Information, Animal Spirits, and the Meaning of Innovations in Consumer Confidence

Information, Animal Spirits, and the Meaning of
Innovations in Consumer Confidence*



Robert B. Barsky†
University of Michigan and NBER

Eric R. Sims‡
University of Michigan

September 3, 2008


Abstract


Innovations to measures of consumer confidence convey incremental information about economic activity
far into the future. Comparing the shapes of impulse responses to confidence innovations in the data with
the predictions of a calibrated New Keynesian model, we find little evidence of a strong causal channel
from autonomous movements in sentiment to economic outcomes (the “animal spirits” interpretation).
Rather, these impulse responses support an alternative hypothesis that the surprise movements in
confidence reflect information about future economic prospects (the “information” view). Confidence
innovations are best characterized as noisy measures of changes in expected productivity growth over a
relatively long horizon.


†: barsky@umich.edu
‡: ericsims@umich.edu










* The authors gratefully acknowledge helpful comments and suggestions from Susanto Basu, Alan Blinder, John
Cochrane, Olivier Coibion, Daniel Cooper, Mike Elsby, Yuriy Gorodinchenko, Christopher House, Peter Ireland,
Lutz Kilian, Miles Kimball, Guido Lorenzoni, Serena Ng, Matthew Shapiro, Robert Solow, Frank Vella, two
anonymous referees, and seminar participants at Boston College, Johns Hopkins, Michigan, the Wharton School, the
Russell Sage Foundation, and the NBER Summer Institute in Monetary Economics. We also thank John Fernald
for providing his quarterly utilization-corrected TFP measure. Robert Barsky acknowledges generous support as a
visiting scholar at the Russell Sage Foundation and the Bank of Israel.

---

I. Introduction
In the popular press and much of the business community it continues to be an article of
faith that “consumer confidence” has an important role – both prognostic and causal – in
macroeconomics. On the other hand, the stance of the rather limited academic literature on
confidence is far more ambiguous. The judgments range from the conclusion that confidence
measures have an important role both in prediction and understanding the cause of business
cycles, to the view that they contain important information but have little role in the assignment
of causality, to the verdict that they have no value even in forecasting.
There are, broadly speaking, two contrasting approaches to the role of confidence in
macroeconomics. The first, which we will refer to as the “animal spirits” view, posits
autonomous fluctuations in beliefs and consumption that in turn have causal effects on economic
activity. In the proceedings of a symposium on the causes of the 1990-1991 recession, both Hall
(1993) and Blanchard (1993) regard exogenous movements in consumption as a cause of
business cycles.1 Indeed, Blanchard proposes that the cause of the recession was a powerful,
long-lasting negative consumption shock associated with an exogenous shift in pessimism that
had a causal effect on consumption and overall aggregate demand. While not fully pursuing the
idea in his brief paper, Blanchard proposes that one might be able to test this hypothesis on the
basis of the observation that such an exogenous shift in pessimism ought to have only temporary
effects on consumption.2
The second view of confidence – what we will call the “information view” – suggests
that a relationship between innovations in measures of consumer confidence and subsequent
macroeconomic activity arises because confidence measures contain fundamental information
about the current and future states of the economy. For example, Cochrane (1994b) proposes
that consumption surprises proxy for news that consumers receive about future productivity that
does not otherwise show up in econometricians’ information sets. His attempt to reconcile VAR

1 In an interesting but almost forgotten early contribution, Hall (1986) – partially repudiating Hall (1978) – argues
that an important fraction of the random walk in consumption comes not from the expectational surprise in the Euler
equation but from a second disturbance that he has more recently referred to as “spontaneous consumption”. In Hall
(1993), this is interpreted as a shock to the taste for consumption relative to leisure.

2 In some ways, a limiting case of animal spirits appears in the “sunspot” literature. Though pinned down only by
extrinsic coordinating variables, expectations in the equilibria of these models are self-fulfilling, and thus not
irrational (see Farmer (1999)). The existence of sunspot equilibria depend on strong increasing returns, supply
externalities, or other mechanisms that are typically not accepted as empirically plausible. The notion of animal
spirits in this paper does not encompass sunspots.

