theory, Harvard
University Press, and Ma, J. and M. Wohar, 2009, “Real and Nominal Business
Cycles: New Evidence from a Generalized Unobserved Components Model,” Working
Paper.
[6] Milton
Friedman, 1953, “The Methodology of Positive Economics” in Essays in
Positive Economics, The University of Chicago Press.
[7] Cogley, T.
and J.M. Nason, 1993, “Impulse Dynamics and Propagation Mechanisms in a Real
Business Cycle Model, Economics Letters 43,
77-81 and Cogley, T. and J.M. Nason, 1995a, “Output Dynamics in
Real-Business-Cycle Models,” American Economic
Review 85, 492-511.
[8] It is
conceivable that technology actually evolves according to the same AR(1)
process as output growth. But even if this is so, there is no sense in which
the behavior of technology is being explained by the endogenous economic
mechanisms of the model (i.e., capital accumulation). Thus, in essence, the
model provides little or no additional insight into the behavior of real GDP
beyond that of a univariate AR(1) time-series model for output growth.
[9] As an early
example of expressed concern about relying on serially-correlated shocks,
consider the following discussion from a classic paper in the RBC literature by
Robert King, Charles Plosser, and Sergio Rebelo: “But along other dimensions,
the basic model seems less satisfactory. In particular, the principle serial
correlation in output – one notable feature of economic fluctuations – derives
mainly from the persistence of technology shocks.” See King, R.G., C.I.
Plosser, and S.T. Rebelo, 1988, “Production, Growth, and Business Cycles: I.
The Basic Neoclassical Model,” Journal of
Monetary Economics 21, 195-232.
[11] Without
taking too long of a digression into the literature on trend/cycle
decomposition and spurious cycles, it is worth mentioning another paper by
Timothy Cogley and James Nason. They consider the implications of HP filtering
for RBC models, but their conclusions apply equally well for linear detrending
and BP filtering for DSGE models: “When applied to integrated processes, the HP
filter can generate business cycle periodicity and comovement even if none are
present in the original data…We show that RBC models can exhibit business cycle
dynamics in HP filtered data even if they do not generate business cycle
dynamics in pre-filtered data. The combination of a unit root or near unit root
in technology and the HP filter is sufficient to generate business cycle
dynamics. Propagation mechanisms are unnecessary, and in many RBC models they
do not play an important role.” See Cogley, T. and J.M. Nason, 1995b, “Effects
of the Hodrick-Prescott Filter on Trend and Difference Stationary Time Series:
Implications for Business Cycle Research,” Journal
of Economic Dynamics and Control 19, 253-278.
[12] There is a
more subtle and technical issue with the model comparison. Smets and Wouters
consider Bayesian estimation and use “Bayes factors” to evaluate the relative
forecasting ability of models. Bayes factors are, by construction, sensitive to
priors. On the one hand, macroeconomic theory provides a sensible way to
formulate priors about the parameters of the model. On the other hand, if
priors are being informed by the data, the exercise is one of so-called
“empirical Bayes”, which is highly problematic (by putting different weights on
data than what is implied by Bayes rule). In the literature on estimated DSGE
models, the priors are often influenced by what has been found in other studies
that consider the same data. This will tend to bias results in favor of DSGE
models relative to models with less informative priors.
[13] See, for example,
Favero, C. and D. Hendry, 1992, “Testing the Lucas Critique: A
Review,” Econometric Reviews 11,
265-306. Meanwhile, the implied change in demand and supply relationships given
a change in policy regime may not even occur within
DSGE models when allowing for indeterminate equilibria. See
Farmer, R., 2003, “Why Does Data Reject the
Lucas Critique?” Annales d’économie et de statistique special issue on
“The Econometrics of Policy Evaluation”, 67-68, [pdf]. As
for the economic relevance of the Lucas critique under the assumption that a
DSGE model is reality, there are different views. Glenn Rudebusch argues that
the Lucas critique does not invalidate the use of reduced-form models to
forecast aggregate output and inflation, while Thomas Lubik and Paolo Surico
argue that it is more relevant when considering the effects of policy regimes
on macroeconomic volatility. See Rudebusch, G.D., 2005, “Assessing the Lucas
Critique in Monetary Policy Models,” Journal
of Money, Credit, and Banking 37, 245-272, [pdf]
and Lubik, T.A. and P. Surico, 2010, “The Lucas Critique and the Stability of
Empirical Models,” Journal of Applied
Econometrics 25, 177-199, [pdf].
[15] There, of
course, exist DSGE models that incorporate some of these additional details
found in large-scale macroeconometric models. But for reasons of computational
tractability, they never come close to incorporating all of the details at the
same time. As for the notion that DSGE models can, in principle, be made as complicated
as necessary to capture reality, there is a problem that the microfoundations
are likely to be so complicated that the models would never be operational for
forecasting or policy analysis. For example, the large disconnect between the
complicated micro-founded models on the existence of money and the practice of
monetary policy comes to mind. Indeed, by focusing on a micro-founded
“medium-of-exchange” role for money, the models may be reducing rather than
increasing their links to reality. See Goodhart, C.A.E., 2009, “The Continuing
Muddles of Monetary Theory: A Steadfast Refusal to Face Facts,” Economica 76, 821-830, [pdf].
