(0.048 and 1.571, respectively). Similarly, among the three recommendation revi-

sion indicators associated with negative target price revisions (column 6), the

one associated with recommendation downgrades is statistically and economic-

ally the largest. Another intriguing ¢nding is that, when partitioned by the sign

of the target price change, recommendation reiterations are associated with a

14

McNichols and O™Brien (1997), for example, report that coe„cient estimates on recom-

mendation upgrades are smaller in absolute value than those for recommendation down-

grades. The evidence from columns 2 and 3, however, suggests that these estimates are not

economically di¡erent. The reason for this discrepancy is our conditioning on the issuance

of a target price. In unreported results, we ¢nd that when we do not condition on the presence

of a target price, investor reaction to recommendation downgrades is larger in absolute value

than to upgrades.

15

The magnitude of the slopes on the target price revision in columns 5 and 6 are both low-

er than the slope reported in column 1. Whereas in column 1 we specify a linear relation be-

tween target price revisions and abnormal returns, the regressions in columns 5 and 6 allow

the coe„cient estimates on the stock recommendation to vary depending on the sign of the

target price revision, thus exploiting the information in target price revisions as well. Indeed,

when we estimate a regression similar to the one in column 1, but with the inclusion of a

slope indicator variable that assumes the value of 1 for positive target price revisions and 0

for negative ones, the coe„cient estimates on positive target price revisions is 3.095 compared

to 4.375 on negative target price revisions (the di¡erence is statistically signi¢cant, with a p-

value ¼ 0.001).

An Empirical Analysis of Analysts™ Target Prices 1949

large and signi¢cant abnormal return. For instance, when issued along with a

positive target price revision, reiterated recommendations are informative as

the associated intercepts are economically and statistically signi¢cant

(a3 ¼ 1.571, t-statistic ¼ 29). Similar evidence is obtained for reiterations accompa-

nied by negative target price changes.

In additional unreported tests we have also examined the informativeness of

target price revisions using event-day abnormal volume. Following Holthausen

and Verrecchia (1990), who argued that abnormal volume and abnormal returns

are equally relevant means of assessing information content, we have calculated

for every ¢rm and event in our sample an abnormal volume measure and then

repeated the analysis reported in Table III. Speci¢cally, we regress the absolute

value of abnormal volume on target price revisions controlling for recommenda-

tion and earnings forecast revisions. Consistent with our earlier results we ¢nd

that changes in target prices are positively related to abnormal volume. Indeed,

target price revisions lead to the highest abnormal volume when issued with re-

commendation downgrades. Moreover, abnormal volume is highest when the di-

rection of the target price and recommendation revisions coincide.

Taken together, the evidence presented in Table III supports the hypothesis

that target prices are informative, both unconditionally and conditional on stock

recommendation and earning forecast revisions. We ¢nd that target price revi-

sions are deemed more informative when they are negative and when associated

with recommendation downgrades. We also ¢nd that investor reaction is the

strongest when the direction of the target price and recommendation revisions

coincide rather than when they di¡er.16

C. Postevent Abnormal Returns

The preceding analysis is based on the assumption that investors respond

quickly and rationally to the information conveyed in the analyst reports. Since

some studies (e.g., Stickel (1995),Womack (1996), Barber et al. (2002)) have shown

that market reaction to announcements of recommendation changes is incom-

16

We have performed three additional tests. First, we estimate regressions in which we con-

dition on the sign of the earnings forecast revision. We ¢nd that when earnings forecasts are

revised downward, the abnormal return associated with target price revisions is larger than

when earnings forecasts are either unchanged or revised upward. This is consistent with the

analysis in which we conditioned on the type of recommendation changes in columns 2 ^ 4.

