Heterogeneous Fore asts and Aggregate Dynami s

Antulio N. Bom m

Division of Monetary A airs, Federal Reserve Board, Washington, DC, 20551, USA
January 19, 2000
Abstra t

Motivated by issues raised in both the nan e and e onomi s literatures, I onstru t a
dynami general equilibrium model where agents use di ering degrees of sophisti ation when
fore asting future e onomi onditions. All agents solve standard dynami optimization problems and fa e strategi omplementarity in produ tion, but some solve their inferen e problems
based on simple fore asting rules of thumb. Assuming a hierar hi al information stru ture
similar to the one in Townsend's (1983) model of informationally dispersed markets, I show
that even a minority of rule-of-thumb fore asters an have a signi ant e e t on the aggregate
properties of the e onomy. For instan e, as agents try to fore ast ea h others' behavior they
e e tively strengthen the internal propagation me hanism of the e onomy. The quantitative
results are obtained by alibrating the model and running a battery of sensitivity tests on key
parameters. The analysis highlights the role of strategi omplementarity in the heterogeneous
expe tations literature and quanti es many qualitative laims about the aggregate impli ations
of expe tational heterogeneity.

Classi ation: E32, D82, C61
Keywords: expe tations, business y les, strategi omplementarity, propagation, bounded rationality
JEL 

E-mail: abom mfrb.gov, Fax: (202) 452-2301, Tel.: (202) 736-5619.

I re eived valuable omments from an anonymous referee, Robert King, Mi hael Binder, Satyajit Chatterjee, Russell
Cooper, Frank Diebold, John Haltiwanger, Mi hael Waldman, and seminar parti ipants at the 1997 North Ameri an
Winter Meetings of the E onometri So iety. Steve Sumner provided ex ellent resear h assistan e in earlier stages
of this proje t. The opinions expressed here are not ne essarily shared by the Board of Governors of the Federal
Reserve System or any other members of its sta .

1

Introdu tion

Although the rational expe tations assumption is a standard xture of most work in modern
ma roe onomi s and nan e, a growing number of studies has raised on erns about its theoreti al
and empiri al impli ations. For instan e, Frydman and Phelps (1983), Board (1994), and Arthur
(1994) argue that human rationality is bounded, and thus the rational expe tations assumption
imposes extreme informational and omputational requirements on agents (see also Sargent, 1993).
In addition, other resear hers base their reservations on on erns about the observable hara teristi s of rational expe tations equilibria. For example, De Long, S hleifer, Summers, and Waldmann
(1990) onstru ted a model in whi h behavior based on irrational noise trading helps explain a
number of observed phenomena in nan ial markets, su h as the ex ess volatility of asset pri es
and the equity premium puzzle. Also in the nan e literature, Roll (1997) mentions in omplete
(bounded) rationality as a possible explanation for the observation of large trade volumes in debt
markets. Whereas these reservations have led some e onomists to dis ard the rational expe tations
hypothesis altogether, others have sought to re onsider it in the ontext of environments with less
demanding informational assumptions and more plausible observable impli ations. Lu as (1975)
and Townsend (1983) were among the rst to take up this line of inquiry. Townsend, in parti ular,
analyzed a model where agents form heterogeneous expe tations be ause their fore asts are onditioned on di erent subsets of the relevant data. He showed that, as agents attempt to \fore ast
the fore ast of others," the e onomy onverges to a rational expe tations equilibrium.
Subsequently, the work of Haltiwanger and Waldman (1985, 1989) brought a new perspe tive
to the analysis of fore ast heterogeneity. Rather than working with dynami models with no externalities (the Lu as-Townsend approa h), Haltiwanger and Waldman analyzed environments that
allowed for strategi omplementarity. In addition, instead of fo using on agents with di erent
a ess to the data, they built models where fore ast heterogeneity arises be ause some agents may
use more sophisti ated fore asting methods than others, an idea that is formalized in the work of
Evans and Ramey (1992) and Sethi and Franke (1995). Unlike the results obtained by papers in the
Lu as-Townsend tradition, the lass of simple, stati models analyzed by Haltiwanger and Waldman gave rise to environments where fore ast heterogeneity did ast some doubt on the aggregate
impli ations of the rational expe tations hypothesis. They showed that, with strategi omplemen1

1
Cooper and John (1988) de ne strategi omplementarity and dis uss its impli ations for ma roe onomi s. For
the purpose of my paper, strategi omplementarity involves a situation where an individual's output de ision is
in reasing in the level of aggregate output.

