Download Analysis of Post-Period Predictive Ability of Price Recommendations by Analysts and more Study notes Finance in PDF only on Docsity!
Can analysts pick stocks for the long-run?
Forthcoming, Journal of Financial Economics
Oya Altınkılıç a, Robert S. Hansenb,*^ , Liyu Ye b
ba School of Business, George Washington University, Washington, D.C. 20052, USA A.B. Freeman School of Business, Tulane University, New Orleans, LA 70118, USA c U.S. Treasury, Fannie Mae, Washington D.C., 20016, USA
Abstract
This paper examines post-revision return drift, or PRD, following analysts’ revisions of their stock recommendations. PRD refers to the finding that the analysts’ recommendation changes predict future long- term returns in the same direction as the change (i.e., upgrades are followed by positive returns, anddowngrades are followed by negative returns). During the high-frequency algorithmic trading period of 2003-2010, average PRD is no longer significantly different from zero. The new findings agree with improved market efficiency after declines in real trading cost inefficiencies. They are consistent with areduced information production role for analysts in the supercomputer era.
JEL classification: G02, G14, G24. Keywords: Analysts’ forecasts, Financial analysts, Financial markets, Investment banking, Market efficiency, Security analysts, Behavioral finance. First draft: January, 2010 Latest draft: February, 2015 We thank the editor, Bill Schwert. We thank an anonymous referee for very constructive comments, and for helpfulcomments received on early drafts of the paper we thank colleagues at The Freeman School of Business, Tulane University and The School of Business at George Washington University.
- (^) Corresponding author contact information: Tulane University, New Orleans, LA 70118. Tel +504-865-5624. E-mail addresses (L. Ye). : oaltin@gwu.edu (O. Altınkılıç), rob.hansen@tulane.edu (R.S. Hansen), yeliyuyeah@gmail.com
1. Introduction
For decades researchers have examined average long-run stock returns after sell-side security analysts revise their recommendations for buying and selling stocks. The universal finding is that the recommendation changes predict future long-term returns in the same direction as the change (i.e., upgrades are followed by positive returns, and downgrades are followed by negative returns). This phenomenon is known as post-revision return drift (PRD). This result has supported the hypothesis that PRD persists because investors typically underreact to analysts, responding partly at their revision announcements and slowly thereafter, perhaps taking months. It has also underpinned the nested hypothesis that security analysts are better-informed, skilful at information discovery from non-public sources (e.g., from insiders) and from neglected public information in inefficient markets, as noted by Grossman and Stiglitz (1980).^1 This article provides new evidence about PRD that extends the literature in a number of ways. The primary contribution is the finding that average PRD is no longer persistently different from zero in the May 2003 through 2010 sample post-period. A second contribution is new results that show a causal relationship between analysts’ revisions and PRD is not supported in many tests of the PRD cross-section. A third contribution is new evidence from the PRD cross-section regarding the investor underreaction hypothesis and the informed analyst hypothesis. Results from tests for underreaction that use proxies suggested by other researchers do not support the underreaction hypothesis in the post-period. For instance, one finding in this article shows there is no significant association between PRD and analysts’ coverage, a widely used proxy for underreaction. Tests of the informed analyst hypothesis that employ proxies for better-informed analysts used in prior research, do not support the idea that analysts typically supply new information that correctly picks stocks for the long run. One example is that the PRD cross-
(^1) Givoly and Lakonishok (1979), Womack (1996), Hong, Lim, and Stein (2000), Gleason and Lee (2003), Jegadeesh, Kim, Krische, and Lee (2004), and Loh (2010) discuss underreaction to analysts.