Second, we examine whether the inclusion of prior stock returns a¡ects the results reported

above, since high (low) prior returns might proxy for unusual events in the recent past that

might prompt analysts to revise their beliefs regarding ¢rm value. We ¢nd strong evidence

that target price revisions are correlated with prior returns. We also ¢nd, however, that in a

regression such as that reported in column 1 of Panel A in Table III, the inclusion of prior one-,

two-, three-, or six-month market-adjusted abnormal returns does not alter any of our conclu-

sions. Third, we examine the e¡ects of the coincidence of target price issuance with earnings

announcements. We repeat the abnormal return regressions (conducted in Section II) sepa-

rately for those events in which earnings announcements occurred within the previous ¢ve

days and those events in which no such recent announcement occurred. We found that the

slope coe„cient on the target price revision is signi¢cant and is similar in magnitude across

the two scenarios.

1950 The Journal of Finance

plete, we conclude this section with an exploration of postevent abnormal re-

turns in which we ask whether investor reaction to target price revisions is un-

biased. Accordingly, we extend the postevent window from event-day þ 3

through six months after the event and examine the abnormal returns in this

period.

We begin by calculating equal-weight size and book-to-market adjusted cumu-

lative abnormal returns (CAR) for each event in our sample. Speci¢cally, we ¢rst

obtain the market capitalization and book-to-market ratio for each ¢rm prior to

an event. Then, using the Fama and French 25 size and book-to-market sorted

portfolios, we ¢nd the portfolio with the matching characteristics. Finally, we

calculate the six-month CAR as the cumulative return on the event ¢rm begin-

ning in the ¢rst month after the event, minus the cumulative matching portfolio

return over the same time period. To avoid possible cross-correlation problems

arising from identical return observations, all but one of the identical return ob-

servations within each portfolio are deleted.

Consider ¢rst the average abnormal returns for subsamples of events classi¢ed

by recommendation upgrades, reiterations, and downgrades.Within upgrade and

downgrade recommendation groups we present abnormal return estimates for

the highest and lowest tercile portfolios, sorted by the magnitude of the analyst™s

target price revision at the time of the event. For the reiterated recommendation

category we report abnormal return estimates for the highest and lowest decile

portfolios, since the number of observations in this case is more than an order of

magnitude larger than in the other two categories. In this manner it is possible to

observe whether target price revisions contain information for future abnormal

returns above and beyond that provided in the associated recommendation. We

calculate standard errors for our CAR estimates using the sample standard de-

viation of the abnormal returns. For example, inferences regarding the six-month

CAR are based on the cross-sectional standard deviation of the event ¢rms™ six-

month cumulative abnormal returns.

Table IV presents our results. Consider ¢rst the sample of target price revi-

sions that were issued along with recommendation upgrades. From the row la-

beled ˜˜ ll target price revisions™™ we learn that, on average, target prices are

A

revised upwards by 10 percent relative to the pre-event stock price. The average

abnormal return is 1.03 percent for the ¢rst month after the event (t-

statistic ¼ 4.1), and increases to 3.08 percent (t-statistic ¼ 4.7) six months after

the event. The next two rows correspond to the abnormal return estimates for

the two subsamples that are sorted based on the magnitude of the price-scaled

target price revision. For events in the highest target price revision group (in

which revisions averaged 37 percent), the average abnormal return through event

month þ 1 is 1.97 percent (t-statistic ¼ 4.2), and it increases to 5.21 percent (t-

statistic ¼ 4.2) through event month þ 6. When we examine events whose target

price revision is in the lowest tercile (in which revisions average À 20 percent),

we ¢nd that abnormal returns are in general negative and insigni¢cant and that

by event month þ 6, equal À 0.38 percent (t-statistic ¼ À 0.4).

