Diagnostic Expectations and Stock Returns

• Author: ANDREI SHLEIFER,NICOLA GENNAIOLI,RAFAEL LA PORTA,PEDRO BORDALO
• Published date: 01 December 2019
• DOI: http://doi.org/10.1111/jofi.12833
• Date: 01 December 2019

THE JOURNAL OF FINANCE •VOL. LXXIV, NO. 6 •DECEMBER 2019

Diagnostic Expectations and Stock Returns

PEDRO BORDALO, NICOLA GENNAIOLI, RAFAEL LA PORTA,

and ANDREI SHLEIFER∗

ABSTRACT

We revisit La Porta’s finding that returns on stocks with the most optimistic ana-

lyst long-term earnings growth forecasts are lower than those on stocks with the

most pessimistic forecasts. We document the joint dynamics of fundamentals, expec-

tations, and returns of these portfolios, and explain the facts using a model of belief

formation based on the representativeness heuristic. Analysts forecast fundamentals

from observed earnings growth, but overreact to news by exaggerating the probability

of states that have become more likely.We find support for the model’s predictions. A

quantitative estimation of the model accounts for the key patterns in the data.

LAPORTA (1996)SHOWS THAT STOCK market analysts’ expectations of the long-

term earnings growth of the companies they cover have strong predictive power

for these companies’ future stock returns. Companies whose earnings growth

analysts are most optimistic about earn poor returns relative to companies

whose analysts are most pessimistic about. Betting against extreme analyst

optimism has thus been a good idea on average. La Porta (1996) interprets

this finding as evidence that analysts, as well as investors who follow them or

think like them, are too optimistic about stocks with rapidly growing earnings

and too pessimistic about stocks with deteriorating earnings. As a result, the

former stocks are overvalued, the latter are undervalued, and the predictability

of returns is driven by the correction of expectations.

In this paper, we revisit this puzzle using a model of Kahneman and Tver-

sky’s (1972) representativeness heuristic developed by Gennaioli and Shleifer

(2010) and Bordalo et al. (2016). Relative to the existing work, we make three

innovations. First, we rely on a portable and psychologically founded model

that has been used to describe the dynamics of beliefs in a variety of settings,

not only in the one at hand. Second, we assess the qualitative and quantita-

tive performance of this model in explaining not only a cross-section of stock

∗Pedro Bordalo is with Sa¨

ıd Business School, University of Oxford. Nicola Gennaioli is with

Bocconi University and IGIER. Rafael La Porta is with Brown University. Andrei Shleifer is with

Harvard University. Nicola Gennaioli thanks the European Research Council (GA 647782) and

Andrei Shleifer thanks the Pershing Square Venture Fund for Research on the Foundations of

Human Behavior for financial support of this research. We are grateful to seminar participants at

Brown University and Sloan School, and especially to Josh Schwartzstein, Jesse Shapiro, Pietro

Veronesi, Yang You,and the referees for helpful comments. We also thank V. V. Chari, who encour-

aged us to confront our model of diagnostic expectations with the Kalman filter. The authors have

no relevant or material financial interests that relate to the research described in this paper.

DOI: 10.1111/jofi.12833

2839

2840 The Journal of Finance R

returns, but also the paths of fundamentals leading to overvaluation and of an-

alyst expectations. Third, we test new predictions that distinguish our model

from mechanical models of extrapolation such as adaptive expectations. These

new predictions follow from the fact that in our model expectations contain a

“kernel of truth” in that they exaggerate true features of reality. Beliefs thus

move in the correct direction, but by more than the right amount.

As a first step, we look at the data (described in Section I). In Section II,we

confirm with 20 additional years of data La Porta’s (1996) finding that stocks

with low long-term earnings growth forecasts (LLTG stocks) outperform stocks

with high long-term earnings growth forecasts (HLTG stocks). Wethen present

three additional facts. First, HLTG stocks exhibit fast past earnings growth,

which slows down going forward. Second, forecasts of future earnings growth

of HLTG stocks are too optimistic, and are systematically revised downward

later. Third, HLTG stocks exhibit good past returns but their returns going

forward are low. The opposite dynamics obtain for LLTG stocks, but in a less

extreme form. We do not account for this asymmetry in our model.

