How Skilled Are Security Analysts?

• Author: ALAN CRANE,KEVIN CROTTY
• Published date: 01 June 2020
• DOI: http://doi.org/10.1111/jofi.12890
• Date: 01 June 2020

THE JOURNAL OF FINANCE •VOL. LXXV, NO. 3 •JUNE 2020

How Skilled Are Security Analysts?

ALAN CRANE and KEVIN CROTTY∗

ABSTRACT

The majority of security analysts are identified as skilled when the cross-section of

analyst performance is modeled as a mixture of multiple skill distributions. Analysts

exhibit heterogeneous skill—some are high-type, and some are low-type. On average,

the recommendation revisions of both types exhibit positive abnormal returns. The

heterogeneity stems from differential ability to produce new information; all analysts

can profitably process news. Top analysts outperform because more of their recom-

mendations are influential (i.e., associated with statistically significant returns) and

both their influential and noninfluential recommendations are more informative. A

majority of research firms are also identified as skilled.

AFUNDAMENTAL QUESTION IN FINANCIAL economics is whether financial profes-

sionals are able to systematically identify misvalued securities. This question

is important not only for understanding the nature of market efficiency,but also

for understanding the role and value of financial intermediaries. A large body

of research speaks to this question by examining the stock recommendations

of security analysts. Overall, this literature finds that analyst recommenda-

tions and revisions to recommendations contain information about future re-

turns (e.g., Stickel (1995), Womack (1996), Barber et al. (2001), Barber,Lehavy,

and Trueman (2010)), although whether analysts produce new information or

merely interpret contemporaneous news is subject to debate (Altınkılıc¸and

Hansen (2009), Altınkılıc¸etal.(2013), Bradley et al. (2014), Li et al. (2015)).

Missing from prior literature, however, is evidence on how widespread such

skill is across analysts. The fact that recommendations contain information on

average does not necessarily mean that the median analyst is skilled, and there

is little evidence concerning the value of the best and worst analysts. Whether

analyst skill concentrates in a small subset of analysts or is ubiquitous in

the industry speaks to the efficiency of markets as well as to the value of

equity research for market participants. The value of analysts depends on

both the extent of their skill and the nature of their informational ability

(i.e., information processing versus information production). Such valuations

∗Alan Crane and Kevin Crotty are with the Jones Graduate School of Business, Rice University.

The authors thank Stefan Nagel (the Editor), an anonymous associate editor, Brad Barber, Cam

Harvey, K. Ramesh, and an anonymous referee for helpful comments. The authors thank Seyed

Kazempour for excellent research assistance. The authors have no conflicts of interest to disclose.

Correspondence: Kevin Crotty, Jones Graduate School of Business, Rice University, 6100 Main

Street MS 531, Houston, TX 77005; e-mail: kevin.p.crotty@rice.edu.

DOI: 10.1111/jofi.12890

C2020 the American Finance Association

1629

1630 The Journal of Finance R

are particularly important in light of recent regulation, namely, MiFID II in

Europe (effective January 2018), which requires institutional investors to pay

directly for equity research.

In this paper, we estimate the fraction of equity analysts who are skilled or

unskilled at making stock recommendations. We also estimate the distribution

of the economic magnitude of analyst skill. Identifying skilled analysts is chal-

lenging. Even in a world in which analysts make recommendations randomly,

we would expect some variation in measured ability simply due to chance. A

given analyst’s recommendations may outperform due to good luck rather than

superior information. Similarly, skilled analysts may underperform due to bad

luck rather than true inability. To determine the fraction of skilled analysts,

accounting for noise due to luck, we model analyst performance as a mixture

distribution, using methods similar to those employed in recent studies on mu-

tual funds (Harvey and Liu (2018)) and hedge funds (Chen, Cliff, and Zhao

(2017)). These studies find that 10.6% of mutual funds and 66.6% of hedge

funds are skilled. In our implementation, analysts belong to one of two skill

groups. The abnormal performance of each of these skill groups is centered at

a different level of abnormal returns, which is also estimated.1Our estimation

uses information from the cross-section of analyst performance to reduce the

noise inherent in estimating analyst-level performance.