1

tarity, less sophisti ated fore asters may have a sizable e e t on the evolution of aggregate output,
e e tively driving the e onomy away from its pure rational expe tations equilibrium.
In this paper I introdu e and solve a model that brings together important issues stemming
from both the Lu as-Townsend and Haltiwanger-Waldman approa hes to analyzing heterogeneousexpe tations models. Capturing the key insights of Haltiwanger and Waldman, my analysis allows
for strategi omplementarity and heterogeneous fore asting rules. However, to bring the dis ussion more into the urrent stage of ma roe onomi thought|whi h emphasizes dynami rather
than stati frameworks|I extend the Haltiwanger-Waldman analysis to a ri her dynami general
equilibrium model, whi h I solve with a methodology that is very lose in spirit to Townsend's.
In hoosing a spe i dynami modeling framework, I opted for the lass of models in the
real business y le (RBC) tradition. The advantages of this hoi e are two-fold. First, the RBC
framework is well understood by the profession, and the results I obtain under fore ast heterogeneity
and strategi omplementarity an be dire tly and quantitatively ompared to those generated by
standard RBC models. Se ond, by introdu ing fore ast heterogeneity into the RBC framework, I
am able to address a re urring theme in the literature: the weak internal propagation me hanism
that underlies many equilibrium business y le models (Cogley and Nason, 1995).
The paper's main results and methodology an be summarized as follows. With a suÆ iently
strong degree of strategi omplementarity, I show that even if only a small subset of agents fore asts
a ording to a simple but reasonable rule of thumb, aggregate output exhibits more persisten e
than warranted by either (i) the degree of serial orrelation in the produ tivity sho k pro ess, or
(ii) the share of rule-of-thumb fore asters in the total population. More to the point, be ause
agents engage in fore asting ea h others' fore asts, rule-of-thumb fore asters have a quantitatively
important impa t on the serial orrelation properties of the business y le. The results are obtained
by alibrating all standard RBC parameters and then running a battery of sensitivity tests on the
fore ast-heterogeneity and strategi - omplementarity parameters. The sensitivity tests highlight
the role of strategi omplementarity in the fore ast-heterogeneity debate and pre isely quantify
the qualitative laims made by Haltiwanger and Waldman and others.
2

The Model

Apart from the issue of heterogeneity, the model is very lose to that of Baxter and King (1991).

2

2.1

Te hnology and Preferen es

Individual output is a fun tion not only
of inputs and a produ tivity sho k, but also of an index of per apita aggregate output. This index
is a geometri average of the per apita output de isions of two types of agents:
Strategi Complementarity and the Produ tion Fun tion. 

R (1
Yt = YR;t
YS;t

(1) 

R )

where R is the population share of rule-of-thumb fore asters, and YR;t is their per apita output
de ision. (Rule-of-Thumb fore asting is dis ussed in the next se tion.) The S subs ript denotes
variables pertaining to the sophisti ated fore asters.
An agent of type i fa es the produ tion fun tion,
yi;t = exp(At )F (ki;t 1 ; ni;t )Yt

i = R; S

(2)

where yi;t denotes individual output, and ki;t and ni;t are the apital and labor inputs. At is
assumed to apture sto hasti shifts in produ tivity; it follows a rst-order autoregression with
auto orrelation oeÆ ient  and innovation at, whi h is normally distributed with zero mean. F (:)
is a Cobb-Douglas produ tion fun tion, F (ki;t ; ni;t)  ki;tk ni;tn , whi h is homogeneous of degree
1. The  parameter in equation (2) embodies the omplementarity assumption; it determines the
extent to whi h individual output, yi;t, depends on aggregate output, Yt. (0   < 1)
Evolution of the Capital Sto k. Output not onsumed onstitutes gross investment, ii;t . With ki;t
representing the apital sto k at the end of period t, and assuming that this sto k depre iates at
the rate Æ, 0  Æ < 1,
ki;t = (1 Æ)ki;t + ii;t
(3)
1

1

1

1

Preferen es are homogeneous throughout the e onomy. The momentary utility fun tion of a representative agent is de ned in terms of onsumption ( i;t ) and leisure (li;t),

Preferen es.

u( i;t ; li;t ) = log( i;t ) + l log(li;t )

where leisure is expressed as a proportion of the unit time endowment.

3

(4)

2.2

Individual Behavior

The e onomy is populated by utility-maximizing, in nitely-lived, forward-looking agents who dis ount the future at the rate . As in Townsend (1983), we an think of individual behavior as the
out ome of two separate problems: dynami optimization and inferen e. The solution to the rst
problem yields the perfe t-foresight equilibrium laws of motion of all hoi e variables, whi h express
the evolution of these variables as a fun tion of past, urrent, and future states of the e onomy. The
assumption of heterogeneous expe tations amounts to saying that agents use the same me hanism
to solve their dynami optimization problem, but not their inferen e problem.
The Dynami Optimization Problem. Given equations (1) through (4) and the usual time and
goods onstraints, the dynami optimization problem redu es to solving
max

1 n
X
u(
t

t=0

i;t

; 1 ni;t) + t [exp(At )F (:)Yt i;t

o

ki;t + (1 Æ)ki;t 1 ℄

(5)

subje t to ki; , the transversality ondition limt!1 t t ki;t = 0, and for given fAt ; Yt g1
t . (t is
the dis ounted Lagrange multiplier relevant for time t.)
The perfe t-foresight equilibrium paths of onsumption, investment, and labor e ort are given
by the solution to the system of Euler equations that orresponds to (5). Given that generally there
is no losed-form solution to this system, my fo us is on an approximate solution, obtainable by loglinearizing the system around its steady state. After imposing a symmetry ondition that says that
individuals who rely on the same fore asting me hanism make identi al de isions in equilibrium,
the (approximate) equilibrium laws of motion take the form
1

