Unformatted text preview: Opinion-producing agents: career concerns and
Stanford Graduate School of Business
June 2001 1 Address: Stanford Graduate School of Business, Stanford, CA 94305. Email: zitze- [email protected] The author would like to thank Glenn Ellison, Bengt Holmstrom, Dean Karlan, Leigh Linden, Sendhil Mullainathan, Nancy Rose, Jean Tirole, Ezra
Zuckerman and seminar participants at Berkeley, Harvard, London Business School, Michigan, MIT, Northwestern, Stanford, Washington University-St. Louis, and Yale for helpful
comments and suggestions. Particular thanks are due to my primary advisor, Robert Gibbons. All errors are my own. The …nancial support of an NSF Graduate Research Fellowship is gratefully acknowledged. I also thank I/B/E/S for providing me with access to their
analyst data. Abstract
This paper models the incentives created by career concerns for opinion-producing
agents. We …nd that career concerns can create an incentive for exaggeration or
anti-herding, since high-ability agents will have opinions that are more di¤erent from
the consensus on average and potential clients will learn more quickly about how
di¤erent an agent’s opinions are from the consensus on average that about whether
or not they are exaggerating. The model predicts that agents should exaggerate more
when they are under-rated by their clients, when the realizations of the variables
they are forecasting are expected to be especially noisy, and when they expect to
make fewer future forecasts. We …nd that these predictions are consistent with the
empirical data on equity analyst’s earnings forecasts. 1 Introduction Much of the information in the so-called information economy is not verifyable information as economists normally de…ne it, but is rather opinion. Opinion goods
such as forecasts, consulting advice, and product reviews are sold in markets, and
the production of these and other types of opinions is also the primary job of many
professionals within organizations. Unlike most traditional goods, the quality of information or opinion goods cannot be readily observed prior to purchase. In addition,
opinions that are not veri…able can be manipulated by their producer. Reputational
or career concerns are thus likely to be especially important to opinion producers.
This paper examines the relationship between the reputational or career concerns
of opinion producers and their incentives to engage in a particular type of opinion
manipulation, namely exaggerating their di¤erences with the existing consensus. It
essentially asks the question: do people exaggerate in order to appear smart? Most
opinions can be thought of as forecasts of a random variable that will be realized in
the future. We develop a model in which potential clients attempt to learn the ability
of forecasters from their track record. In the model, an incentive to exaggerate arises
because high-ability agents have access to more private information and thus have
unbiased beliefs that are more di¤erent from the prior consensus on average. Clients
learn more quickly about how di¤erent an agent’s forecasts are from the consensus
on average than about whether or not they are exaggerating, and thus agents can
temporarily raise estimates of their ability by exaggerating.
The model also yields cross-sectional predictions about when we should expect
agents to exaggerate more. This is potentially useful to consumers of opinions, since
they will want to back out expected exaggeration to form unbiased beliefs. One source
of variation arises from the fact that the di¤erence in learning speed discussed above
is more pronounced when forecast variable realizations are noisier, so agents should
1 exaggerate more under these circumstances. Likewise, agents expecting a shorter
future career length should exaggerate more. Agents should also exaggerate more
when they have had bad luck in the past (and are consequently under-rated by the
market) in order to increase the weighting on future observations of their ability. We
…nd that both the general …nding of exaggeration and these cross-sectional predictions
are consistent with the empirical evidence when we examine the exaggeration of equity
analysts forecasting earnings using the methodology developed in Zitzewitz (2001).
The model of career concerns and exaggeration in this paper draws on the literature on career concerns, reputational concerns for producers of goods of unobservable
quality, and herding theory.1 It is particularly related to three recent papers. Scharfstein and Stein (1990) provide a model of sequential investment decisions in which a
herding equilibrium exists. A key assumption they make is that agents do not know
their ability and thus low-ability agents (who receive noisier signals) do not know
that their signals are noisier and do not know to reduce their reliance on their private information and therefore make reports that deviate more from the consensus.
Repeating the consensus is therefore a signal of ability, and a herding equilibrium
can result. Avery and Chevalier (1999) extend this equilibrium to allow agents to
learn their own ability and …nd that anti-herding will be an equilibrium rather than
herding once agents are expected to have a su¢ciently precise knowledge of their own
1 For general career concerns theory see Fama (1980), Holmstrom (1982), and Gibbons and Mur- phy (1992). The industrial organizations literature on reputational concerns includes Nelson (1974),
Schmalensee (1978), Klein and Le-er (1981), and Shapiro (1983). Herding models include information cascade models (Banerjee, 1992; Bikhchandani, Hirshleifer, and Welch, 1992; Welch, 1999),
incentive-concavity models (Holmstrom and Ricart i Costa, 1986; Zweibel, 1995; Chevalier and Ellison, 1997 and 1999; Laster, Bennett and Geoum, 1999), and career-concerns models (Scharfstein
and Stein, 1990; Brandenburger and Polak, 1996; Trueman, 1994; Ehrbeck and Waldmann, 1996;
Prendergast and Stole, 1996; Avery and Chevalier, 1999; Ottaviani and Sorensen, 2000; E¢nger and
Polborn, 2000). 2 ability. In Prendergast and Stole (1996), agents know their own ability perfectly and
they anti-herd or exaggerate in order to signal ability. Prendergast and Stole (1996)
also allow the set of possible investments to be continuous rather than restricting it
to having only two possible values.
