Does Analyst Optimism Fuel Stock Price Momentum?

Abstract

Researchers have struggled to find rational risk factors that explain momentum profits derived from buying recent winners and shorting recent losers. Behavioral explanations have been offered that focus on the tendencies of investors to underreact to news and recommendations. Our study provides an alternative explanation centered on the behavior of sell-side analysts. We find a change in consensus recommendation from a hold to a buy is accompanied by an increase in momentum profits of 3.40% annually. Momentum profits fall, yet remain material, after the passage of Reg FD and the enactment of the Global Analyst Research Settlement. Our results support a behavioral explanation of investor cognitive biases fueled by analyst regency and optimism biases.

Keywords:

• Stock price momentum
• Sell-side analysts
• Cognitive bias
• Confirmation bias
• Analyst recommendation optimism

JEL CLASSIFICATION:

• G1
• G14
• G24

Acknowledgments

We gratefully acknowledge participants in seminars at University of North Georgia, Colorado State University, University of Texas at Arlington, the Financial Management Association meetings, the Eastern Finance Association Meetings, the Southwest Finance Association meetings, and the Midwest Finance Association meetings for their many helpful comments on earlier versions of our paper.

Notes

1 Although Cornett et al. (2007) find investor reaction to revisions in analyst recommendations decreased after the enactment of Regulation Fair Disclosure (Reg FD), the authors show that the effect remains significant and robust.

2 Related research includes Givoly and Lakonishok (1979), Stickel (1990), Womack (1996), Chan et al. (1996), and Dichev and Piotroski (1999). Barberis and Shleifer (2003) develop a behavioral investor demand function modeled with momentum preferences and offer a rich set of implications for correlations and autocorrelations of returns.

3 We borrow the phrase from Jegadeesh et al. (2004, page 1086), who contend analysts may spur momentum “by virtue of their own stature as opinion makers.”

4 We thank an anonymous referee for suggesting the analysis of the effects of the Global Analyst Research Settlement.

5 Lee et al. (2011) note covered stocks that are unreported might be mistakenly classified as uncovered stocks, especially during earlier sample periods. Our paper, however, is free from this bias as we classify stocks by cross-referencing analyst coverage records with the IBES Earnings Summary History file. Our definition coincides with Jegadeesh et al. (2004).

6 Our recommendation breakpoints match Barber et al. (2003). Our classification scheme improves on methods that classify stocks into equal sized quintiles (based on consensus recommendation scores). The latter method encounters issues when many consensus recommendations are identical and choices must be made of how to break ties at the quintile breakpoints.

7 Mean returns are slightly lower for REC 5 portfolios versus REC 1 portfolios. When we remove the observations when portfolio size is less than 5, we find mean returns are slightly higher for REC 5 versus REC 1 in most cases (similar to Jegadeesh et al. (2004) after controlling for momentum effects). There were 6 occasions when the portfolio size was less than 5.

8 If confirmation bias is strong and if analysts' weakest recommendations (REC 1) offer unambiguously pessimistic signals, one might expect the shape of the momentum profits to follow a swoosh shape across the REC levels. For example, if weak recommendations for prior losers offers strong confirmation for investors, then the average return for stocks at the intersection of prior return quintile 1 (weakest recent performers) and REC 1 will be low. And, returns for stocks at the intersection of prior return quintile 5 (best recent performers) and REC 1 will not be low; e.g., consistent with a confirmation bias, investors ignore pessimistic signals for prior winners. In this scenario, momentum profits will be higher for REC 1 versus REC 2 where analyst signals are less clear (essentially a hold recommendation). This is exactly what we find – the momentum strategy return is higher for REC 1 versus REC 2 for 3 and 6 month holding periods (e.g., 2.48% versus 1.85% and 3.65% versus 3.36%, respectively). The differences are not larger because most of the REC 1 recommendations do not offer strong confirmation for prior loser stocks; e.g., the average REC 1 recommendation is midway between a sell and hold. As an additional test, we split REC 1 in halves where REC 1a consists of the weaker half of REC 1 and REC 1b consists of the remaining half of REC 1. Mean consensus recommendations equal 2.15 and 2.74 in REC 1a and REC 1b, respectively. The subsequent 3-month mean return for the stocks falling at the intersection of prior return quintile 1 and REC 1a equals -0.26% versus 0.24% for stocks falling at the intersection of prior return quintile 1 and REC 1b. Three-month momentum profits are 45 basis points higher for REC 1a versus REC 1b, exactly the trend we would expect when confirmation signals for prior loser stocks are stronger for stocks in REC 1a versus REC 1b.

9 The annualized momentum factor returns from Professor Kenneth French’s website over our sample period equal 5.13%, which equates more closely to the effect of a change from a hold to a strong buy in our tests. The momentum factor returns are derived from 2-pass portfolios based on below 30th and above 70th percentile prior return (months -12 to -2) and above and below median NYSE firm size, and are rebalanced monthly. Despite the differences in methodology, the momentum factor return provides a helpful reference point. The momentum factor returns are available from: https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html.