=0

2

xi;t = i;k K^ i;t 1 + i; ~ i;t + i;eei;t ;

i; j = S; R; j 6= i

(6) 

i h (Fi; ei;t h + Fi; ei;t h ).
where xi;t  [N^i;t ; K^ i;t ; C^i;t℄0 , ei;t  [At ; Y^j;t℄0 , and ~i;t  P1
h
A \ aret" over a symbol denotes that the variable is expressed in per entage deviations from
the steady state (e.g., Y^i;t  log(Yi;t =Yi)), and upper- ase letters denote per- apita variables. The
matri es i and Fi, as well as the i parameter, are fun tions of various steady-state properties of
the model, su h as the steady-state apital-output ratio and the labor share of total in ome.
The Inferen e Problem. To fully hara terize individual behavior, I still need to expli itly address
the issues of un ertainty and expe tational formation. For ea h period t, I assume that all agents
=0

2

1

+ +1

2

+

The log-linear approximation method used here is des ribed in detail in King, Plosser, and Rebelo (1990).

4

follow a two-stage de ision pro ess. At the beginning of the period, the rst stage takes pla e:
agents make their labor supply and apital a umulation de isions before being able to observe the
urrent value of the produ tivity shifter, At, or the urrent output de ision of the other agents in
the e onomy. The fa tor-allo ation de ision rules take the form:
3

h

zi;t = i;k K^ i;t 1 + i; E (i) ~ i;t j
t

1

i

+ i;eE i [ei;t j
t ℄
( )

1

(7)

where zi;t  [N^i;t ; K^ i;t℄0 , and the i parameters orrespond to the appropriate elements of the i
matri es from equation (6).
t is the information set available at the beginning of period t; it
ontains the whole history of the e onomy up to period t 1. E i [:j
t ℄ denotes the expe tation
of a type i agent onditioned on
t .
Given last period's apital sto k and this period's fa tor-allo ation de isions, produ tion and
apital a umulation take pla e, and the agents move on to the se ond and last stage of their
de ision making pro ess. Assuming that both sophisti ated and rule-of-thumb fore asters observe
ea h others' output as soon as produ tion o urs, ea h agent an use its knowledge of the produ tion
fun tion to dedu e the urrent value of the produ tivity shifter (At ). Therefore, the onsumption
de ision is based on a larger information set,
;t  f
t ; At ; YS;t; YR;t g.
Equation (7) makes expli it two points advan ed earlier in this paper. First, agents of di erent
types are informationally linked: To generate their own de ision rules they must fore ast the
behavior of the other agents in the e onomy|re all that Y^j;t is an element of ei;t . Se ond, equation
(7) highlights the hannel through whi h agents' expe tations a e t their behavior: The di erent
expe tational rules embedded in E S and E R an lead to potentially di erent responses to the
same fundamental sho ks.
1

( )

1

1

0

( )

3

1

( )

Expe tational Heterogeneity

Sophisti ated agents make use of full knowledge of the stru ture of the model, in luding the expe tational behavior of rule-of-thumb agents. For any given variable t , their expe tations an
be formally de ned as E S [ t j
t ℄ = E [ t j
t ℄, whi h is the mathemati al expe tation of t
onditioned on the information set
t and the true stru ture of the entire model.
( )

1

1

1

3

Kydland and Pres ott (1982) assume a similar information stru ture.

5

3.1

Rule-of-Thumb Fore asting

Three riteria guided the spe i ation of an illustrative fore asting model for rule-of-thumb agents.
First, to apture the on erns of the ostly-implementation literature|e.g. Evans and Ramey
(1992) and Board (1994)|the rule must be simple to implement. Se ond, the fore asting rule
should generate \reasonable" fore asts, i.e., it should be onsistent with well-known hara teristi s
of the e onomy. Finally, the expe tational model must not be at odds with the agents' ability to
solve their dynami optimization problem. In other words, when solving their dynami optimization
and inferen e problems, rule-of-thumb agents must rely on a single, onsistent pool of information
about the behavior of the e onomy.
A ording to (7), a type i agent must fore ast urrent and future movements in total fa tor
produ tivity (At ) and the per apita output of the other agents in the e onomy (Y^j;t). For illustrative
purposes, suppose that rule-of-thumb agents rely on an autoregressive fore asting model:
E (R) [At+h j
t 1 ℄ = h+1 At 1
E (R) [Y^S;t+h j
t 1 ℄ = h+1 Y^S;t 1

(8)
(9)