Next, we examine events associated with recommendation reiterations. While

there is little trace of an economically meaningful drift for all reiteration events,

An Empirical Analysis of Analysts™ Target Prices 1951

Table IV

Postevent Cumulative Abnormal Returns

Size and book-to-market adjusted cumulative abnormal returns (CARs) are calculated for each

event in our sample as follows. First, we obtain the market capitalization and book-to-market

ratio for each ¢rm prior to the event. Then, each ¢rm is matched with a benchmark portfolio

return from the Fama and French 25 size and book-to-market sorted portfolios.Third, a t-month

CAR (t ¼ 1,y,6) is calculated by cumulating the event ¢rm return beginning in the ¢rst month

after the event through event-month þ t minus the cumulative matching portfolio return over

the same time period.We present CARs for subsamples of events classi¢ed by recommendation

upgrades, downgrades, and reiterations.Within each recommendation upgrade and downgrade

groups, we present CARs for tercile portfolios sorted based on the magnitude of the analyst™s

target price revision at the time of the event. For recommendation reiterations, we present

CARs for decile portfolios sorted based on the magnitude of the analyst™s target price revision

at the time of the event. We winsorize monthly return observations at the 2nd and 98th percen-

tiles to mitigate the possible e¡ect of extreme observations.To avoid a possible cross-correlation

problem caused by identical return observations, we delete all but one of identical return obser-

vations within each portfolio. Standard errors for the CAR estimates are obtained using the

sample standard deviation of the abnormal returns. For example, inferences regarding the six-

month CAR are based on the cross-sectional standard deviation of the events-¢rms™ six-month

CARs. The resulting t-statistics are presented below the CAR estimates. Within each possible

recommendation/target price classi¢cation we also report the average of the target price revi-

sions (scaled by preannouncement stock price). For example, for all events in which recommen-

dations were upgraded, the average target price revision was 10 percent.

Postevent Month

Average

Target Price

þ1 þ2 þ3 þ4 þ5 þ6

Revision

Recommendation upgrades

All target price revisions 10% 1.03 1.45 1.73 2.66 2.82 3.08

4.1 4.0 3.9 5.1 4.8 4.7

Most favorable target price revisions 37% 1.97 2.83 3.21 4.23 4.98 5.21

4.2 4.1 3.7 4.2 4.4 4.2

À 20% 0.08 À 0.23 À 0.20 À 0.13 À 0.38

Least favorable target price revisions 0.08

0.1 À 0.3 À 0.2 À 0.1 À 0.4

0.2

Recommendation reiterations

À 1%

All target price revisions 0.31 0.73 0.96 1.09 0.90 1.08

3.7 6.1 6.3 6.1 4.5 5.0

Most favorable target price revisions 32% 1.19 3.49 4.22 4.77 5.24 6.22

5.8 10.9 10.4 10.1 10.0 11.0

À 58% À 0.68 À 1.07 À 1.10 À 1.16 À 1.74 À 1.88

Least favorable target price revisions

À 4.3 À 4.5 À 3.9 À 3.5 À 4.7 À 4.6

Recommendation downgrades

À 31% À 0.80 À 0.50 À 0.41 À 0.17 À 0.53 À 0.36

All target price revisions

À 3.0 À 1.3 À 0.8 À 0.3 À 0.9 À 0.5

À 1.14 À 0.90 À 0.66 À 0.86 À 0.11

Most favorable target price revisions 13% 0.13

À 2.7 À 1.5 À 0.9 À 1.0 À 0.1 0.1

À 91% À 0.63 À 0.40 À 0.08 À 0.23 À 1.43 À 1.19

Least favorable target price revisions

À 1.2 À 0.5 À 0.1 À 0.2 À 1.2 À 0.9

1952 The Journal of Finance

we ¢nd large and signi¢cant postevent drifts when the information content in

the target price is used. Indeed, the CAR of recommendation reiterations asso-

ciated with target price revisions in the highest (lowest) tercile is 6.22 ( À 1.88)

percent by event month þ 6. Finally, when we examine events associated with

recommendation downgrades, we ¢nd little evidence of drift.