1

---

evidence with theory closely anticipates the “news approach to business cycles” of Beaudry and
Portier (2004, 2006). They analyze models where agents become aware of changes in future
productivity orthogonal to current productivity, and argue that stock price innovations proxy for
future technological improvement not reflected in current technology. The “information view”
of confidence supposes that confidence innovations might contain similar information.
In Section II of the paper, we first show that unexplained innovations in several variables
representing survey responses to forward-looking questions from the Michigan Survey of
Consumers have powerful predictive implications for the future paths of macroeconomic
variables. In particular, within the context of augmented consumption-income VARs, we show
that unexplained innovations in the responses to several consumer confidence questions have
significant, slowly building, and apparently permanent implications for output and consumption.
Confidence is not highly Granger-caused by income or consumption, nor are its innovations
highly correlated with innovations in those variables. Responses to little-used survey questions
on “news heard” do help to somewhat explain confidence innovations, but with only a very
modest incremental R2. These observations point to the conclusion that these measures of
consumer confidence are not merely noise, nor are they simply reflections of macroeconomic
news reports or innovations in other variables with which they are correlated.
In Section III we attempt to distinguish the hypothesis that these impulse responses
indicate a causal channel from sentiment to economic outcomes (the “animal spirits” view) from
the alternative interpretation that the surprise confidence movements summarize information
about economic prospects known to consumers (the “information” view). To provide a
framework for distinguishing these alternative views of confidence, we present a highly stylized
New Keynesian model with three kinds of shocks. The first shock is an immediate and
unexpected improvement in productivity (a “level shock”). The second is a reflection of genuine
news that productivity will grow more rapidly for a substantial period of time into the future (a
“growth shock”, also to be referred to as an “information shock” because it conveys information
about future productivity that cannot be fully inferred from current productivity).3 We only
permit households to observe a noise-ridden signal of the information shock to technology. We
interpret the noise innovation in the signal as an “animal spirits shock” as it is associated with

3 We employ the term “information” in the same way Cochrane (1994b) and Beaudry and Portier (2006)
use the word “news”. Lorenzoni (2008), somewhat confusingly, uses the term “news” to refer to noise in a public
signal, which functions much like our animal spirits shock.

2

---

erroneous consumer optimism or pessimism. This shock can be given alternative less structural
interpretations, and in equilibrium its implications are similar to those of an exogenous
innovation to the Euler equation. Regardless of the particular interpretation, a series of positive
animal spirits shocks might capture the putative “irrational exuberance” of the 1920s or 1990s,
while a predominance of negative shocks would usher in a period of excessive pessimism.
The model has clear implications for the response of the endogenous variables to each of
the three shocks. “Animal spirits” shocks behave as aggregate demand shocks – they are
associated with transitory increases in output that attenuate over time, and they produce both
inflation and increases in real interest rates. “Information shocks” regarding future productivity
and shocks to current productivity are followed by gradual movements in the macroeconomic
variables that are not subsequently reversed. Both of these fundamental shocks are also
associated with rising real interest rates. Thus, the model yields two primary criteria by which to
distinguish animal spirits from fundamental shocks: positive animal spirits shocks are followed
by transitory movements in real activity and increases in inflation, while favorable fundamental
shocks may result in permanent movements in activity and may be either inflationary or
deflationary.
In Section IV, we estimate an expanded VAR with the variables implied by the model
augmented with a measure of confidence. As in the three variable systems of Section II, the
results show that confidence innovations are associated with little immediate response of real
activity but prolonged growth in consumption, income, and measured productivity. There is no
evidence of reversion in these variables – in particular, the point estimates suggest that income
and consumption are higher by more than two-thirds of a percent in the long future in response to
a confidence innovation, with the confidence bands associated with these impulse responses
lying above zero at horizons in excess of ten years. Confidence innovations are associated with
transitory increases in real interest rates and hours of work, and also lead to a large and persistent
reduction in inflation. These empirical responses are not at all similar to the implications of
animal spirits shocks in our model, nor are they particularly consistent with the theoretical
responses to level shocks.
We next postulate a structural equation in which surprise movements in confidence are
attributable to the signal agents receive about the growth rate and to the innovation in the current
state of productivity. We estimate a subset of the structural parameters of the model via a

3

---

modified version of simulated method of moments. We are able to resoundingly reject the
hypothesis that animal spirits shocks (as specified in this paper) are an important source of the
observed relationships between confidence innovations and macroeconomic variables. On the
other hand, we do find convincing evidence in favor of the information interpretation of
consumer confidence. The implications of confidence innovations for output and spending at
short horizons are far too small for confidence to be primarily a reflection of changes in current
fundamentals, yet the longer horizon implications are far too large and significant for confidence
innovations to not be conveying information about fundamentals. Our results suggest that there
are information shocks about future productivity not wholly reflected in current productivity, and
that these shocks account for a significant fraction of the innovation in measured confidence.