This paper is closest in setup to Prendergast and Stole (1996); its main departure
is to examine an environment in which there is learning from investment outcomes
(or, in the language of this paper, from the realizations of the variables being forecast). This seemingly trival extension yields the cross-sectional predictions about the
degree of exaggeration mentioned above that we can test empirically. The extension comes at a cost, however, since modelling the learning from realizations is too
complicated to handle in a fully Bayesian way. Rather than assuming that potential
clients conduct a Bayesian estimate of an opinion seller’s ability, we instead assume
that they use an econometric methodology that approximates a maximum-likelihood
estimate of ability. The results from this model can therefore be viewed either as
an approximation of the results from a Bayesian model or as the results from a behavioral model in which we make the arguably realistic assumption that clients use
econometric approximations when faced with a Bayesian problem that is too di¢cult
to them to solve.
The remainder of the paper is organized into two sections. The …rst develops the
model and the cross-sectional predictions. The second section presents evidence that
equity analyst’s earnings forecasts are consistent with these predictions. A conclusion
follows. 2 Career-concerns model In this section, we model the incentives created by the career concerns of an agent
who forecasts a series of random variables. We …rst present a two-period model in
3 which agents issue forecasts in the …rst period and receive revenue in the second
period proportional to clients’ valuation of their forecasts based on their …rst period
performance. We then extend the model to three periods in order to examine how
past forecasting performance a¤ects an agent’s incentives to exaggerate. From the
two-period model we conclude that agents have an incentive to exaggerate and this
incentive to exaggerate will be greater when earnings realizations are expected to be
noisy or when agents expect to make a limited number of forecasts in the future. From
the three-period model, we conclude that agents will also exaggerate more when they
have had bad luck in the past and thus are underrated by their clients. 2.1 Two-period model The model has two periods. In the …rst period, the agent issues forecasts of J random
variables after observing a private signal and a common prior. In the second period,
the agent sells early access to their forecasts to clients, receiving revenue proportional
to clients’ valuation of their forecasts based on their …rst-period performance. Agents
also face an exogenous incentive for forecast accuracy, and thus receive revenue proportional to b ¡ ¸ ¢ M SE , where b is the estimate of the value of new information in
the agents forecasts and ¸ is the size of the incentive for accuracy. The role of the incentive for accuracy is to make the language analysts use relevant; we can think
of this incentive as the result of clients’ costs of translating exaggerated forecasts
into unbiased expectations or of industry institutions that measure analysts based on
mean squared error.
The timing of the model is as follows:
1. Nature chooses an ability, a, and an accuracy incentive, ¸, for the agent from a
prior distribution g (a; ¸). Both parameters are observed by the agent but are
unknown to their potential customers.
4 2. The agent forecasts a series of J random variables Aj . For each Aj , the agent
observes a public consensus prior that Aj is distributed N (Cj ; §¡1), where §
is the precision of the prior. Since Cj is public, we can think of the agent as
forecasting yj = Aj ¡ Cj , i.e. the di¤erence between the actual realization and
its consensus prior expectation.
3. In addition, the agent observes an independent private signal s j » N (Aj ; p¡1 ),
where p = a§2
(1¡a§) > 0 is the precision of the signal. p is an increasing function of a; higher ability agents receive higher precision private signals.
4. Based on the consensus priors and her private signals, the agent issues forecasts
using a forecasting rule x = x(s; C; p; ¸), where xj = Fj ¡ Cj is a forecast of yj .
All forecasts are made before any of the earnings variables are realized.
5. Clients with early access to forecasts can make investments in securities with
returns that are proportional to yj .
6. Current and potential clients observe the forecasts x and realizations y and
estimate the value of the information content of the agent’s forecasts using a
valuation rule v = v(x; y ).
b 7. Agents sell early access to their second-period forecasts and receive revenue 2.1.1 proportional to b ¡ ¸ ¢ M SE. Agents consume and experience linear utility.
Forecast information content and clients valuation of forecasts After the agent observes the consensus prior and the private signal, her posterior
expectation of yj is
E( yj jC j; s j ) = (sj ¡ C j) 5 p
p +§ (1) where a higher-precision private signal receives a higher weight. Notice that the
variance of the di¤erence between an agent’s unbiased beliefs and the prior consensus
is increasing in p:
V ar[ E(yj js j; C j)] = p
(p + §)§ (2) That is, higher-ability agents have opinions that are more di¤erent from the consensus,
on average. We will also refer to this variance as the information content of an agent’s
forecasts, since it is equal to the reduction in mean-squared error in the expection of
yj or Aj due to the agent’s private information: 2
V ar[ E(yj jsj ; Cj )] = V ar (yjj Cj ) ¡ V ar [yj ¡ E(yj js j ; Cj )]. (3) A mean-variance client with early access to a forecast investing in a security with
returns that are linear in yj will invest to maximize
max Ij ¢ E(yj jxj ) ¡ Ij ¢
¢ V ar(yj jxj ),
2 (4) where Ij is the client’s exposure to yj and r is the coe¢cient of absolute risk aversion.