analysts’ revisions in the other Group of 7 countries: Canada, France, Germany, Italy, Japan, and the UK.^3 The findings show that drift after analysts’ revisions in these countries also is not informative in the post- period, supporting similar findings for U.S. analysts. PRD is examined from several perspectives, reflecting different ways that researchers have measured PRD, and a variety of samples that are employed for different tests. One PRD measure uses an event study approach in which the revisions are aligned on their announcement date, similar to that used by Womack (1996) and by Jegadeesh, Kim, Krische, and Lee (2004). This measure is examined in the Event- Time Sample (see Appendix A.1 for sample descriptions). A second measure evaluates PRD from a portfolio perspective in calendar time and examines the returns on buy portfolios of upgraded stocks and sell portfolios of downgraded stocks, and compares their differences. This drift measure is similar to that employed in Barber, McNichols, and Trueman (2001) and utilizes the Portfolio Sample. PRD is examined from a third viewpoint first introduced in this article, which aligns firms on their earnings report announcement dates, and compares the drift for firms with upgrades to the drift for the other firms with continuations (i.e., those with unchanged recommendations), and similarly for the downgrades. This method controls for the influence of post-earnings announcement drift (PEAD) and uses the Earnings Sample. Revisions in each of these three Samples are examined in both the post-period and in the pre- period. This provides opportunities to replicate findings from the earlier studies, and to compare the pre- and post-period PRD behavior side-by-side. PRD is also examined in a sample of consensus recommendations within each period.^4 Although average transaction costs are lower in the post-period, it is unlikely that they have entirely disappeared (for example, see Beneish, Lee, and Nichols 2013; Boehmer, and Wu, 2013). Under the transaction cost rationale, some PRD is likely present for stocks with relatively high transaction costs. In
(^34) We thank the referee for suggesting this out of sample test.
(2004) also study PRD. Cowles (1933, 1944) does not find evidence of PRD in a much earlier sample period.Dimson and Marsh (1984), Elton, Gruber, and Grossman (1986), Stickel (1992), and Mikhail, Walther, and Willis
agreement, after sorting the Event-Time Sample into trading volume deciles, some statistically significant average PRD exists in the lowest decile, or 10% of the revisions. Significant average PRD is also present in the lowest deciles in sorts by firm size and by analysts’ coverage of the firm. The VSC revisions which are common to the lowest deciles for all three characteristics, and make up 3% of all revisions, are expected to have high transaction costs. In agreement VSC revision stocks have a number of characteristics that are consistent with high transaction costs. Their stock prices are among the lowest, so trading a certain weight of these shares in a given portfolio will be more costly (i.e., requiring the sale of many more shares). They are twice as likely to be listed on the Nasdaq, where bid-ask spreads are larger than for NYSE-listed firms in the post-period (Angel, Harris, and Spatt 2012). Furthermore, VSC revision stocks are among the smallest firms, with other firms’ average equity valued 50 times higher. Also, their trading volume is the lowest, showing they are in limited short-run demand and supply. The findings also allow that some PRD could be the unintended result of the way in which the long- run abnormal returns are measured. For example, the small sample evidence of PRD is ever-present in the lowest deciles in a number of sorts for the Event-Time Sample , while the evidence is inconsistent in the lowest deciles of the Portfolio and Earnings Samples. While PRD may come from analysts’ new information discovery and asset pricing model effects, this article documents a third potential source: drift that is associated with other recent news and events about the covered firm. A number of studies find that analysts often piggyback their reports on recent events and news, which contribute to magnifying return reactions measured during the days around the revision announcements, and to masking how much of the return reaction can be credited to analysts’ new information (Altınkılıç and Hansen, 2009; Altınkılıç, Balashov, and Hansen 2013; Kim and Song, 2014). Similarly, recent events and news themselves can be associated with their own return drift which could confound the average PRD and consequently raise the question of how much of the PRD can be credited to analysts’ revisions. This issue is addressed in this article through examination of concurrent event drift from the notable reported earnings event using the Earnings Sample , which allows for control of the well-
analysts’ revisions, as well as the predictability of long-run returns based on the revisions. The disappearance of PRD therefore exemplifies how rational pricing of securities interacts with changes in real inefficiencies to extend market efficiency. One alternative explanation for the lack of significant PRD in the post-period, which conserves the view that analysts supply ample new information, is that investors no longer underreact to new information from analysts’ revision announcements. Perhaps the profound economic forces of the supercomputer era, which greatly accelerated trading frequency at minimal cost, also elevated investor attention to all stocks, thereby significantly expanding investor awareness and hastening their reaction to news that affects stocks’ prices. An important implication of this increased awareness theory, if the revisions supply analysts’ new information, is that announcements of analysts’ revisions should be met quickly with widespread investor reaction that quickly impounds most of the new information, if not all, into stock prices. Thus, in the post-period, the average return reaction to revision announcements should contain economically significant evidence of analysts’ new information, all else the same. However, this implication is not supported by recent studies, which report that that the announcement period price reaction to analysts’ reports contains little new information that can be attributed to the analysts themselves (Altınkılıç and Hansen, 2009; Loh and Stulz, 2011; Altınkılıç, Balashov, and Hansen 2013). The findings in this article widen the empirical evidence indicating little average investor reaction to analysts’ reports. The remainder of the article proceeds as follows; Section 2 describes the sample; Section 3 examines PRD in side-by-side comparisons of the post-period and the pre-period; Section 4 examines evidence of PRD in the 3% sample of the Event-Time Sample revisions; Section 5 examines the underreaction hypothesis and the informed analyst hypothesis; Section 6 examines PRD in a sample of international stocks; the article concludes with final thoughts and implications for future research in Section 7.