To see how the above fore asting stru ture fares with the three riteria listed above, note that,
rst, espe ially for lose or equal to , the fore asting model in equation (8) is not only simple,
reasonable, and onsistent with their ability to solve their dynami optimization problem, but also,
for = , perfe tly rational. Se ond, while it is obvious that (9) is only an approximation to the true
pro ess governing the evolution of Y^S;t, it aptures a well-known feature of traditional RBC models:
the fa t that output persisten e is tightly linked to the degree of serial orrelation in the produ tivity
sho k series. Thus, rather than taking the time and resour es to ompute a fully model- onsistent
fore ast for Y^S;t, a pra ti e that would onsiderably ompli ate the solution to the model, rule-ofthumb agents use the simple model given by (9). Finally, note that the ability of rule-of-thumb
agents to solve their dynami optimization problem with the same of level of sophisti ation as the
other agents is not in onsistent with the relatively unsophisti ated methods they use in solving
their inferen e problem. The solution to the dynami optimization problem requires stru tural
information only about one's own onstraints and opportunities; by assumption, both types of
agents use this information. However, to solve their inferen e problem in a way onsistent with
the rational expe tations hypothesis, individuals also need omplete stru tural information on the
onstraints and opportunities fa ing the other agents in the e onomy; by assumption, sophisti ated
6

agents pro ess this information, rule-of-thumb fore asters do not.
In several aspe ts, the expe tational assumptions made so far are similar to Townsend's (1983)
des ription of a hierar hi al informational stru ture. Sophisti ated fore asters are pla ed higher in
the hierar hy: In addition to the information that enables them to solve their dynami optimization
problem, they also know the pre ise nature of the dynami optimization and inferen e problems
being solved by the rule-of-thumb agents. Thus, their fore asts in orporate stru tural information
about the whole e onomy. In ontrast, the stru tural information embodied in the fore asting
behavior of the rule-of-thumb fore asters is self- ontained: They use information about their own
onstraints and opportunity sets, but not those of sophisti ated agents.
3.2

De ision Rules under Heterogeneous Expe tations

Given the hierar hi al information stru ture, the model an be solved sequentially in two steps.
Starting at the bottom of the hierar hy, I rst derive the de ision rules of rule-of-thumb agents and
then use the results to ompute the more ompli ated de ision rules of the sophisti ated fore asters.
Rule-of-Thumb Agents. The fa tor-allo ation de ision rule of rule-of-thumb agents,
zR;t = R;k K^ R;t

1

+ R;AAt + R;Y Y^S;t
1

1

(10)

is obtained by substituting their expe tations formulae|equations (8) and (9)|into the perfe tforesight equilibrium law of motion of zR;t |equation (7). Given the fa tor-allo ation de isions,
produ tion and onsumption take pla e.
Sophisti ated Agents. Like the rule-of-thumb fore asters, sophisti ated agents make their fa torallo ation and onsumption de isions in two states and subje t to the same information sets,
t
and
;t. Thus, their fa tor-allo ation de isions are made before they an observe either the output
de ision of the other agents or the urrent state of produ tivity. However, to form expe tations
about YR;t , sophisti ated agents look not only at their own dynami optimization problem, but also
at the de isions and fore asting models used by rule-of-thumb agents. This information is subsumed
in equations (6), through (10), whi h an be grouped together with the appropriate Euler equations
to form a two-sided matrix di eren e equation that an be solved for the de ision variables.
Aggregation. Having outlined above how di erent types of agents go about solving their respe tive
utility maximization problems, my ultimate interest lies on the analysis of the dynami s of the
e onomy as a whole. As it turns out, the evolution of aggregate variables an be easily derived
1

0

7

from the individual de ision rules omputed above|see equation (1).
4

Quantitative Business-Cy le Analysis

To assess whether the introdu tion of fore ast heterogeneity a e ts the dynami properties of a realbusiness- y le e onomy, I run what Kydland and Pres ott (1996) all a omputational experiment
(see also King, 1995). The nal steps of su h an experiment involve alibrating the model to allow
for meaningful quantitative analysis and then running the experiment itself.
4.1

Model Calibration

With the ex eption of the fore ast-heterogeneity and omplementarity parameters, [R ; ; ℄, all
model parameters are alibrated as in King, Plosser, and Rebelo (1988).
Strategi omplementarity. The alibration of the strategi omplementarity parameter () is
guided by the empiri al work of Baxter and King (1991), Caballero and Lyons (1989, 1992), Cooper
and Haltiwanger (1996), and Basu and Fernald (1995, 1997). However, even a asual look at these
papers reveals a very wide range of estimates for . For instan e, Baxter and King and Caballero
and Lyons report estimates that range from from 0.1 to 0.49, and, while the results obtained by
Basu and Fernald point towards the lower range of these estimates, Cooper and Haltiwanger report
even larger estimates. For the purposes of this paper, rather than making a ase for any parti ular
estimate of , I run my experiments using a range of values for  that is onsistent with the diverse
ndings of the empiri al literature and then tra e the onsequen es of these di erent estimates for
the aggregate e e ts of expe tational heterogeneity.
Serial orrelation and volatility of te hnology sho ks. Two parameters that do not a e t the steadystate properties of the model, but play a ru ial role in aggregate u tuations, are the innovation
varian e and the autoregressive oeÆ ient of A^t, (a and , respe tively). To alibrate a , I
simply set it to a value that makes the model's output varian e equal to its empiri al ounterpart.
Nonetheless, as long as my fo us rests on the propagation me hanism, my results are invariant to
the parti ular parameterization of a .
The parameterization of  is designed to highlight a re urring weakness of the standard RBC
model with homogeneous expe tations: the la k of a quantitatively important internal propagation
me hanism (Cogley and Nason, 1995). As is well known, in order to be able to mimi the degree of
serial orrelation in the data, most RBC models require near-unit root pro esses for At , e e tively
2