The evidence reported in Table IV that prices drift in the direction of the target

price revision for both recommendation upgrades and reiterations suggests that

target prices contain information regarding future abnormal returns. Since

these ¢ndings are subject to some methodological concerns (see Fama (1998)

and Barber and Lyon (1997)), we now turn to a calendar-time portfolio regression

approach, which has been advocated by Fama (1998) and applied by Ja¡e (1974),

Mandelker (1974), and Brav and Gompers (1997). This approach is conducted by

forming a portfolio that includes all events that are announced within the pre-

vious t periods (in this paper we set t equal to six months). In our setting, we form

the month t portfolio return by either equal weighting or value weighting ¢rm

returns for events that occur within the previous six months. Once a ¢rm has been

added to the portfolio, if analysts issue additional reports on the ¢rm, we refrain

from adding it again to the portfolio.The equal-weight portfolio returns in excess

of the risk-free rate are then benchmarked relative to the maintained asset pri-

cing model, and evidence for abnormal performance is based on the magnitude

and signi¢cance of the regression intercept. It is well known that the portfolio

approach eliminates the problem of cross-sectional dependence among the sam-

ple events and is not susceptible to misleading rejections owing to compounding

of single-period returns (Mitchell and Sta¡ord (2000)).17

We address the choice of a benchmark model by relying on Carhart™s (1997) four-

factor model, which is an extension of the three-factor model of Fama and French

(1993).18 Thus, the regression framework is given by

rp;t À rf;t ¼ a þ b1 Á RMRFt þ b2 Á SMBt þ b3 Á HMLt þ b4 Á PR12t þ et °2Þ

and we focus our inferences on the magnitude and statistical signi¢cance of the

intercept, a.

TableV presents the regression results for portfolios in which monthly returns

are weighted equally. Consider ¢rst the regression results in which portfolios are

formed alternatively based on the three recommendation revision categories

without conditioning on target price revisions (denoted ˜˜ ll™™). In contrast to the

A

17

Mitchell and Sta¡ord (2000) point out that the portfolio approach has several potential

problems that arise from the changing composition of the portfolio through time, which can

potentially lead to heteroskedasticity. We have veri¢ed that heteroskedasticity alters none of

the conclusions drawn below.

18

The ¢rst factor, RMRF, is the excess return on the value-weighted market portfolio. The

second factor, SMB, is the return on a zero-investment portfolio formed by subtracting the

return on a large ¢rm portfolio from the return on a small ¢rm portfolio. The third factor is

the return of another mimicking portfolio, HML, de¢ned as the return on a portfolio of high

book-to-market stocks less the return on a portfolio of low book-to-market stocks. The fourth

factor, PR12, is formed by taking the return on high return stocks minus the return on low

return stocks over the preceding year.

An Empirical Analysis of Analysts™ Target Prices 1953

TableV

Calendar Time Regressions

The sample is all target price announcements between January 1997 and December 1999. Port-

folios are formed by including all events that were announced within the previous six months.

The portfolios™ equally weighted monthly returns, in excess of the risk-free rate, are regressed

on the following four factors: RMRF, the excess return on the value-weighted market portfolio;

SMB, the return on a zero investment portfolio formed by subtracting the return on a large ¢rm

portfolio from the return on a small ¢rm portfolio; HML, the return on a portfolio of high book-

to-market stocks less the return on a portfolio of low book-to-market stocks; and PR12, formed

by taking the return on high momentum stocks minus the return on low momentum stocks.We

report regression results for portfolios classi¢ed by recommendation upgrades, reiterations,

and downgrades. We form tercile (decile) portfolios for recommendation upgrades/downgrades

(reiterations) based on the magnitude of the analyst target price revision, which we then re-

gress on the four factors. For example, conditional on a recommendation upgrade, we construct

a portfolio that includes ¢rms whose target price revision occurred within the pervious six

months and was in the top or bottom tercile at the time it was announced.We winsorize monthly

return observations at the 2nd and 98th percentiles to mitigate the possible e¡ect of extreme

observations. Adjusted R2 and t-statistics are presented for each regression.

PR12 Adjusted R2

Intercept RMRF SMB HML

Recommendation upgrades

All target price revisions 0.373 1.150 0.537 0.171 0.005 95.5%

1.69 23.37 8.86 2.56 0.09

0.748 À 0.169 À 0.022

Most favorable target price revisions 0.799 1.170 95.1%

8.83 À 1.81 À 0.32

2.59 17.01

Least favorable target price revisions À 0.112 0.344 À 0.061