II. Income, Consumption, and Confidence
(a) Cochrane’s Bivariate VAR
We begin with the dynamics of income and consumption as implied by the bivariate
vector autoregression discussed by Cochrane (1994a). In particular, we estimate a two variable
system consisting of the log of real GDP and the log of real consumption of services plus non-
durables, both in per capita terms after dividing by the civilian non-institutionalized population
aged sixteen and over. The data are seasonally adjusted measures at a quarterly frequency from
the first quarter of 1960 to the third quarter of 2007. The data strongly suggest that the variables
are cointegrated, and the estimated cointegrating vector is sufficiently close to [1,-1] that we
follow Cochrane and others in imposing it. While popular information criteria generally favor a
small number of lags (one or two), we take a conservative stance and estimate the VAR with four
lags.
Cochrane orthogonalizes the innovations so that consumption is ordered first. This
ordering is implied by a simple permanent income model in which all information is immediately
reflected in consumption.4 However, the line of inquiry in his subsequent paper (Cochrane
(1994b)) suggests a focus on the alternative ordering with income first; there the focus is on the
information about future income embodied in consumption but not in current income. Figure 1

4 The consumption ? income ordering also splits income fairly neatly into permanent and transitory components,
as any innovation to income not reflected in consumption ought to be transitory under a partial equilibrium view of
the PIH.


4

---

presents impulse responses under both orderings, with the solid line referring to the ordering
with consumption first and the dashed line to the orthogonalization with income ordered first.
The key feature of these impulse responses is that innovations in consumption – whether or not
they are orthogonalized with respect to income – are associated with powerful and prolonged
subsequent increases in income. At the shorter horizons, most of the movement in income is
explained by its own innovation, but the “effects” of a consumption innovation build over time
so that much or all of the permanent component of GDP appears to be captured by innovations in
consumption. In short, results from this two variable VAR suggest that “consumption shocks”
convey news about income many periods into the future.
As Cochrane (1994b) stresses, a natural explanation for the finding that consumption
innovations predict much of future output is that agents have some advance knowledge about
future income that they use when making consumption decisions. Forward-looking questions on
surveys of consumer expectations and attitudes might potentially provide a direct measure of
such information, and thus a direct test of Cochrane’s hypothesis. Is much or most of the
information embodied in consumption picked up by survey expectations of future output? Do
the survey data indicate, on the other hand, that consumers receive a great deal of news that is
not reflected in current consumption? We turn to these questions now, introducing some
expectational measures from the Michigan Survey of Consumers and augmenting the bivariate
consumption-income VARs with these variables.

(b) Confidence Data
The survey measure that we will make the most use of in this paper, which we call E5Y,
summarizes responses to the following question: “Turning to economic conditions in the country
as a whole, do you expect that over the next five years we will have mostly good times, or
periods of widespread unemployment and depression, or what?” The variable is constructed as
the percentage giving a favorable answer minus the percentage giving an unfavorable answer
plus one hundred.5 Our particular affinity for this question arises from the fact that it is aimed at
gauging expectations over a relatively long horizon, and because of its specificity as to the

5 Thus a value of 100 is a “neutral” position, while a value of 140 means that the fraction of responses reflecting
optimism about the future exceeds the fraction reflecting pessimism by forty percentage points.

5

---

relevant time frame.6 However, its correlation with the response to a similar question specifying
a horizon of only twelve months (a variable we call E12M) is 91 percent, and its correlation with
another concerning expected changes in personal financial situation over the next twelve months
is 85 percent. The correlation of E5Y with the overall expectations component of the Michigan
index exceeds 95 percent. Our results in this section are essentially unchanged by the
substitution of either of these alternative expectations variables. The alternative questions are
described in more detail in the Appendix.
Figure 2 plots E5Y and E12M against time. Both series undergo repeated dramatic
swings though (as we would expect) the twelve-month-ahead expectations are more volatile than
the expectations over a five year horizon. Both variables are quite stationary. The cross-
correlogram between E5Y and the conventional Hodrick-Prescott detrended GDP (not shown)
indicates that the expectations are by no means a reflection of current output; the
contemporaneous correlation between detrended GDP and E5Y is essentially zero. E5Y is
negatively correlated with the output gap lagged several periods, and positively correlated with
the gap several quarters ahead.