The optimal investment is:
Ij = E (yj j xj)
r ¢ V ar(yj jxj ) (5) and such an investment has an ex-ante certainty-equivalent value of
CE = [E (yj j xj )] 2
2r ¢ V ar(yj jxj ) (6) The value of early access to an agent’s forecasts will be proportional to the variance
of E(yj jxj ). If an agent’s forecasts are fully revealing of her signal, this will be the
same as V ar [E (yj js j ; Cj )] = a, otherwise it will be a less a discount for uncertainty
regarding an agent’s forecasting strategy. For simplicity, we will assume that the
2 This is true since by the law of iterated expectations, y j ¡ E (y j jsj ; Cj ) must be uncorrelated with E (y j jsj ; Cj ) ¡ E (yj jCj ). 6 i nformation lost in communicating the signal is small in the long run and thus that
clients are interested in estimating a as the long-run value of a client’s forecasts, so
b = b.3
2.1.2 Solution We will look for a consistent-exaggeration forecasting equilibrium in which agents use
the forecasting rule: xj = b(a; ¸) ¢ E (yj j sj ; Cj ) with some constant exaggeration factor
b that can depend on a or ¸. This forecasting rule implies that agents exaggerate
their di¤erences with the consensus when b > 1, herd when b < 1, and report their
expectation when b = 1.
The standard solution approach would be to solve for a Bayesian equilibrium in
which the analyst chooses her forecasts to maximize expected utility, given the clients’
valuation rule, and the valuation rule produces an unbiased and e¢cient estimate,
given the analyst’s forecasting rule. Unfortunately, the standard Bayesian estimate of
v(x ; y ) = E[ ajx ; y; g () ; b()] is very intractable and conjugate prior distribution families that improve tractability do not exist. Instead of assuming that clients make a
such a di¢cult calculation, we will instead assume that they use an econometric estimation approach that is consistent but not necessarily e¢cient, with the ine¢ciency
coming from the incorporation of the information from the prior distribution in an
approximate rather than a fully Bayesian way. We will describe an equlibrium in
which clients do econometric estimation that anticipates consistent exaggeration and
then show that consistent exaggeration is in fact optimal for the analysts.4 This ex3 If we relaxed this assumption, agents would face an additional incentive for limiting the un- certainty regarding their exaggeration strategy, since uncertainty about exaggeration creates a gap
between a = V ar[E (yj jsj ; C j )] and v = V ar[E ( yj jx j )]. Relative to the solution described below,
the agent would choose an exaggeration factor slightly closer to their clients’ prior expectation of
4 Actually, a consistent-exaggeration equilibrium will be the only equilibrium whenever ¸ > 0 . To 7 ercise can be viewed as the derivation of an Bayesian equilibrium in which clients are
constrained by bounded rationality and thus use an unbiased and tractable but less
e¢cient estimation approach. Alternatively, it can be viewed as merely an analysis of
the incentives created for analysts if clients estimate their ability using a particular
We proceed by specifying the clients’ estimator of analyst ability, solving the
analyst’s problem given the clients’ valuation rule, and verifying that consistentexaggeration forecasting is an equilibrium.
Clients’ problem If analysts follow the consistent-exaggeration forecasting rule
described above, they will issue forecasts such that:
( s ¡ Cj ) » N (0; a)
(sj ¡ Cj ) » N (0; b2 a)
p+§ E(yj jsj ; Cj ) =
xj yj = b¡1 xj + "j
"j = yj ¡ E(yj js j; C j) .
As a result, xj and yj are distributed joint normally:
6 yj 7
6 a + V + (§ + p)
5 » N (0; 4
b Note that a is the information content of the analyst’s forecasts V ar[ E(yj jxj )]; clients
are therefore trying to estimate a.
see this consider an alternative equilbrium where clients expect an analyst to exaggerate by b ¢ f (j ) ,
where f varies predictably from observation to observation. In this case, the …rst-order condition in
(12) below will be b = [E (b CL ) + f ¡ 1 ¸][E (¯ CL )2 ¡ °¯ 2 + ¸ ]¡1 , so the analyst will choose a lower
0 b when f is high and vice versa, i.e. will choose a strategy closer to consistent exaggeration than
they are expected to. It follows that a consistent exaggeration is the only equilibrium (when clients
deviate from Bayesian estimation as described above). 8 Given the regression-like setup, it will be convenient to discuss ¯ = b¡1 as the
change in the expectation of y for a given change in x. A natural set of estimators
for ¯ = b¡1 and a are the classical statistical estimators that are used to estimate
exaggeration and forecast information content in Zitzewitz (2001):
b CL = P=1 xjyj
j =1 xj
( J=1 xjyj )2
bCL = ¯ CL ¢ V ar(xj ) =
( J ¡ 1) J=1 x2
(8) These estimators are consistent and unbiased, but they are ine¢cient if clients have
prior information about a or b(a; ¸). A client can improve the e¢ciency of an estimate
of a by averaging the observed bCL with the mean of her prior distribution. In
addition, since the estimate bCL is likely to be noisier than the estimate V ar(xj ),
especially when V is high, clients can improve on the e¢ciency of their estimate b
by constructing their b with an average of the observed ¯ CL and ¯ 0, client’s prior
expectation of b¡1 given the distribution g (a; ¸) and the function b(a; ¸) . We therefore
assume that clients estimate analysts’ ability as:
(¯ CL + ° ¢ ¯ 2) ¢ V ar (xj ) + ± ¢ a0
1+°+± (9) where ± is the weight given the prior expectation of ability and ° > 0 is the weight
placed on ¯ 0.