10 As pointed out by Daniel and Moskowitz (2016), momentum strategies experience occasional crashes. In particular, momentum strategies failed notably during parts of the 2008-2009 financial crisis, as also illustrated in Figure 1.

11 We ran additional tests using a composite measure using Zhang’s (2006) information uncertainty variables. Conclusions remain unchanged. As pointed out by Zhang, firm size serves as a good proxy for information uncertainty.

12 For example, from Table 1, the mean recommendations in REC 1, REC 2, REC 3, and REC 4 are 2.48, 3.17, 3.67, and 4.10, respectively. Using linear interpolation, a score of 3 equates to REC 1.75 and a score of 4 equates to REC 3.76. Therefore, a change from a hold to a buy is equivalent to a recommendation level change of approximately 2.

13 We ran a robustness test based on analyst earnings forecast dispersion. Altinkiliç and Hansen (2009) question the contributions of analysts, contending analyst recommendations revisions often “piggy back” corporate events. For example, positive (negative) recommendations follow positive (negative) corporate announcements such as beating (missing) earnings expectations. Therefore, analyst recommendation revisions have little informational value, but simply reflect positive (or negative) corporate news. Intuitively, the herding of analyst revisions proceeding corporate events should correspond to lower dispersion of earnings forecasts. Also, McNichols and O’Brien (1997) argue analysts tend to cover stocks about which they are most optimistic. Therefore, an endogeneity concern arises in which analysts limit their coverage to stocks with optimistic prospects. Endogeneity concerns imply analysts would rather not cover firms with pessimistic outlooks than actually voicing a negative recommendation. In these cases, analyst forecast dispersion would be relatively low. To address endogeneity issues, we split the sample into dispersion terciles and run regressions with a dispersion interaction term equal to one for stocks in the upper dispersion tercile. We remove stocks in the middle dispersion tercile. The interaction term (REC · D, where D equals one for stocks in the upper dispersion tercile) is small and statistically insignificant.

14 As pointed out by Irvine and Liu (2016), private interactions between analysts and management have persisted in the post Reg FD era. Soltes (2014) cites Thomson Reuters and Bank of New York findings that investor relations officers and senior management still spend considerable time meeting with sell-side analysts to discuss nonmaterial information, which can be pieced together with other information sources to form material information mosaics.

15 The factor returns are taken from Professor Kenneth French’s website: https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html.

16 Similar conclusions are reached from tests on 3-month and 12-month holding periods.

17 To control for event effects, we repeated the analysis by removing all IPOs and delistings. In a separate test, we repeated the analysis after removing discontinued analyst coverage observations. And, in another test, we removed all observations in which the consensus recommendation changed. There were no material changes to our overall conclusions of a significant and positive relation between analyst recommendation and subsequent momentum profits.

18 We also examine alternative methods including simple random one-to-one matching, the caliper matching criterion, Mahalanobis metric, stratification, local linear weights, and the Gaussian kernel function. Results from these alternative methods are consistent with those we present in this paper, adding to the robustness of our tests.

19 According to Daniel et al. (1997), the advantages of characteristic-based matching are multifold: characteristic-based measures are simple and straightforward, they have more statistical power to detect abnormal returns, and characteristics are better predictors of stock returns. The characteristic matching technique is a good fit for our tests because it is designed to produce samples that are matched on multiple dimensions, which, in turn, are used to produce projected recommendations based on public information for uncovered stocks.

20 The projection of the covered stocks into the uncovered stocks subspace can be achieved by two methods: the direct projection method (parallel projection) or the regression method (orthogonal projection). The regression method produces least square error prediction, but loses substantial information in the projection. The regression method also is subject to potential outlier problems. For these reasons, we use the parallel projection method to assign recommendation scores to uncovered stocks.

21 Our approach is analogous to the methods applied by So (2013) who compares analyst earnings growth forecasts against naïve models of earnings forecasts based on prior year company characteristics, summarized by earnings, accruals, change in total assets, dividends, book-to-market, and stock price. The results support the proposition that investors tend to favor analyst earnings forecasts over firm fundamentals.

22 Our matching method includes industry controls. Therefore, to the extent that analyst recommendations are based on private information regarding the industry, then our matching methods will apply appropriate controls.

23 We use changes in analyst recommendations instead of recommendation levels because analyst recommendation levels exhibit strong evidence of non-stationarity for our sampled stocks. Also, prior research shows changes in consensus recommendations affect stock returns Stickel (1995), Womack (1996), Francis and Soffer (1997), Barber et al. (2001), Jegadeesh et al. (2004), Loh and Mian (2006), Green (2006), and Jegadeesh and Kim (2010).