(c) Augmenting the Bivariate VAR with Confidence Measures

We begin by augmenting Cochrane’s income-consumption VAR with E5Y. As before,
the system is estimated allowing cointegration between consumption and income, with four lags
of each variable. Because confidence measures are clearly stationary, E5Y cannot enter into the
long run equilibrium relationship, and we once again impose that the cointegrating vector
between consumption and income is [1, -1].7 It is necessary to make some choices as to how to
orthogonalize the innovations. It is important to understand that alternative orthogonalizations in
this context are not to be thought of as minimum delay restrictions that delineate alternative
structural models; in almost any sensible model, innovations in the underlying structural shocks

6 Some might argue as well that this question gives the animal spirits hypothesis its “best shot”. One argument is
that individuals are likely to be more sober-minded in assessing family resources than in forming expectations about
the national economy. Another is based on animal spirits models that focus on strategic complementary; in those
models beliefs about the economic activities of other agents are central.

7 Formally, the system features E5Y in levels, consumption and income in first differences, and the lagged (log)
difference between consumption and income as an exogenous variable. Our results are virtually identical when
estimating the full system in levels.


6

---

should affect all three variables instantaneously. Attempts to think about ordering should instead
focus on “assigning” the common component of the information in innovations to one or another
variable so as to provide upper and lower bounds for the amount of information content in each
of the series.
To begin to assess the extent to which the “consumption shocks” in the bivariate VAR
are in fact “information shocks” that are well captured by innovations in the survey expectations,
we compute impulse responses with E5Y ordered first. As in Cochrane (1994a), income is
ordered last, though our results from the augmented consumption-income VARs are largely
invariant to the placement of income in the ordering. Figure 3(a) presents the impulse responses
to E5Y and consumption innovations under this orthogonalization. The dashed lines represent
90 percent bias-corrected bootstrap confidence bands.8 As in Cochrane’s two variable system,
consumption behaves roughly like a random walk in response to its own innovation. In response
to a consumption innovation output jumps up on impact, follows a slight hump-shape, and levels
off at roughly 0.4 percent higher than its pre-shock value. Though not shown, output displays a
large and significant response to its own innovation that dissipates rather quickly. The part of
the output innovation that is orthogonal to consumption predicts no significant movement in
consumption at any horizon.
An innovation to E5Y has very small (though statistically significant) implications for
both consumption and output on impact. The small impact effects are followed by slowly-
building, statistically and economically significant, and apparently permanent responses of both
consumption and output. In particular, a one standard deviation innovation to E5Y predicts
levels of output and consumption that are roughly 0.7 percent higher forty quarters hence;
further, the long run responses of both consumption and GDP to an E5Y innovation are both
statistically significant at better than the 90 percent level. E5Y responds significantly neither to
income nor consumption innovations; its own innovation accounts for more than 95 percent of its
forecast error variance at all horizons under this ordering.
E5Y innovations thus clearly convey important information about the future time paths of
real variables, with “effects” that show no tendency to attenuate even at long horizons.
However, to what extent are innovations in E5Y simply reflective of information contained in

8 In particular, we generate the confidence bands from the empirical distribution of impulse responses based on
2000 bootstrap draws using bias-corrected OLS slope coefficients as proposed by Kilian (1998).