The ° term in (9) captures an important feature of any optimal estimator, namely
that prior information about exaggeration should be relied on when estimates of
exaggeration are noisy. Equation (9) is similar in structure to a maximum-likelihood
estimator of ability; an analysis of the maximum likelihood estimator in Appendix A
yields some intuitive predictions about the determinants of ° . The weight ° placed
on ¯0 should be higher when the estimate ¯ is noisier i.e. when earnings realizations
are noisier (and V is higher) or when the number of observations J is low.5
5 We can also think of ° as capturing the potential for exaggeration to signal high or low ability. 9 Analyst’s problem Analysts choose their xj to maximize their expectation of
bP ¡ ¸ ¢ M SE , their clients’ estimation of their forecast value plus their incentive for absolute accuracy. If clients use the estimation approach outlined above, the analyst’s
( J=1 xj yj )2
(yj ¡ xj )2
max E[ + °¯2 j=1 ¡ ¸E [ j=1,
(J ¡ 1)
(J ¡ 1)
(J ¡ 1) j =1 x2
j (10) where the expectations are the analyst’s before y is realized. The …rst-order condition
for each xj is:6
xj = b
E (yj ¯ CL) + E (yj) ¢ ¸
E (¯ CL ) + ¸
= E(yj ) ¢
E(¯ CL)2 ¡ °¯ 2 + ¸
E(¯ CL) ¡ °¯2 + ¸
0 (11) So a consistent-exaggeration strategy of the type assumed above is in fact optimal.
The analyst chooses b such that:
¯ =b ¡1 E (b CL)2 ¡ °¯2 + ¸
¯ CL + ¸ (12) From (7) above we can see that rational expectations on the part of the analyst imply that E (b CL ) = ¯ . This condition together with (8) implies the following
¯ relationships between accuracy incentives (¸), the weight placed on ¯ 0 (° ), and exaggeration ( ¯):
² When ° = 0 (no prior information about ¯) and ¸ = 0 (no incentive for accuracy), any value of ¯ is possible. This is a cheap talk result: if there is no
incentive for absolute accuracy, any language is as good as the next so long as
the clients do not have prior beliefs about ¯ .
When the distribution g( a; ¸ ) and the function b (a; ¸) are such that the priors on a and b (a; ¸ ) are
positively (negatively) correlated, then exaggeration signals low (high) ability. Clients can account
for this in their estimation by lowering (raising) ° relative to its value when the priors on a and
b (a; ¸ ) are uncorrelated.
6 Although it might appear that this simpli…cation was made using the incorrect assumption that
E (yj b C L) = E (y j ) ¢ b CL and E (¯ CL ) = E (¯ CL )2 , actually, the two cross terms exactly cancel.
¯ 10 ² When ° = 0 and ¸ > 0, ¯ = 1. Adding even a small incentive for absolute
accuracy to the cheap talk situation makes unbiased forecasting optimal.
² When ° > 0 and ¸ · °¯2 , ¯ = 0 and agents exaggerate by an in…nite factor.
With no or a limited incentive for absolute accuracy and with clients placing
some weight on their prior belief, analysts can always increase estimates of their
ability by exaggerating more. Analysts essentially report a binary forecast (i.e.,
above or below the consensus) and cannot credibly communicate the strength
of their beliefs.7
² When ° > 0 and ¸ > °¯ 2, ¯ = ¸¡1(¸ ¡ °¯ 2) < 1. Some exaggeration occurs,
but it is limited to a …nite amount by the incentive for absolute accuracy. As
this incentive increases, the amount of exaggeration decreases. Likewise, as J
decreases or random variable realizations become noisier and thus clients rely
more on their prior beliefs about ¯0 , exaggeration increases.8 2.2 Three-period model In the two-period model above, the analyst makes all J forecasts before seeing any
of the realizations. In this section, we analyze how an analyst’s past forecasting
performance a¤ects her future exaggeration. We extend the model by assuming that
agents observe the realizations of the …rst J variables and then forecast a second group
8 This result is similar to the in…nite exaggeration result in Ottaviani and Sorensen (2000).
Notice that, given the estimation approach assumed above, b( a; ¸) is a function of only ¸ . This implies that when clients’ prior beliefs g (a; ¸) imply that a and ¸ are uncorrelated, then exaggeration
will signal neither high or low ability. If instead clients’ believe that high-ability agents face more
(less) exogenous incentives to exaggerate, then exaggeration will signal high (low) ability. When
ability (a) and exogenous incentives to exaggerate (¸ ) are more positively correlated, exaggeration
signals high ability and ° and the equilibrium amount of exaggeration increase. 11 of K variables. Clients expect consistent exaggeration within a group of random
variables, but not necessarily across groups.
Clients estimate ability as before, except that they allow for di¤erent exaggeration
in the two sets of observations: P
( J=1 wxj yj + J=J +1 xjyj )2
bP = (1 + ° + ±)¡1 PJ
PJ +K 2
j=1 w xj +
j=J +1 xj
PJ +K 2
j=1 xj +
j=J +1 xj
1 +° + ±
(J + K ¡ 1)
E (¯K j b J )
¯ where ¯J and ¯ K refer to the ¯ for the …rst J and the second K random variables,
respectively. The weight w is the ratio between the observed exaggeration in the
…rst set and the clients’ expectation of exaggeration in the second set of random
variables, conditional on the observed ¯ J . The …rst order condition for the agent
when forecasting the second set of variables reduces to:
¯K = b¯
wE(¯ JK j bJ )2 ¡ °¯ 2 + ¸
b JK jb J ) + ¸
wE (¯ ¯ (13) where b JK is the exaggeration factor estimated across both sets of observations, which
will be an average of w ¡1b J and ¯ K .