2. Sample description
The samples used in this article draw from real time or no source code needed recommendation revisions from First Call Historical Database (FCHD) from 1997 through 2010 (batch recommendations, initiations, and resumptions are excluded), which have the control variables used in this study as identified from the literature. Recommendation levels range from 1 to 5 respectively: strong buy, buy, hold, sell, and strong sell. If the new recommendation level is lower (higher) than the previous recommendation level, it is an upgrade (downgrade). Recommendations issued after 4 PM are taken as issued the next trading day. If one brokerage issues multiple recommendations on the same day for the same company, only the latest is retained. If there are multiple upgrades (downgrades) for the same company on the same day, the upgrade (downgrade) is counted only once. Conflicting revisions for the same stock on the same day and on stocks with no earnings announcement in the prior year are deleted. CRSP provides daily stock prices and volume. Industry classifications are from Ken French’s website.
[Insert Table 1 about here] Revision annual frequency is usually higher over 2005-2007 for both upgrades and downgrades. Covered firms and brokerage firms in the sample follow a similar pattern (Table 1).
3. Post-revision return drift
The first measure of PRD is in revision event-time, using the long-term average abnormal buy and hold returns aligned on the third trading day after the revision announcement day. Measuring drift starting three days after the announcement avoids return shocks from confounding events immediately around the revision announcement. Initially, PRD is calculated two ways that use different asset pricing models: the Market Model, denoted MM , and the characteristics model of Daniel, Grinblatt, Titman, and Wermers (1997), denoted DGTW. The sample for these estimations is the Event-Time Sample.
[Insert Figure 2 about here] A further evaluation of the return performance uses a measure that is conditioned on the direction of analysts’ revisions; called the “up-less-down” strategy, which seeks to capture the difference between PRD after upgrades and PRD after downgrades. In effect, this is a strategy that invests long in the upgraded stocks and short in the downgraded stocks, providing a gross assessment of the return from analysts’ information. Note the up-less-down measure is not capable of evaluating expected profits; that is the gross value from the entire investment strategy net of the likely transaction costs from executing the strategy. When the gross value from this up-less-down strategy is significantly positive, the average PRD only reveals that the recommendations have behaved, in combination, as if they are informative for the investment duration. For the post-period, the up-less-down MM PRD generally is slightly negative over the 120-day after the revisions are announced, and so is the DGTW PRD (Figure 2, Panel A). This does not agree with the informed analyst view. In contrast, in the pre-period the strategy yields positive PRD for both drift measures, which does not agree with the informed analyst view. While this positive drift resembles qualitatively similar findings from prior studies in earlier sample periods, the only source for this positive up-less-down drift in Figure 2 is the long upgrade leg of the portfolio, because the short downgrade leg earns negative returns (Figure 1). In other words, a strategy of going long stocks after both upgrades and downgrades, contrary to the strategy that strictly follows analysts’ advice, would have earned a higher return than the up-less-down strategy. The PRDs computed using value-weighted returns are not materially different from the equally weighted case reported in Figures 1 and 2 (unreported).