2

8

2

implying that persisten e is exogenously imposed on the system, rather than explained by it. To
isolate the role of expe tational heterogeneity in the persistent generation pro ess, I start by setting 
at 0.5, about half the usual parameterization adopted in standard RBC models.
Rule-of-Thumb Fore asting. If there were some pre ision on erns surrounding the available estimates of the strategi omplementarity parameter, we are hard pressed to nd any estimates,
however impre ise, of the expe tational parameters of the model (R and ). Thus, I shall treat 
R and as semi-free parameters and experiment with a wide range of values for ea h of them.
A ordingly, I view my results as a mapping from the magnitudes of the , R and parameters to
the properties of the arti ial time series generated by the model. The rest of this se tion provides
a sensitivity analysis that re e ts this mapping.
4

4.2

Expe tational Heterogeneity and the Propagation Me hanism

I start by examining an e onomy without strategi omplementarity ( = 0). Table 1-A summarizes the auto orrelation fun tion of aggregate output under alternative values of the expe tational
parameters. In parti ular, the table reports the results of omputational experiments that explore di erent degrees of rule-of-thumb agents' misper eption about the persisten e of te hnologi al
sho ks|for  = 0:5, is set to 0.3, 0.5, and 0.7|and di erent shares of rule-of-thumb agents in
the total population|R varies from 0 to 0.9.
Three main results are evident Table 1-A. First, the impa t of rule-of-thumb fore asters on the
persisten e of output is potentially negligible, even if we allow these agents to make up the vast
majority of the population (R = 0:9). Se ond, though quantitatively small, the parti ular expe tational model used by rule-of-thumb fore asters has a noteworthy property: Their misper eptions
about the persisten e of the te hnologi al sho k are re e ted in the a tual serial orrelation of
output: Whenever they expe t the sho ks to be more [less℄ persistent than warranted by the data
generating pro ess, aggregate output ends up slightly more [less℄ persistent than otherwise. Third,
when there are no misper eptions about the size of , the heterogeneous expe tations model essentially redu es to a standard RBC model. Taken together, the results suggest that the in lusion of
unsophisti ated fore asters in the RBC model without strategi omplementarity produ ed quantitatively insigni antly e e ts: With  set to zero, small deviations from the rational expe tations
assumption produ e only small deviations from the standard RBC results.
4

I am grateful to an anonymous referee for suggesting this ourse of inquiry.

9

Table 1-B summarizes the results
of omputational experiments identi al to the ones des ribed above, ex ept that now I set the
strategi omplementarity parameter to 0.49. The results stand in stark ontrast to the ones
reported in Table 1-A: Even when the rule-of-thumb fore asters represent only a minority of the
population, the serial orrelation properties of output may be a e ted in a quantitatively important
way. For instan e, if these agents over-estimate the persisten e of the te hnologi al sho k ( = 0:70, 
= 0:50), aggregate output be omes noti eably more persistent than in the ase of purely rational
expe tations even with R = 0:30. Thus, unlike the ase of no strategi omplementarity, even
small deviations from the rational expe tations hypothesis an lead to signi ant deviations from
standard RBC results when external returns are high .
Tables 1-A and 1-B highlight an important feature of the model. Note that, for given and 
, the e e t of rule-of-thumb fore asting on the persisten e of aggregate output does not generally
monotoni ally in rease with R . For instan e, for = 0:70, Table 1-B shows that output a tually
be omes slightly less persistent as the share of rule-of-thumb fore asters in the population in reases
from 0.60 to 0.90. This nding has important impli ations for the study of the aggregate e e ts
of expe tational heterogeneity. What it says is that persisten e is not simply being exogenously
generated as a result of the introdu tion of rule-of-thumb agents. In addition to this exogenous
fa tor, the expe tations-indu ed propagation me hanism featured in this paper has an important
endogenous omponent: Strategi omplementarity strengthens the informational links between
the two types of agents, whi h in turn leads the sophisti ated fore asters to in orporate into their
de ision making pro ess the autoregressive nature of the fore asting models used by the ruleof-thumb agents (see subse tion 4.4 for additional dis ussion of the intera tions between the two
types of agents). Therefore, as the population share of sophisti ated agents de reases, the aggregate
impli ations of their optimal responses to rule-of-thumb fore asting is dampened, and this explains
why the relative e e t of rule-of-thumb fore asting eventually de reases at very high values of R .
Thus far I have reported on the e e ts of rule-of-thumb fore asting under what might be alled
two polar assumptions about the magnitude of the strategi omplementarity parameter: the zero
lower bound featured in most RBC models and 0.49, whi h is at the high end of the range of
estimates dis ussed in the previous se tion. A question of interest is what happens at intermediate
values of . Indeed, I ran the omputational experiments reported in Table 1-B for all point estimates of  obtained by Baxter and King (1991) and Caballero and Lyons (1992)| = 0.10, 0.32,
0.45, and 0.49. (I did not in lude the higher estimates reported by Cooper and Haltiwanger, 1996,
as they would violate the model's stability onditions.) The impa t of the type of expe tational
Expe tational Heterogeneity under Strategi Complementarity.