¯ Proposition 1 Equation (13) implies that agents will choose a ¯ K that is between
the one-period optimal ¯ 1 = ¸¡1 (¸ ¡ °¯ 2) and their client’s expectation E(¯ K j¯ J ):
Proof. In Appendix B Agents who have had bad luck in the past and realized a lower b J than the ¯J
¯ they intended will choose a lower ¯ K . This can be interpreted as the agents who have
had bad luck in the past and are thus under-rated will exaggerate more in order to
increase the relative weight of the later observations.9
9 This prediction that under-rated agents should exaggerate more is also made by a di¤erent model in Graham (1999). 12 3 Testing predictions of the model The model in section 2 has three cross-sectional predictions about when we should
expect more exaggeration. Agents should exaggerate more when they are underrated
by their clients, when earnings realizations are expected to be noisy, and when they
expect to make a limited number of future forecasts. In this section of the paper we
test these predictions using the I/B/E/S analyst earnings forecast dataset and the
methodology for measuring exaggeration outlined in Zitzewitz (2001). 10 Speci…cally,
we estimate the average exaggeration coe¢cient for a speci…c group of forecasts using
A ¡ C = ® + ¯ (F ¡ C ) + " (14) as in the classical estimator described in section 2.1.2, where A is the I/B/E/S actual
earnings for a given …rm-quarter combination, F is a forecast of earnings, and C is an
econometric expectation of earnings based on prior forecasts for that …rm-quarter.11
We test the predictions for how exaggeration should vary with a speci…c variable by
interacting the right-hand side of (14) with the variable of interest.
As Zitzewitz (2001) shows, the regression in (14) produces an unbiased estimate of
the inverse of the exaggeration factor, ¯ = b¡1, because the error term is the analyst’s
10 As Zitzewitz (2001) argues, inferring exaggeration or herding from forecast dispersion, a popular methodology (e.g., Hong, Kubik, and Solomon, 2000), is problematic in that it does not control for
forecast information content. In other words, a forecaster can forecast close to the consensus for
two reasons: 1) if she is herding, but also 2) if she does not have very much independent private
information. Since the amount of independent private information many vary with our cross-sectional
variables, we use the methodology in Zitzewitz (2001) which controls for it.
The exaggeration measurement methodology, including the methodology for measuring C , is
described in more detail in Zitzewitz (2001). A non-econometric methodology for estimating C
(such as using the mean of all outstanding or the most recent forecasts) can be substituted without
materially a¤ecting the results. All earnings variables are normalized by the share price. 13 expectational error at time of forecasting, " = A ¡ E (Aj s; C ), and expectational errors
must be mean zero with respect to all variables known at time of forecasting, including
F ¡ C . In order for interaction versions of this regression to be valid, the analyst’s
expectational error must be mean zero with respect to the interaction variable as
well. This must be true for all variables that are known at time of forecasting, but
for variables that incorporate the econometricians knowledge of the future, we will
need to verify that the orthogonality condition still holds. 3.1 Under-rated analysts Since it is impossible to directly observe which analysts have true ability that is higher
than their measured ability, we are forced to use our knowledge of the future to help
identify under-rated analysts. In particular, we divide analysts with a given past
forecast information content into those whose performance eventually rises and those
whose performance falls and assume that, on average, the analysts whose performance
rises were under-rated in the past.
Speci…cally, we rank analysts with at least 50 past forecasts based on two variables: their past forecast information content from observation 1 to j ¡ 1 and the
di¤erence between their past information content and their information content from
observation j + 1 to j +50 (or fewer if the analyst leaves the sample). We then interact
these rankings with the right-hand side of equation (14).
Table 1 presents the results of such an interaction regression. The results suggest
that under-rated analysts exaggerate more, whether or not past performance is controlled for. The results also do not change if we control for the analyst’s career length
or the size of their brokerage, both of which have signi…cant positive e¤ects on ¯ .
In constructing this test, we took two steps to avoid violating the orthogonality
condition discussed above. First, we used di¤erent observations to measure perfor14 mance improvement (observations j + 1 to j + 50) and exaggeration (observation j ).
This is important since if an analyst “gets lucky” and gets surprised in the direction
of their deviation from the consensus (i.e., A ¡ E and F ¡ C positively correlated), ¯
will be overestimated, exaggeration will be underestimated, and analyst performance
will be overestimated. Using di¤erent observations avoids this potential problem.
Second, we used bCL = ¯ 2V ar(x) as our performance measure, a measure that is
a robust to exaggeration, so even if exaggeration in observation j were correlated with
exaggeration in observations j + 1 to j + 50 or 1 to j ¡ 1, this would not create a
correlation with our measure of the change in performance. 3.2 Expected earnings uncertainty We test for whether analysts exaggerate more when earnings are uncertain using a
two-step process (Table 2). In the …rst step, we predict the average absolute earnings
surprise (actual less consensus) for a particular …rm-quarter based on market cap, the
standard deviation of prior outstanding forecasts, and the prior average absolute earnings surprise for the …rm in question. We hypothesize and …nd that average absolute
earnings surprise is higher for small-cap …rms, when past forecasts are dispersed, and
for …rms for which average earnings surprise has been large in the past. In the second
step, we test the e¤ect of expected earnings surprise on ¯ using an interaction regression, …nding that there is signi…cantly more exaggeration when predicted absolute
earnings surprise is higher. Notice that in this analysis all of the interaction variables
are known at time of forecasting; thus the orthogonality condition should be satis…ed. 3.3 Expected career length The prediction that analysts should exaggerate less when they expect to make more
future forecasts is more di¢cult to test. We can use the actual number of future
15 forecasts as our interaction variable, and when we do this, we …nd less exaggeration
by analysts who make more future forecasts (Table 3, Panel A). A problem with this
analysis, however, is that analysts who have good luck should both survive longer
and have measured exaggeration that is less than what they intended.