The data suggest that analysts use a firm’s past performance to infer its

future performance, but overreact. This updating mechanism arises naturally

from overweighting of representative types. Gennaioli and Shleifer (2010)and

Bordalo et al. (2016) model a type tas representative of a group Gwhen it occurs

more frequently in that group than in a reference group −G. For instance,

after a positive medical test, the representative patient is t=“sick” because

sick people are truly more prevalent among those who tested positive than in

the overall population. After such a positive test, the representative sick type

quickly comes to mind and the doctor inflates its probability too much, which

may still be objectively low if the disease is rare (Casscells, Schoenberger, and

Graboys (1978)). There is a kernel of truth in departures from rationality: the

doctor overreacts to the objectively useful information from the test.

In Section III, we incorporate this model of representativeness into a problem

of an analyst learning about a firm’s unobserved fundamentals in light of a

noisy signal such as current earnings. This approach yields a distorted Kalman

filter, which we refer to as the “diagnostic Kalman filter,” that overinflates

the probability of future earnings growth realizations whose likelihood has

objectively increased the most in light of recent news. After strong earnings

growth, the probability that the firm is the next “Google” goes up. This type

becomes representative and analysts inflate its probability excessively, even

though Googles remain rare in absolute terms. As good news stops arriving,

overoptimism cools off. Rapid earnings growth thus causes overvaluation and

subsequent disappointment, leading to reversals of optimism and low returns.

In Section IV, we show that this model makes predictions consistent with the

evidence in Section II. In Section V, we perform a first pass quantitative assess-

ment of the model. We first show that expectations data point to overreaction

to news, in the sense that revisions of the forecast of LTG negatively predict

errors in that forecast. We then use the simulated method of moments (SMM)

to estimate the parameter controlling the strength of representativeness by

matching observed overreaction and the empirical moments of the earnings

Diagnostic Expectations and Stock Returns 2841

process. We find that forecasters react about twice as much to information as is

objectively warranted. This estimate lies in the ballpark of those obtained us-

ing the expectations of professional forecasters about credit spreads (Bordalo,

Gennaioli, and Shleifer (2018)) and many macroeconomic variables (Bordalo

et al. (2018)). We also find that stocks overreact to news over a time scale of

about three years. Notably, while the estimation uses expectations and fun-

damentals alone, it predicts the observed return spread between LLTG and

HLTG stocks reasonably well.

In Section VI, we test three novel predictions of the model, all following from

the kernel-of-truth property as applied to the present dynamic setting. First,

we show that the HLTG group contains a fat right tail of future exceptional

performers, and that analysts attach an excessively high probability that HLTG

firms are in that tail. There are very few Googles, but they are concentrated

in the HLTG group, and analysts exaggerate their frequency. This is precisely

what the kernel-of-truth property of expectations predicts.

Second, we show that the return spread between LLTG and HLTG stocks

widens among firms with more volatile or persistent fundamentals. This is

also in line with the kernel of truth: in both cases, good news is even more

informative about strong future performance, which renders Googles even more

representative. Third, we show that expectations about HLTG (LLTG) stocks

revert downward (upward) even in the absence of bad (good) news. This is

again in line with the kernel of truth: analyst forecasts reflect the true mean

reversion in earnings growth.

Mechanical models of beliefs, such as adaptive expectations, cannot yield

these facts, which are due to the forward-looking but not fully rational learning

that we see in the data.

Our paper follows extensive research on overreaction and volatility that be-

gins with Shiller (1981), De Bondt and Thaler (1985,1987), Cutler, Poterba, and

Summers (1990,1991), and De Long et al. (1990a,1990b). This work often uses

mechanical rules for belief updating such as adaptive expectations or adaptive

learning (e.g., Barsky and De Long (1993), Barberis et al. (2015), Adam, Marcet,

and Beutel (2017)).1Barberis, Shleifer, and Vishny (1998) is closest in spirit,

though not in formulation, to our current work, since it is also motivated by rep-

resentativeness. The authors present a model of Bayesian learning in which the

decision maker is trying to distinguish models that are all incorrect. Because

they do not model representativeness explicitly, their specification does not al-

low for a tight link between measurable reality and measurable beliefs that is

central to our theory and evidence. Daniel, Hirshleifer, and Subramanyam

(1998) and Odean (1998) model investor overconfidence, the tendency of

1Other papers include Barberis and Shleifer (2003), Glaeser and Nathanson (2015), Hong and

Stein (1999), Lakonishok, Shleifer, and Vishny (1994), Marcet and Sargent (1989), and Adam,

Marcet, and Nicolini (2016). In Adam, Marcet, and Beutel (2017), agents learn the mapping between

fundamentals and prices, but they are rational about fundamentals. Pastor and Veronesi (2003,

2005,2009) present rational learning models in which uncertainty about the fundamentals of

some firms can yield predictability in aggregate stock returns. This approach does not analyze

expectations data or cross-sectional differences in returns.