10

heterogeneity examined in this paper is still quite sizable for  = 0:45, but the results are not as
dramati for the two lower values of the external returns parameter,  = 0:10 and  = 0:32. These
results highlight the fa t that more pre ise estimates of the a tual degree of strategi omplementarity are ru ial for a more de nitive assessment of the aggregate (quantitative) impli ations of
fore ast heterogeneity.
4.3

Model Evaluation

Although I annot fully alibrate all parameters of the model and ompare its time series properties
with sele ted moments of the data, it is still useful to verify whether a plausible parameterization
of the heterogeneous-expe tation RBC model with strategi omplementarity an make it roughly
onsistent with the data.
Table 2-A, extra ted from King et al. (1988), summarizes the sele ted moments of the U.S. data
that the model will try to mat h. The orresponding model moments are shown in Table 2-B. The
results reported in this table are obtained by assuming a relatively high degree of omplementarity
( = 0.50), while potentially allowing for only a limited role for rule-of-thumb fore asting (R =
0.30). To highlight the internal propagation me hanism oming from heterogeneous expe tations
under strategi omplementarity, I arbitrarily set the persisten e parameter () at 0.70, lower than
what a standard RBC model would require to apture the serial orrelation observed in the data.
Assuming no misper eptions from the part of rule-of-thumb agents ( = ), the model repli ates
well the serial orrelation of the data, espe ially for onsumption and output. The observed relative
volatilities of output, onsumption, and investment are also largely aptured by the model, though
onsumption and investment are a bit too volatile and hours do not vary as mu h as in the data.
To ompare the internal propagation of the model with the standard RBC framework, Table
2-C shows the same sele ted moments shown in Table 2-B, but now the omplementarity and
expe tational parameters are both set to zero. As expe ted, without strong serial orrelation in
the sho ks, the standard RBC model fails to apture the observed serial orrelation of output.
4.4

Disaggregated Dynami s

In a statisti al (mean-squared-error) sense, the fore asts formed by rule-of-thumb agents are less
eÆ ient than the model- onsistent expe tations of the sophisti ated fore asters. The goal of this
subse tion is to informally he k whether the rule-of-thumb agents' relian e on statisti ally suboptimal fore asts makes them behave substantially di erent from the sophisti ated fore asters.
11

Consider, for example, that di erent agents may fa e di erent osts of forming sophisti ated fore asts. If the behavioral di eren es between sophisti ated and rule-of-thumb fore asters are large,
either these (unspe i ed) osts are very sizable or there is something inherently irrational about
rule-of-thumb fore asting. On the other hand, if behavioral di eren es between the two types of
agents are small, then even small osts asso iated with forming sophisti ated fore asts ould potentially motivate fore asting based on rules of thumb. Moreover, under su h onditions, rule-of-thumb
behavior may a tually be optimal, in that it may satisfy Akerlof and Yellen's (1985) on ept of near
rationality. The analysis performed in this se tion is suggestive, however, and only looks at the
resulting behavior of ea h type of agent; neither the osts of be oming a sophisti ated fore aster
nor the potential utility loss from fore asting with rules-of-thumb are modeled expli itly.
The left panel of Figure 1 plots the simulated output paths for representative sophisti ated and
rule-of-thumb agents using the same parameter settings as in Table 2-B. As shown in this panel,
despite signi ant di eren es in the way they form expe tations, sophisti ated and rule-of-thumb
fore asts behave in an almost identi al manner. In parti ular, as sophisti ated agents anti ipate
and rea t to the imperfe tions embedded in the fore asting s hemes of rule-of-thumb agents, they
e e tively end up mimi king their a tions. In the presen e of omplementarities, it pays to produ e
more [less℄ whenever aggregate output is higher [lower℄, even if the rise [de line℄ in output is largely
due to the suboptimal fore asts of rule-of-thumb agents (and not warranted by true fundamentals).
The right panel of Figure 1 shows that even in their investment de isions, whi h orrespond to
mu h more volatile series, the a tions of sophisti ated and rule-of-thumb agents are remarkably
similar.
Figure 1 suggests that the potential utility losses from being a rule-of-thumb fore aster are
likely small, implying that this ourse of a tion might well onstitute a near-rational strategy.
Thus, the aggregate e e ts of rule-of-thumb fore asters annot be explained by simply looking at
their a tions in isolation; a more interesting and important fa tor lies in the endogenous response
that their a tions eli it from the sophisti ated fore asters, a phenomenon that was also shown to
play an important role in the model's persisten e generation me hanism.
5