7

---

consumption? To address this possibility, we re-order the variables in the system such that E5Y
is orthogonalized with respect to consumption. As before, output is ordered last in the system.
Figure 3(b) presents impulse responses with this particular ordering.
The qualitative features of the impulse responses are unaffected by the alternative
orthogonalization. In particular, E5Y innovations orthogonal to consumption still predict
slowly-building and permanent responses of both output and consumption. The point estimates
are slightly smaller than in the case with E5Y ordered first, with a one standard deviation
innovation to E5Y prognostic of long run increases in both consumption and output of slightly
more than 0.5 percent (as opposed to 0.7 percent with E5Y ordered first). E5Y also responds
significantly (in the statistical sense) to a consumption innovation, but the point estimate is small
and the response is statistically significant only for a few quarters.
Figure 4 graphically depicts the variance decompositions of consumption, income, and
E5Y under both orthogonalizations. Regardless of ordering, own innovations account for the
bulk of the forecast error variance of output at short horizons and virtually none at longer
horizons. Ordered first, E5Y innovations account for more than 60 percent of the forecast error
variance of income and consumption at long horizons. Even after orthogonalization with respect
to consumption, innovations to E5Y still account for more than 30 percent of the long horizon
forecast error variance of both income and consumption. We can thus fairly easily reject the
hypothesis that E5Y simply reflects information available in consumption. Rather, innovations
in E5Y and in consumption each convey news about future output that is not subsumed in the
other.
We now examine several variations on the three variable VAR using alternative measures
of consumer confidence. First, we substitute the relative score from the question on the
Michigan Survey concerning expected personal financial situation (PFE) in place of E5Y. This
question gauges expectations analogously to E5Y and E12M, although it specifically asks for
expectations concerning personal situations as opposed to aggregate expectations.9 The second

9 Dominitz and Manksi (2004) express doubt that consumers can give meaningful responses to survey questions
concerning aggregate as opposed to individual expectations, and they point to the higher volatility of responses to
questions like E5Y versus questions like PFE as support. Given the structure of the questions, however, we would
in fact expect aggregate questions to have greater volatility even if individuals are equally capable of answering both
kinds of questions accurately. For example, even in severe recessions most people do not personally experience
layoffs. The typical respondent who says that the national economy will exhibit “periods of widespread
unemployment or depression” is predicting that a significant minority of others will experience layoffs while his or
her own income is stable by comparison.

8

---

modification is to use the Index of Consumer Sentiment (ICS) in place of the purely forward-
looking survey questions. While the ICS is the most reported measure of consumer confidence
(both by the press and in the academic literature), it is an average of survey responses to both
forward-looking and retrospective questions, and thus its interpretation is unclear. For a more
detailed description of these alternative confidence measures and their statistical relationships
with E5Y, the interested reader is referred to the Appendix.
Figure 5 presents impulse responses to confidence innovations in our three variable
system with three alternative measures of confidence: E5Y, PFE, and ICS. We order the
confidence measure first in the system, impose cointegration between consumption and output,
and employ a lag structure of four.10 There is very little qualitative or quantitative difference
between the results using E5Y or any of the other broad confidence measures. The seeming
disparity between some of our results and others in the academic literature thus does not appear
to be attributable to different measures of confidence.11 Use of other alternative confidence
measures – such as E12M or the expectations index of the Michigan Survey – and alternative
measures of consumption and output (for example, durable goods consumption or private sector
GDP) also produce very similar impulse responses.
In summary, innovations in expectational variables from the Michigan Survey of
Consumers are powerful predictors of changes in output and future spending that last for the
foreseeable future. This finding obtains regardless of whether or not the confidence innovations
are orthogonalized with respect to current spending. In Section III we will argue, based on
model with both shocks to information and animal spirits shocks, that the apparent permanence
of the impulse responses of consumption and output to confidence shocks is more consistent with
an information view of confidence than it is with an animal spirits interpretation.
Our finding that unexpected increases in confidence imply predictably higher subsequent
consumption is somewhat related to the results of Carroll, Fuhrer, and Wilcox (1994), who focus

10 Alternative orderings with the confidence measure after consumption also produce quite similar results.

11 Among papers in this literature that find a small role for consumer confidence measures in predicting the future
time path of economic variables are Mishkin (1978), Leeper (1992), Mehra and Martin (2003), and Croushore
(2005). Matsusaka and Sbordonne (1995) and Howrey (2001) report a much stronger prognostic role for
confidence, while Ludvigson (2004) takes something of a middle ground. Souleles (2004) analyzes the micro data
underlying aggregate confidence data used in the present paper. However, the most important difference between
our results and the results in these papers is that by looking at impulse responses to confidence innovations many
periods into the future, we are able to recover the longer run implications of confidence innovations that are in fact
more powerful that are the short run business cycle “effects”.

9

---