[16] Prakken,
J., 2002, “Deep Parameters and the Consumption Function,” Technical Note, Macroeconomic Advisers. In addition to
depending on the interest and tax rates, which are accounted for directly in
the estimated consumption function, the marginal propensities to consume also
depend on the rate of time preference, both the inter- and intra-temporal
substitution elasticities, the intensity factors in the utility function, the
expected rate of growth in the consumer’s real wage, the age of the consumer,
life expectancy, and retirement age.
[17] Somewhat
ironically, the literalism of estimated DSGE models for short-run dynamics does
not generally extend to long-run relationships. In particular, by estimating
linear time trends or filtering the data prior to estimation of the
log-linearized system, the implied long-run relationships between the levels
data can be contrary to any cointegration relationships implied by theory.
[18] Some
estimated DSGE models allow for “measurement error”. However, Smets and Wouters
celebrate the fact that their model does not rely on such shocks to “explain”
the data (while, of course, measurement error does not explain actual economic
activity). Meanwhile, simply interpreting the technology and other shocks
broadly does not serve the same purpose as including residuals. Again, as discussed
in footnote 4, this is because a model forces the same endogenous reaction to
all shocks given the same label. Thus, estimates of the endogenous reaction to
the true more-narrowly-defined shocks will be distorted.
[19] The Macroeconomic
Advisers model also implicitly assumes that firms are forward-looking in
their investment decisions. Of course, this distinction between implicit and
explicit assumptions, which is related to how literally the theory is taken, is
the most essential difference from the DSGE framework. Meanwhile, the Phillips
curve in the model makes use of survey measures of inflation expectations. This
is forward-looking, but does not require identification via rational
expectations in the sense of economic agents — unlike lowly econometricians — knowing
the exact structure of the economy, including all parameter values, and
observing all variables and shocks in real time. This assumption about
expectations is also an important difference from DSGE models.
[21] There is an
old argument dating back (at least) to the Cowles Commission in the 1940s and
1950s that if different models fit the data equally well, their relative
usefulness for policy evaluation can only be determined on a priori theoretical grounds. However, even if models are
observationally equivalent within a given sample period, the very fact that
they can produce different policy predictions implies that they will have
different out-of-sample forecasting performances (at least assuming the
forecasters can observe the policy interventions). Thus, an empirical
evaluation of the usefulness of a model for predicting the effects of policy is
possible by tracking its real-time out-of-sample forecasting performance over
periods in which policy has changed.
[22] Benati, L.,
2008, “Investigating Inflation Persistence across Monetary Regimes,” Quarterly Journal of Economics 123,
1005-1060, [pdf].
The parameter is the “indexation” parameter for Calvo-style pricing that
corresponds to the fraction of firms that cannot re-optimize prices in a given
time period. Also see Glenn Rudebusch’s analysis of the Lucas critique
mentioned in footnote 13 in which he discusses some other models with supposed
deep structural parameters that have turned out to be unstable.
[23] Lucas, R.E.
and T. Sargent, 1981, “After Keynesian Macroeconomics,” in Rational
Expectations and Econometric Practice, University of Minnesota Press,
295-319. This quote is, of course, somewhat disingenuous. Lucas and Sargent
emphasize that success in terms of fitting the data or general short-term
forecasting is no guarantee of success at forecasting conditional on a change
in the policy environment. They clearly see the failures of large-scale
macroeconometric models in the 1970s as empirical validation of this point and
strongly believe micro-founded models will provide stable parameters. But their
point about the empirical nature of whether a model is structural has often
been forgotten in modern macroeconomics.
[25] A young
macroeconomist who shall remain nameless once responded to evidence of
nonlinearity in the form of business cycle asymmetry by saying (without any
hint of embarrassment) that a DSGE model could easily explain such a phenomenon
— all that was necessary was to assume the appropriate nonlinear shock process!
[26] Consider
the following claim by V.V. Chari, Partrick Kehoe, and Ellen McGratten: “[M]ost
macroeconomists now
analyze policy using some sort of dynamic stochastic general equilibrium (DSGE)
model. This type of model can be so generally defined that it incorporates all
types of frictions, such as various ways of learning, incomplete markets,
imperfections in markets, and spatial frictions. The model’s only practical
restriction is that it specify an agreed-upon language by which we can
communicate, a restriction hard to argue with. An aphorism among
macroeconomists today is that if you have a coherent story to propose, you can
do it in a suitably elaborate DSGE model.” See Chari, V.V., P.J. Kehoe, and
E.R. McGratten, 2008, “New Keynesian Models: Not Yet Useful for Policy
Analysis,” Federal Reserve Bank of Minneapolis Staff Report, no. 409, [pdf].
Click here to contact Macroeconomic Advisers