3.2. The decline of PRD
This section documents in more detail the behavior of PRD in the post-period and in the pre-period.
3.2.1. The Event-Time Sample
[Insert Table 2 about here] Consider first the univariate regression of PRD on the revision upgrade indicator variable, UP , that is one if the recommendation is an upgrade and zero if it is a downgrade. The regression equation is thus ܴܲܦ (^) ,ௗగ^ ܷܲߚ ߙ ൌ (^) ߝ ,ௗ (2)
where ߚ registers the additional drift that is associated with analysts’ upgrades relative to their downgrades. In the post-period the regression estimate ߚመ^ shows the MM measure of PRD is positive over the first 20 days after day 3, although the coefficient estimates are not economically large at around 27-28 bp (with significance at the 5% level). The two longer duration drifts are not significant, diminishing in magnitude or turning negative (Table 2, Panel A). Qualitatively similar results are recorded for the DGTW PRD (Table 2, Panel A). In contrast, the pre-period upgrade average PRD is significantly positive relative to the downgrade PRD, at all durations, typically at the 1% level, for both the MM and the DGTW measures. Moreover, the magnitude of the UP coefficient estimates for the pre-period are often large, over five times larger than the estimates for the post-period. The findings from the pre-period agree with the informed analyst hypothesis and are qualitatively similar to findings reported in earlier studies.
[Insert Table 2 about here]
The second regression model is a multivariate specification that includes twelve exogenous binary variables that control for several other effects affecting the return behavior, similar to the measure employed by Jegadeesh, Kim, Krische, and Lee (2004)
ܴܲܦ (^) ,ௗ ߙ ൌ (^) ܷܲߚ (^) ∑^ ଵଶܺ, ߛ ߝ ,ௗ (3)
3.2.2. Portfolio Sample
The next measure of PRD uses a calendar portfolio approach, similar to Barber, Lehavy, and Trueman (2007), and focuses on three portfolio classes: the Buy portfolio of all upgraded stocks, the Sell portfolio of all downgraded stocks, and the Buy-less-Sell portfolio, which is the difference between the upgraded and downgraded stocks. This sample is the Portfolio Sample. Portfolio drift is measured using the Carhart (1997) extension of the Fama and French (1992) expected returns model,
ܴ ௧ ܴെ (^) ௧ ߙ ൌ ߚ ܴ൫ (^) ௧௩^ ܴെ (^) ௧ ൯ ݏܵ ܤܯ (^) ௧ ݄ (^) ܮܯܪ௧ ݓ ܹ ܮܯ (^) ௧ ߝ ௧ (4)
where for date t the dependent variable is the return on a portfolio j of recommendations less the risk-free rate, and the right-hand side variables are respectively: the return on the value-weighted market index less the risk-free rate; the return on a portfolio of small-cap stocks less the return on a portfolio of large-cap stocks; the return on a portfolio of high book-to-market value stocks less the return on a portfolio of low book-to-market value stocks; and the return on a portfolio of recent high return stocks (winners) less the return on a portfolio of low recent return stocks (losers). The factors and the risk-free interest rate are collected from Ken French’s website. The expected signs of the coefficients in equation (4) are: small cap and small value portfolios respectively have higher expected returns than the large cap and growth portfolios. The holding period returns are measured over the 60 trading days starting on day three after the revision announcement. Drift performance is measured by the intercept estimate, ߙො, in equation (4). The estimation results are reported in Table 3.
[Insert Table 3 about here] In the post-period, estimates for both the Buy portfolio and the Sell portfolio indicate no significant causation between the revisions and PRD. Similarly, Buy-less-Sell portfolio PRD is not significant (Table 3, Panel A). Neither result is expected by the informed analyst view
In the pre-period the Buy portfolio drift estimate is positive and statistically significant, in agreement with the informed analyst hypothesis and some prior studies (Table 3, Panel B). However, the Sell portfolio estimate indicates no significant post-revision return drift, which is inconsistent with the view that downgrades anticipate lower prices in the future. The Buy-less-Sell portfolio PRD remains significant, however, registering the strong relative performance of the Buy portfolio. Results from further tests in the post-period, after removing revisions issued during the financial crisis period (September 2008 through the sample period end), are weaker than the full post-period results (Table 3, Panel C), as none of the portfolios’ PRD is significant.