Summary and Con luding Remarks

The model des ribed in this paper suggests strong quantitative e e ts at the aggregate level from
allowing even a minority of agents to form expe tations a ording to simple rules of thumb. Indeed, ombined with strategi omplementarity, fore ast heterogeneity an strengthen the internal
12

propagation me hanism of the model, as hypothesized by Haltiwanger and Waldman (1989) and
Oh and Waldman (1994). The intuition is straightforward: As sophisti ated agents try to fore ast
the fore asts (and a tions) of others, they e e tively end up reinfor ing the per eptions of less sophisti ated fore asters, even if these per eptions are not entirely onsistent with the stru ture of the
e onomy. This o urs be ause the sophisti ated agents ultimately are about the aggregate state
of the e onomy, and thus they will produ e more whenever gains in aggregate output are foreseen,
regardless of whether these gains are a result of the unsophisti ated fore asts of the rule-of-thumb
agents.
The above ndings are quantitatively relevant only if the degree of strategi omplementarity
is suÆ iently high, and there is still mu h un ertainty surrounding the empiri al measures of the
omplementarity parameter at the aggregate level. Indeed, while the work of Basu and Fernald
(1995, 1997) favors a low value for the strategi omplementarity parameter, many of the (se toral)
estimates obtained by Cooper and Haltiwanger (1996) would pla e  well above the range of values
analyzed in this paper. Nonetheless, apart from on erns related to the magnitude of , this paper
an be interpreted as a relatively onservative approa h to the analysis of the e onomi impli ations
of fore ast heterogeneity: After all, the rule-of-thumb fore asters depi ted herein are still highly
sophisti ated individuals. In parti ular, they are smart enough to orre tly solve their dynami
optimization problem and suÆ iently well informed to know the exa t nature of the sto hasti
pro ess for te hnology sho ks and to observe the ontemporaneous a tions of all the agents in the
e onomy. Indeed, given how mu h the rule-of-thumb agents are allowed to know, one might even
be surprised as to the extent to whi h su h a small limitation in their behavior mattered as mu h
as it did for any degree of strategi omplementarity.
5

5

Krusell and Smith (1996) analyze an arti ial e onomy where agents are allowed to adopt simple savings rules
of thumb instead of basing their savings behavior on the solution of a standard dynami optimization problem. The
time series properties of the rule-of-thumb e onomy are substantially di erent from those of a traditional arti ial
e onomy.

13

Referen es

Akerlof, G. and J. Yellen, 1985, Can small deviations from rationality make signi ant di eren es
to e onomi equilibria?, Ameri an E onomi Review 75, 708{20.
Arthur, W. B., 1994, Indu tive reasoning and bounded rationality, Ameri an E onomi Review
Papers and Pro eedings 84, 406{411.
Basu, S. and J. Fernald, 1995, Are apparent produ tive spillovers a gment of spe i ation error?,
Journal of Monetary E onomi s 36, 165{188.
Basu, S. and J. Fernald, 1997, Returns to s ale in U.S. produ tion: Estimates and impli ations,
Journal of Politi al E onomy 105, 249{283.
Baxter, M. and R. King, 1991, Produ tive externalities and business y les, Working paper, University of Ro hester.
Board, R., 1994, Polynomially bounded rationality, Journal of E onomi Theory 63, 246{270.
Caballero, R. J. and R. K. Lyons, 1989, The role of external e onomies in U.S. manufa turing,
NBER Working paper No. 3033.
Caballero, R. J. and R. K. Lyons, 1992, External e e ts in U.S. Pro y li al Produ tivity, Journal
of Monetary E onomi s 29, 209{225.
Cogley, T. and J. Nason, 1995, Output dynami s in real-business- y le models, Ameri an E onomi
Review 85, 492{511.
Cooper, R. and A. John, 1988, Coordinating oordination failures in Keynesian models, Quarterly
Journal of E onomi s 103, 441{63.
Cooper, R. and J. Haltiwanger, 1996, Eviden e on ma roe onomi omplementarities, 1996, Review
of E onomi s and Statisti s 78, 78{93.
DeLong, J. B., A. Shleifer, L. Summers and R. Waldmann, 1990, Noise trader risk in nan ial
markets, Journal of Politi al E onomy 98, 703{738.
Evans, G. and G. Ramey, 1992, Expe tation al ulation and ma roe onomi dynami s, Ameri an
E onomi Review 82, 207{224.
Frydman, R. and E. Phelps, 1983, Individual fore asting and aggregate out omes:`Rational expe tations' examined (Cambridge University Press).
Haltiwanger, J. and M. Waldman, 1985, Rational expe tations and the limits of rationality: An
analysis of heterogeneity, Ameri an E onomi Review 85, 326{340.
14