An alternative approach is to use variables that are known at time of forecasting
that predict an analyst’s longevity. Probit regressions that predict an analyst’s leaving
the sample and not returning for at least 2 years after a given forecast …nd longer
survival is expected for analysts who have made a large number of past forecasts,
analysts who work at larger (usually the more prestigous) brokerages, and analysts
who have had better forecast accuracy in the past (Table 3, Panel B).12
In Table 1, we found that analysts with more forecasting experience and analysts at larger brokerage …rms exaggerated less. The probit regressions suggest that
one potential explanation for this result is that these analysts have longer expected
careers, and the optimal exaggeration rate for these analysts is lower. Alternative
explanations exist, however. Analysts may become less overcon…dent in their own
information or better calibrated with experience. In the model we assumed that analysts’ utility is linear in the market valuation of their forecasts; if the concavity of
analyst’s incentives varies with brokerage size or career length, this may also explain
the results. Inexperienced analysts may face greater outside options and thus more
convex incentives (i.e., they can gamble and then leave if it does work out), and this
may explain their greater exaggeration. Analysts at larger …rms may be given more
concave incentives by their …rm to reduce exaggeration, such as a risk of getting …red
12 Forecast information content, in turn, does not appear to play a role in predicting exits from the I/B/E/S sample. One potential explanation for this result is that analysts can leave the I/B/E/S
sample for either good reasons (moving to lucrative proprietary researcg positions) or bad reasons
(getting …red). We do …nd that for analysts who stay in the profession, forecast information content
helps explain which analysts are ranked highly by Institutional Investor (Zitzewitz, 2001, Table 7). 16 for deviating from the consensus and being wrong that is not fully compensated by
the reward for deviating from the consensus and being right. A model in which …rms
have a collective reputation for exaggeration might predict this, since larger …rms will
make more forecasts in the future than smaller …rms and thus would prefer that their
analysts exaggerate less.
In summary, the empirical evidence that is available is consistent with the prediction that analysts who expect to make more future forecasts should exaggerate less,
but alternative explanations for the results exist. 4 Conclusion The evidence presented in Zitzewitz (2001) suggests that there are persistent di¤erences in analyst’s exaggeration factors and forecast information content and that the
best predictor of the future value of an analyst’s forecasts is the value of her past
forecasts. This suggests that potential clients should use an analyst’s track record
to determine how much to pay her. This paper investigates the incentives for exaggeration created by clients attempting to learn ability from forecasting record in a
…nancial market environment where forecasts are valuable for their new information
We …nd that career concerns can create an incentive for agents to exaggerate,
or overweight their private information. This incentive exists because high-ability
analysts have viewpoints that are more di¤erent from the consensus on average and
since potential clients learn more quickly about an analyst’s average di¤erence with
the consensus than about whether she is exaggerating. The equilibrium exaggeration
rate is …nite so long as there is a su¢ciently large external incentive for absolute
forecast accuracy. The model also predicts that agents should exaggerate more when
earnings are expected to be noisy, when they expect to make a limited number of
17 future forecasts, or when they are under-rated by the market, and we …nd that these
predictions are consistent with the equity analyst forecast data.
Although the evidence in the paper is for equity analysts, the issues examined
in this paper potentially apply to other opinion-producing agents. A large number
of agents produce opinions that can be thought of as forecasts of random variables.