3.2.3. The Earnings Sample
Thus far, the tests follow the custom of comparing PRD after all upgrades with PRD after all downgrades. However, these tests overlook the possible external influence in those revisions when analysts’ decisions to upgrade or downgrade their revisions rely at least in part on publicly known factors that in and of themselves have some power to predict future drift. To the extent that analysts piggyback their revisions on drift predictors in this way, there may be a revision selection bias that could be influential enough to cause spurious agreement between the revisions and subsequent PRD, even when the analyst is not communicating new information. An important example taken up here is piggybacking the revisions on known predictors of PEAD, an enduring long-run drift anomaly (e.g., see Fama, 1998). For example, by upgrading (downgrading) after good (bad) earnings surprises, the revision could associate with subsequent positive (negative) PEAD, independent of any new information the analysts might deliver in the revision announcement. While some control for this bias is possible in the Event- Time Sample tests by including unexpected earnings (SUE) among the right-hand side factors in the PRD regression estimations, spurious correlation could still be operative that biases the revision PRD to agree with the informed analyst view, even though analysts may only be tracking PEAD alone, not seeking to offer new information.
[Insert Table 4 about here] In the post-period the average upgrade increment to PRD is not statistically significantly positive, for any of the three durations of the MM PRD measure (Table 4, Panel A). Similarly, for the downgrades the average increment to PRD is not statistically significantly negative, for any duration, for the MM PRD and DGTW PRD measures (Table 4, Panel A). These findings are not consistent with the informed analyst hypothesis. In light of these findings, it is plausible that evidence of significant PRD reported in the post- period Event-Time Sample could contain bias to the extent that revisions piggyback on predictors of future drift that are not fully reflected in the twelve factors. In the pre-period, by contrast, the PRD for the upgrades is significantly positive, for both MM PRD and the DGTW PRD, for all three durations (Table 4, Panel B). For the downgrades, half of the DOWN coefficient estimates are significantly negative. Some of these findings agree with the informed analyst thesis, and are qualitatively similar to findings that are reported in prior studies whose sample period overlaps the pre-period. For the post-period, further tests are conducted after removing observations during the financial crisis period (September 2008 through the sample period end). These estimates are never significant, qualitatively similar to the full post-period results (Table 4, Panel C). Hereafter, because the reported post-period findings are not altered qualitatively by excluding crisis period revisions, to conserve space only results for the whole post-period sample are reported.
3.2.4. The Matched-Earnings Sample
One concern is that by including all earnings announcements, the Earnings Sample tests might not adequately control for the extent to which firms associated with analysts’ revisions differ from other firms associated with continuations, in ways that are not controlled by the regression model despite the inclusion the twelve factors. For example, perhaps PRD is not linear in some of the factors, or perhaps there are measurement concerns, like the fact that the inclusion of a momentum measure among the
independent variables might not adequately control spurious correlation that could be attributed to earnings announcement news. Further tests, therefore, use the Matched-Earnings Sample. In the Matched-Earnings Sample , each upgrade and each downgrade from the Earnings Sample is matched to a similar continuation in the Earnings Sample , where similarity is determined using the Propensity Score method. Thus, a Probit model is first estimated for the Earnings Sample using the set of twelve predictor variables to obtain propensity scores for all revisions and continuations. The Matched Earnings Sample consequently consists of all upgrades and downgrades and their respective matched continuations using the closest Propensity Score. In the Matched-Earnings Sample the measured drift for the revisions includes PEAD in addition to the incremental PRD triggered by the informativeness of the analysts’ revisions. The continuations, however, contain only PEAD and no incremental information effects from revisions (as in equation 5).^7
[Insert Table 5 about here] For the Matched-Earnings Sample , Panel A of Table 5 reports mean values for the twelve factors used to estimate the Propensity Scores, for the upgrades and their corresponding matched continuations, and likewise for the downgrades. The factor means show the Propensity Score identifies matched continuation firms that are similar to the firms, in terms of the twelve factors.