Our primary result is striking. A large fraction of the almost 5,500 equity

analysts studied between 1993 and 2015 appear skilled. When an analyst’s per-

formance is measured using the average abnormal return associated with their

recommendation revisions, the mixture model indicates that 97% of analysts

have positive true abnormal returns. This estimate is higher than the fraction

of analysts with positive point estimates of performance (85%), so the use of

cross-sectional information in assessing performance implies that some ana-

lysts with negative measured performance are unlucky rather than unskilled.

The average magnitude of analyst skill is also large, with an abnormal return

of 1.5% over a [0,5] trading-day window. Roughly 64% of analysts are estimated

to come from a lower return distribution, but the average performance of these

analysts is still economically large and positive, with the distribution centered

at a 1% abnormal return. The remaining 36% of analysts perform even better,

with abnormal performance centered at 2.4%.

We further find that analyst skill is persistent. In particular, the autore-

gressive coefficient from a regression of analyst performance measured over

the second half of the sample on the performance measured in the first half is

0.24.2This estimate from the data is statistically indistinguishable from the

persistence of 0.27 implied by the structural assumptions and estimates of the

mixture model. These results suggest that the parametric assumptions under-

lying the mixture model do a good job characterizing the skill distribution.

1It is possible to extend the number of skill groups beyond two, but a model with two skill

groups results in a better fit to the empirical distribution of analyst performance than adding an

additional distribution to the mixture.

2For other evidence on analyst performance persistence, see Mikhail, Walther, and Willis (2004)

and Li (2005).

How Skilled Are Security Analysts? 1631

Given how well analysts perform, an obvious question that arises relates

to the source of this performance. In particular, does their performance stem

from an ability to forecast the pricing implications of concurrent news events,

from an ability to provide incremental information to the market, or both?

Recent research by Altınkılıc¸ and Hansen (2009) and Altınkılıc¸etal.(2013)

concludes that recommendations are not informative beyond other concurrent

information releases, but this conclusion is refuted by Bradley et al. (2014)and

Li et al. (2015), who provide evidence that analyst picks do have incremental

information content.3We contribute to this debate by testing whether skill at

the analyst level is driven by “piggybacking” on news events (i.e., information

processing), the production of new information, or both. Using the RavenPack

news database to separately estimate analyst skill from revisions made with

and without concurrent news releases, we find evidence of skill on both di-

mensions. First, all analysts appear to be skilled at forecasting drift associated

with news events. The economic magnitude of this skill is large (2%), and the

only variation across analysts on this dimension is due to estimation error. In

addition, a sizable fraction of analysts (93%) that make revisions on nonnews

days are skilled at information production as well. While the average magni-

tude of information production ability is still large (1.7%), there is considerable

cross-sectional heterogeneity in analyst ability to produce new information—

the cross-sectional standard deviation is 1.2%. Thus, the variation in overall

analyst skill is driven by differences in information production ability.

While our estimates suggest that almost all analysts are skilled, there is

substantial cross-sectional variation in the extent of that skill. Accordingly,

we next examine the analyst characteristics that are associated with higher

levels of skill. Specifically, we model the probability that a given analyst is

drawn from the higher mean distribution as a function of the number of rec-

ommendations the analyst makes, the percentage of their picks that are sells,

the concentration of the analyst’s industry coverage, and the analyst’s years of

experience. Not surprisingly, we find that more experienced analysts are more

likely to be drawn from the higher mean distribution, while those that make

more recommendations per year are more likely to be low-type, consistent with

diseconomies of scale. We further find that analysts covering more industries

are more likely to be identified as high-type analysts, which is inconsistent with

the view that more focused analysts perform better. Analysts issuing a larger

fraction of sell recommendations are significantly more likely to come from the

higher mean distribution, which is consistent with Barber et al. (2006)and

suggests that conflicts of interest may play a role in the performance (Michaely

and Womack (1999)).

The results above imply that analysts are important information intermedi-

aries who help facilitate price discovery in the securities they cover. We next

characterize the economic value of this information by estimating its dollar

3Using intraday data, Bradley et al. (2014) find that about 25% of analyst recommendations

are associated with price jumps. Li et al. (2015) incorporate after-hour revisions as well and find

that only a minority of recommendations are susceptible to the piggybacking critique.