Especially when the actions taken based on the opinions are strategic substitutes, the
value of privileged access to an opinion depends on its information content relative
to the consensus. Whenever the realized values of random variables are noisy, agents
will learn more quickly about the average di¤erence between an agent’s opinion and
the consensus that they will about whether the agent is exaggerating, an the agent
will be able to raise estimates of her ability by exaggerating. This incentive to exaggerate will be greater when the realization of the random variable is expected to be
more noisy, which makes exaggeration harder to detect, when the agent expects to
leave the profession soon, or when the agent perceives that she is under-rated by the
market. These predictions, together with the empirical support for them in the analyst data, are potentially useful for consumers attempting to account for exaggerate
in their interpretation of opinions or for …rms attempting to reduce exaggeration in
the incentives they design for opinion-producers. A Maximum likelihood estimation In this appendix we assume that rather than calculating expectations for a and b,
clients calculate maximum likelihood estimates. We can think about combining the
prior distribution g (a; ¸) and the function b(a; ¸) into a prior on a and ¯ = b¡1, which
we will call f (a; ¯). For tractability, we will also assume that the prior distribution
g (a; ¸) is such that the distribution f (a; ¯ ) is concave in logs for both variables,
i.e. d2 ln f ( a;¯ )
da2 · 0 and d2 ln f (a;¯ )
d¯ 2 · 0 8x, a condition that is satis…ed by the normal
18 and chi-squared distributions, for example. We also assume that V = V ar(yj jxj ) is
Given these assumptions, the log likelihood function is:
ln L(x; y; a; ¯ ) = J
X J ln Á ( j=1 X
y ¡x ¯
ln Á( j 1=2j ) + ln[ f (a; ¯ )],
j=1 where Á (¢) is the standard normal p.d.f. The maximum likelihood estimators of a and
b satisfy the conditions:
¯ PJ fa(bMLE ; bMLE ) b3=2
f (bMLE; ¯ MLE)
V ¡1 j=1 xjyj f¯ (bMLE ; b MLE )
= (1 +
) [ PJ
f (bMLE; b MLE )
j =1 j
¯ MLE We have assumed that fa
j =1 xj is monotonically decreasing in a, so the second term in the expression for a will be positive (negative) when a is less (greater) than the
mode of the prior distribution. The maximum likelihood estimate of a will thus be
between b MLE V ar(xj ) and the mode of the prior. The …rst factor in the maximum
¯ likelihood estimate of ¯ induces a bias toward zero that is a artifact of our taking a maximum likelihood approach to estimation. Ignoring this factor, the maximum
likelihood estimate of ¯ will be between b OLS and the mode of the prior, with more
weight being placed on the prior when V is large. Thus the MLE is similar in structure to the estimating approach assumed in (9):
(¯ CL + ° ¢ ¯ 2) ¢ V ar (xj ) + ± ¢ a0
1+°+± The estimate of ¯ is some weighted average of the observed ¯ CL and the mean of
the prior distribution ¯ 0. Ability is estimated in turn as some weighted average of
¯ MLEV ar(xj ) and the mean of the prior distribution. The weight placed on the prior
belief about exaggeration, ° , is increasing in V and decreasing in J , i.e. it is higher
when realizations are noisy or when the number of observations is small.
19 B Proof of Proposition 1 b
We know that E (b JK j¯ J ) is between w¡1 bJ and ¯ K . Since w ¡1 bJ = E (¯ Kj b J ), this
implies that ¯K will be above E(¯ JK j¯ J ) when it is above client’s expectation based
on the observed bJ . If we de…ne 4¯ = ¯K ¡ E (b JK j¯ J ), we can rewrite the …rst
order condition as E( bJK j bJ ) + 4¯ =
E (bJK j b J ) =
wE (b JK j¯ J )2 ¡ °¯ 2 + ¸
bJK j¯ J ) + ¸
¸(1 ¡ 4¯ ) ¡ °¯2
¸ + w ¢ 4¯ 2
bJ Kj b J ) + 4¯ = ¸ ¡ °¯0 + w (4¯)
= E (¯ ¯
¸ + w ¢ 4¯
bJK j bJ ) ¢ 4¯
= ¯1 ¡
¸ If we de…ne ¯1 = ¸¡1(¸ ¡ °¯ 2) to be the optimal one-period beta, we can multiply
both sides of the above expression by ¸¡1(¸ + w ¢ 4¯ )to get:
¯K = ¯1 ¡ bb
wE(¯ JK j¯ J ) ¢ 4¯
¸ This implies that ¯ K is less than ¯ 1 if and only if it is greater than E (bJK j b J ) and
E (¯K j b J ).
¯ 20 References
Avery, Christopher N. and Judith A. Chevalier, “Herding Over the Career,”
Economics Letters, 1999, 63, 327–333.
Banerjee, Abhijit, “A Simple Model of Herd Behavior,” Quarterly Journal of Economics, 1992, 107, 797–817.
Bikchandani, Sushil, David Hirshleifer, and Ivo Welch, “A Theory of Fads,
Fashion, Custom, and Cultural Change as Informational Cascades,” Journal of
Political Economy, 1992, 100, 992–1026.
Brandenburger, Adam and Ben Polak, “When Managers Cover Their Posteriors:
Making Decisions the Market Wants to See,” RAND Journal of Economics, 1996,
Chevalier, Judith A. and Glenn D. Ellison, “Risk Taking by Mutual Funds as
a Response to Incentives,” Journal of Political Economy, 1997, 105, 1167–1200.
and , “Career Concerns of Mutual Fund Managers,” Quarterly Journal of Economics, 1999, 114, 389–432.
Cooper, Rick A., Theodore E. Day, and Craig M. Lewis, “Following the
Leader: A Study of Individual Analysts Earnings Forecasts,” 1999. Vanderbilt
E¢nger, Matthias R. and Mattias K. Polborn, “Herding and Anti-herding: a
model of reputational di¤erentiation,” 2000. University of Munich Mimeo.
Ehrbeck, T. and R. Waldmann, “Why Are Professional Forecasts Biased? Agency
Versus Behavioral Explanations,” Quarterly Journal of Economics, 1996, 111,
21 Fama, Eugene, “Agency Problems and the Theory of the Firm,” Journal of Political
Economy, 1980, 88, 288–307.
Gibbons, Robert and Kevin J. Murphy, “Optimal Incentive Contracts in the
Presence of Career Concerns: Theory and Evidence,” Journal of Political Economy, 1992, 100, 468–505.
Graham, J. R., “Herding among Investment Newsletters,” Journal of Finance, 1999,
Holmstrom, Bengt, “Managerial Incentive Problems: A Dynamic Perspective,”
Review of Economic Studies, 1999, 66, 169–182.
and Joan Ricart i Costa, “Managerial Incentives and Capital Management,”
Quarterly Journal of Economics, 1986, 101, 835–860.
