extreme changes. It is therefore not surprising that our estimates of amarket,the market price response coeicient, are extremely small.

We conclude this section by presenting, in Table VI, Panel C, parameter estimates sorted by the irms market capitalization (size). To the extent that size differences are associated with cross-sectional differences in information asymmetry or risk, our analysis sheds light on the dynamics of target prices as they relate to this characteristic of firms. We form size terciles based on NYSE capitalization cutoffs and adjust these quarterly. Event firms are then classified based on the market value of their equity at the end of the preceding quarter. Consider irst the estimates of the long-run ratio of target-to-market prices, b. Beginning with irms in the smallest tercile, the average estimate of b is 1.37 and declines monotonically to 1.23 for the firms in the largest tercile. This pattern indicates that analysts expect a higher annual price appreciation for small stocks, which is consistent with asset pricing models, such as the Fama and French three-factor model, that include size as a risk factor. Finally, the information regarding the response coefficients atp and amarket indicates that there are no cross-sectional differences in either analyst or the market reactions to deviations from the long-term relationship across the size terciles. As with the IBM results, the overall evidence supports the interpretation that market prices react to the information conveyed in analyst reports but that any correction to the long-term relation between target and market prices is predominantly made by analysts.

IV. Linking the Evidence from the Short- and Long-term Analyses

The preceding section provides evidence as to the dynamics of target and market prices as well as to their common long-term relation. We now seek to build on these results and link them to the short-term event study conducted in Section II. Specifically, we construct an estimate of the expected one-week-ahead consensus target price and then examine whether investors understand the long-term dynamics of the price series. We do this by testing whether event-day abnormal returns are correlated with the unexpected part of the target price revision (for a similar approach, see Lowry and Schwert (2002)). Toward this end, we irst estimate for each of the 900 firms a one-week-ahead forecast of the consensus target prices, using the sample information that would have been available to investors prior to the release of the analyst reports. We require a minimum of 10 weekly observations to it the cointegration regression. Using the expected consensus target price, we construct two variables. One is the expected target price revision, which equals the scaled difference between preannouncement consensus target price and the forecasted one. The second is the unexpected target price revision, which equals the difference between the announced and the expected target price. We scale both the expected and unexpected variables by the event-irms market price two days prior to the event.

Next, we estimate a regression similar to the one conducted in Section II but by employing the expected and unexpected target price proxies rather than the scaled change in the individual brokerage house target price, DTP/P. If the coin-

tegration setup provides an adequate description of the evolution of market and target prices, we expect that only the unexpected component of the target price revision would be related to event-time abnormal returns. The regression takes the following form:

AR = 1 UPGRADES + (DOWNGRADES + REITERATIONS

The regression results are reported in Table VII. Consider the results for Model I first. Consistent with our prediction, we find that, controlling for the recommendation and earnings forecast revisions, average abnormal returns are significantly associated with the proxy for unexpected revision in target price (slope coefficient = 2.671 with t-statistic = 27.0). We also find no reliable relationship between abnormal returns and the expected revision in target price. This inding is consistent with the view that investors understand the long-term dynamics documented in Section III and thus are able to anticipate some of the analysts revisions.

The association that we estimate between unexpected target price revisions and event-day abnormal returns implicitly imposes the restriction that market participants react symmetrically to unexpected revisions in target prices, irrespective of the current levels of target and market prices. We now relax this restriction by constructing four alternative measures of unexpected target price revisions that condition on the magnitude of the pre-event ratio of target-to-market price relative to the estimate of the firms long-term relation. Specifically, suppose that at the time of the announcement, the pre-event target price to market price ratio is 1.25 and the current estimate of the long-term ratio is 1.20. If the target price revision was away from the long-term ratio of 1.20, we classify it as Unexpected ATP/Pabove/away.A target price revision toward the long-term ratio is classified as Unexpected ATP/P above/towards The remaining two variables, unexpected target price revisions when the pre-event target price to market price ratio is lower than the estimate oflong-term ratio, are deined similarly.

The column labeled Model II in Table VII provides the regression results. We ind that in cases where the pre-event ratio oftarget-to-market prices is below the long-term relation, unexpected target price revisions away from the long-term relation are associated with larger negative abnormal returns than in any of the other cases (z3 = 6.864 with t-statistic = 20.1). Indeed, the latter abnormal return is nearly twice as high as in the case in which the pre-event ratio oftarget-to-market prices is above the long-term relation and the unexpected target price revision is away from the long-term relation cases (z1 = 2.712 with t-statistic = 17.7). The p-value for the hypothesis that the previous two reactions are equal indicates that we can reject this null. Finally, it can be seen that unlike market responses to unexpected target price revisions away from the long-term relation, target price revisions toward the long-term relation are economically and statistically smaller.

TableVII

Informativeness ofTarget Prices Based on the Cointegration Regression

The sample contains all target price announcements between January 1997 and December 1999. The table reports regression results in which the dependent variable is the market-adjusted buy-and-hold abnormal returns around target price announcements. The independent variables are indicator variables for analysts recommendation revisions, earnings forecast revisions, and expected and unexpected target price revisions. The indicator variables assume the value 1 for the relevant recommendation revision and 0 otherwise. The recommendation categories are upgrades, downgrades, and reiterations. Abnormal returns are computed as the difference between the firm buy-and-hold return and the buy-and-hold return on the NYSE/AMEX/Nasdaq value-weighted index over the period beginning two days prior to and ending two days subsequent to the target price announcement. Earnings forecast revision, denoted DF/P, is computed as the percentage change in the brokerage house current and prior annual earnings forecast scaled by preannouncement stock price. Expected target price revisions for each firm and event are constructed as follows. We first estimate a one-week-ahead forecast of the consensus target prices using the sample information that would have been available to investors prior to the release of the analyst report. We require a minimum of 10 weekly observations to it the cointe-gration regression. The diference between the regression forecast and the preannouncement consensus target price, denoted (Expected DTP/P), serves as a proxy for the expected consensus target price revision. Unexpected target price revision, denoted (Unexpected DTP/P), is the difference between the announced and expected target price. Both the expected and unexpected target price variables are scaled by the event-firms market price two days prior to the event. We winsorize all variables at the 1st and 99th percentiles to mitigate the possible effect of extreme observations. In addition, we verify that regression results are not sensitive to inluential observations. The number of observations in each regression is 43,660.