(^7) The Propensity Score matching test is common in financial settings, including underreaction events, where selection bias is a concern. Lee and Wahal (2004) look at IPO underpricing and venture capital reputation; Drucker and Puri (2005) examine SEOs by the firm’s lending bank; Lee and Masulis (2010) look at underwriter and venture capitalist reputations and the firm’s earnings management. For further discussion of the tests, see Heckman (1979), Heckman and Robb (1986), Heckman, Ichimura, and Todd (1997, 1998), and Roberts and Whited (2012).
in the Matched Earnings Sample there is no significant difference between average PRD for upgrades and their matches, or downgrades and their matches, across all three drift durations. In the pre-period Event-Time Sample , the cross-section regression estimates show upgrades have statistically significant impact on PRD. This is confirmed with the Buy-less-Sell findings in the Portfolio Sample , as well as the UP findings and 3-month duration DOWN findings in the Earnings Sample. There is also support for this conclusion among the mixed results using the Matched Earnings Sample , Propensity Score tests. A number of these pre-period results also agree with findings reported in earlier studies.
4. Has post-revision return drift totally disappeared?
If PRD persists because of high transaction costs, then economically significant PRD could still exist in post-period sub-samples that are more likely to have relatively high transaction costs, even if the average PRD is not significant. This section reports results from several tests that are designed to determine if average PRD within sub-sample deciles is significant when the revision firm characteristics suggest large transaction costs are likely to be present. The tests focus on three characteristics: Volume , Firm Size , and Coverage.
4.1. Volume
Everything else the same, low volume firms are likely to have higher costs of illiquidity than other firms, raising their trading cost (Amihud, 2002). To the extent that low volume indicates limited supply of stock at current prices, transaction costs are likely to be even higher when the trading strategies include shorting downgraded stocks. Trading Volume in the firm’s shares is measured by the average daily trading volume over the prior calendar year. For the post-period, in the Event-Time Sample , PRD is insignificant in most trading Volume deciles, however it is distinctly present in the lowest decile using either drift measure, across all
three durations (Table 6, Panel A). This agrees with the transaction cost thesis, to the extent that the lowest Volume decile stocks typically have relatively larger costs. In contrast, for the Portfolio Sample , significant PRD is not consistently present in the lowest decile (Table 6, Panel B); negative PRD is present in the lowest Sell portfolio decile, but significant PRD does not occur in the lowest deciles for the Buy, or the Buy-less-Sell portfolios (Table 6, Panel B). Also, for the Earnings Sample , there is little evidence of significant PRD across durations for the upgrades (Table 6, Panel C), while for the downgrades (Table 6, Panel D), there is evidence in the lowest decile only for the 20-day duration.
[Insert Table 6 about here]
By contrast, in the pre-period Event-Time Sample , significant drift is broadly apparent up and down the Volume deciles. For example, in seven of the 20-day deciles, for both MM and DGTW , PRD is noticeably strong statistically in the lowest deciles across all three durations and both drift measures (Table 6, Panel A). Under the transaction costs thesis, this pattern of significant average PRD over most deciles would suggest relatively high transaction costs range across far more revisions in the pre-period. In the Portfolio Sample deciles, PRD is significant in eight Buy deciles, two Sell deciles, and eight Buy- less-Sell deciles (Table 6, Panel B). In the Earnings Sample , PRD is present in three deciles for most durations, while no obvious cluster pattern is present. For upgrades, PRD is evident in the lowest decile across the two estimation methods (Table 6, Panel C). The evidence is mixed in higher Volume deciles; in some cases it is statistically significant, but in one of those cases there is a wrong sign for the underreaction view (Table 6, Panel D).
4.2. Firm size
Firm size is also a relevant characteristic of trading costs, which are likely to be high for smaller firms. For example, Hendershott, Jones, and Menkveld (2011) find that while quoted and effective bid-