Hong, Harrison, Je¤rey D. Kubik, and Amit Solomon, “Security Analysts’
Career Concerns and Herding of Earnings Forecasts,” RAND Journal of Economics, 2000, 31, 121–144.
Klein, Benjamin and Keith B. Le-er, “The Role of Market Forces in Assuring
Contractual Performance,” Journal of Political Economy, 1981, 89, 615–641.
Kutsoati, E. and D. Bernhardt, “Can Relative Performance Evaluation Explain
Analysts’ Forecasts of Earnings,” 1999. Tufts University Mimeo.
Laster, David, Paul Bennett, and In Sun Geoum, “Rational Bias in Macroeconomic Forecasts,” Quarterly Journal of Economics, 1999, 114, 293–318.
Nelson, Philip, “Advertising as Information,” Journal of Political Economy, 1974,
82, 729–754. 22 Ottaviani, Marco and Peter Sorensen, “Professional Advice,” 2000. University
College London Mimeo.
Prendergast, Canice and Lars Stole, “Impetuous Youngsters and Jaded OldTimers: Acquiring a Reputation for Learning,” Journal of Political Economy,
1996, 104, 1105–1134.
Scharfstein, David and Jeremy Stein, “Herd Behavior and Investment,” American Economic Review, 1990, 80, 465–479.
Schmalensee, Richard, “A Model of Advertising and Product Quality,” Journal of
Political Economy, 1978, 86, 485–503.
Shapiro, Carl, “Optimal Pricing of Experience Goods,” Bell Journal of Economics,
1983, 14, 497–507.
Trueman, Brett, “Analyst Forecasts and Herding Behavior,” Review of Financial
Studies, 1994, 7, 97–124.
Welch, Ivo, “Herding Among Security Analysts,” 1999. UCLA Mimeo.
Zitzewitz, Eric, “Measuring herding and exaggeration by equity analysts,” 2001.
Zweibel, J., “Corporate Conservativism and Relative Performance Compensation,”
Journal of Political Economy, 1992, 103, 1–25. 23 Table 1. Exaggeration by under or over-rated analysts
Dependent variable: (ACT - CONS) This table reports interaction coefficients from a version of equation (14) with the variables listed interacted with the righthand side. Regressions also include (FOR - CONS) and the interaction variables. Regressions include forecasts made by
analysts with at least 50 past forecasts in the 1993-99. The UNDER and INFO variables are rankings, scaled to 0 to 1, of
the analyst based on past performance and performance change over the next 50 forecasts, respectively. Spec.
329,401 Variable defintions
0.133 Interactions with (FOR - CONS)
0.136 -0.046 0.141 0.048
0.040 Actual earnings per share
Analyst's forecast of earnings per share
Prior expectation of earnings per share, as estimated in Table 3
Ranking based on change in analyst forecast value over the next 50 forecasts, scaled 0 to 1.
Ranking of analyst's historical forecast value, scaled 0 to 1.
Forecast number in analyst's career
Number of analysts at analyst's brokerage Standard errors are heteroskedasticity robust and adjusted for clustering within firm-quarter combinations. Table 2. Exaggeration and uncertainty
The effect of expected earnings uncertainty is examined by predicting the absolute
earnings surprise for a firm-quarter combination based on its market cap, the standard
deviation of price-normalized forecasts, and past earnings surprise for the firm, and
average earnings surprise in the prior 90 days. We then interact predicted earnings
surprise with the right-hand side of (2) to measure the effect on exaggeration. The
negative interaction coefficient reported implies more exaggeration when expected
earnings surprise is high. Coeff.
First stage regression
Dependent variable: Abs(ACT - CONS)
Avg past abs(ACT - CONS) for firm
Avg abs(ACT - CONS) in quarter
Interaction coefficient from exaggeration regression
Predicted abs. forecast error S.E. -0.0184
0.007 -0.059 0.022 Notes:
1. Standard errors are heteroskedasticity robust and adjusted for clustering within firmquarter combinations. Standard error in second stage is adjusted for the use of a
predicted value on the right-hand side. Table 3. Exaggeration and expected future forecasts In panel A, the right hand side of (14) is interacted with the actual number of
future forecasts an analyst makes between the current forecast and the end of
1999. Forecasts for the years 1993-97 and for analysts who have already
made 50 forecasts are included in the sample. In Panel B, exit from the
I/B/E/S sample (defined as making a final forecast and not reappearing in the
sample for 2 years) is predicted for each forecast in the 1993-97 period. Past
average relative forecast ranking is the average of a 0-1 ranking of analysts'
relative forecast accuracy for each firm-quarter in which they forecast. Panel A. Interaction regression with actual future forecasts
Dependent variable: actual earnings less consensus Forecast less consensus
(FOR - CONS)*(Actual future forecasts/100)
Constant (in basis points)
Actual future forecasts/100 (in basis points)
198,909 Panel B. Probit regression predicting exit from sample Forecasts in career/100
Ln(Analysts at brokerage)
Ln(Forecast information content)
View Full Document
This note was uploaded on 07/21/2011 for the course BUS 10001 taught by Professor All during the Spring '11 term at Shaheed Zulfiqar Ali Bhutto Institute of Science and Technology.
- Spring '11