Haltiwanger, J. and M. Waldman, 1989, Limited rationality and strategi omplements: The impli ations for ma roe onomi s, Quarterly Journal of E onomi s 104, 463{483.
King, R. G., 1995, Quantitative theory and e onometri s, Federal Reserve Bank of Ri hmond
E onomi Quarterly 81, 53{105.
King, R. G., C. Plosser and S. Rebelo, 1990, Produ tion, growth and business y les: Te hni al
appendix, Working paper, Ro hester Center for E onomi Resear h, University of Ro hester.
King, R. G., C. Plosser and S. Rebelo, 1988, Produ tion, growth and business y les: 1. The Basi
Neo lassi al Model, Journal of Monetary E onomi s 21, 195{232.
Krusell, P. and A. Smith, 1996, Rules of thumb in ma roe onomi equilibrium: A quantitative
analysis, Journal of E onomi Dynami s and Control 20, 527{558.
Kydland, F. and E. Pres ott, 1982, Time to build and aggregate u tuations, E onometri a 50,
1345{1370.
Kydland, F. and E. Pres ott, 1996, The omputational experiment: An e onometri tool, Journal
of E onomi Perspe tives 10, 69{85.
Lu as, R. E., 1975, An equilibrium model of the business y le, Journal of Politi al E onomy 83,
1113{1144.
Oh, S. and M. Waldman, 1994, Strategi omplementarity slows ma roe onomi adjustment to
temporary sho ks, E onomi Inquiry 32, 318{329.
Roll, R., 1997, Nominal interest rates and loan volume with heterogeneous beliefs, 1997, Finan ial
Markets, Institutions and Instruments 6, 1{58.
Sargent, T., 1993, Bounded rationality in ma roe onomi s (Clarendon, Oxford, UK).
Sethi, R. and R. Franke, 1995, Behavioural heterogeneity under evolutionary pressure: Ma roe onomi impli ations of ostly optimization, E onomi Journal 105, 583{600.
Townsend, R., 1983, Fore asting the fore ast of others, Journal of Politi al E onomy 91, 546{588.

15

Table 1 | Expe tational Heterogeneity and the Propagation Me hanisma

E e t of Alternative Expe tational Assumptions on the Auto orrelation of Aggregate Output
A. Model without Strategi Complementarity ( = 0,  = 0:5)
= 0:3
= 0:5
auto(1) auto(2) auto(3) auto(1) auto(2) auto(3) 
R = 0:0 0.70
0.38
0.22
0.70
0.38
0.22 
R = 0:3 0.68
0.37
0.21
0.70
0.38
0.22 
R = 0:6 0.67
0.36
0.20
0.70
0.38
0.22 
R = 0:9 0.65
0.34
0.20
0.70
0.38
0.22
B. Model with Strategi Complementarity ( = 0:49,  = 0:5)
= 0:3
= 0:5
auto(1) auto(2) auto(3) auto(1) auto(2) 
R = 0:0 0.84
0.65
0.55
0.84
0.65 
R = 0:3 0.86
0.63
0.49
0.89
0.71 
R = 0:6 0.85
0.62
0.47
0.91
0.75 
R = 0:9 0.87
0.65
0.53
0.92
0.78

auto(3)
0.55
0.59
0.65
0.70

auto(1)
0.70
0.71
0.72
0.73

auto(1)
0.84
0.92
0.95
0.94

= 0:7

auto(2)
0.38
0.39
0.39
0.40

= 0:7

auto(2)
0.65
0.77
0.87
0.86

a Ea h expe tational assumption is de ned by a ( , R ) pair. is the per eived value of |the true AR(1)
oeÆ ient of the te hnology sho k|and R is the proportion of rule-of-thumb fore asters in the population. The
table entries show the rst 3 auto orrelations of aggregate output for di erent ombinations of and R . All

standard RBC parameters are alibrated as in King et al. (1988).

16

auto(3)
0.22
0.22
0.23
0.24

auto(3)
0.55
0.67
0.81
0.81

Table 2 | Comparing Sele ted Momentsa

Series

Std Dev Rat. SD auto(1) auto(2) auto(3)

Output
Consumption
Investment
Hours

b
A. U.S. Postwar Quarterly Data

5.62
3.86
7.61
2.97

1.00
0.69
1.35
0.52

.96
.98
.93
.94

.91
.95
.78
.85

.85
.93
.62
.74

B. Model with S.C. and Heterogeneous Expe tations

Output
Consumption
Investment
Hours
Prdvty Sho k
Output
Consumption
Investment
Hours
Prdvty Sho k

5.62
4.73
10.96
1.89
0.53

1.00
0.84
1.95
0.34
0.09

.97
.96
.85
.81
.70

d
C. Standard RBC Model

5.62
4.46
16.24
3.19
3.85

1.00
0.79
2.89
0.57
0.69

.85
.16
.67
.66
.70

.91
.95
.68
.59
.49

.86
.95
.54
.42
.34

.63
.21
.44
.43
.49

.47
.25
.29
.26
.34

a The rst olumn of numbers shows the standard deviation of ea h series; the se ond olumn shows

ratios of standard deviations of ea h series with output. Columns 3 through 4 show rst, se ond, and
third auto orrelation oeÆ ients.
b Sour e: King, Plosser, and Rebelo (1988).
R = 0:30,  = 0:50, =  = 0:70. All other parameters alibrated as in King et al. (1988).
d R = 0:00,  = 0:00,  = 0:70. All other parameters alibrated as as in King et al. (1988).

17

Figure 1
Comparing the behavior of Sophisti ated and Rule-of-Thumb agents

Output

Investment
30

6
Sophisticated Agents
Rule-of-Thumb Agents

4

Sophisticated Agents
Rule-of-Thumb Agents

20
2
10
0

-2

0

-4
-10
-6
-20

-8

0

20

40

60

80

100

0

quarters

20

40

60

80

100

quarters

* The harts show the output and investment de isions of representative sophisti ated and rule-of-thumb agents.
The parameterization is the same des ribed for Table 2-B. ( = 

= 0:70,  = 0:50, R = 0:30)

18