Herding among Investment Newsletters: Theory and Evidence

A model is developed which implies that if an analyst has high reputation or low ability, or if there is strong public information that is inconsistent with the analyst's private information, she is likely to herd. Herding is also common when informative private signals are positively correlated across analysts. The model is tested using data from analysts who publish investment newsletters. Consistent with the model's implications, the empirical results indicate that a newsletter analyst is likely to herd on Value Line's recommendation if her reputation is high, if her ability is low, or if signal correlation is high. HERDBEHAVIOR IS OFTEN SAID TO OCCUR when many people take the same action, perhaps because some mimic the actions of others. Herding has been theoretically linked to many economic activities, such as investment recommendations (Scharfstein and Stein (1990)), price behavior of IPOs (Welch (1992)),fads and customs (Bikhchandani, Hirshleifer, and Welch (1992)),earnings forecasts (Trueman (1994)), corporate conservatism (Zwiebel (1995)), and delegated portfolio management (Maug and Naik (1995)). This paper adds to the herding literature by developing and empirically testing a model that examines the incentives investment advisors face when deciding whether to herd. In particular, the paper tests whether economic conditions and agents' individual characteristics affect their likelihood of herding. The results are interpreted as a test of the predictions of the general class of cascade and herding models.1 * Fuqua School of Business, Duke University. I am grateful to David Hirshleifer and Jaime Zender for comments that helped to substantially improve the paper. I would also like to thank Pete Kyle,Alon Brav, Doug Foster, Dan Graham, Rita Graham, Paul Harrison, Eric Hughson, Ron Lease, Mike Lemmon, Ernst Maug, Susan Monaco, Carl Moody, Barb Ostdiek, Drew Roper, Steve Slezak, Ren6 Stulz, Tom Smith, Brett Trueman, Vish Viswanathan, anonymous referees, and seminar participants at Duke, Tulane, and the University of Utah for helpful comments. I am grateful to Mark Hulbert and The Hulbert Financial Digest for providing the newsletter data, to David Hsieh for providing the daily S&P500 index volatility estimates, and to Yunqi Han and the Federal Reserve Bank of Philadelphia for providing the data on Treasury bill forecasts. I am responsible for all remaining errors. The theoretical part of the paper was a chapter of my doctoral dissertation at Duke University. The empirical work was started while I was at the University of Utah. Welch (1996) also tests implications from the general class of herding models. He finds that brokerage recommendations are influenced by the consensus opinion of many brokers, especially in bullish market conditions or when the consensus proves to be wrong. He interprets the latter condition as being consistent with the implications from models that show that herding is sometimes based on little or no information (e.g., Scharfstein and Stein (1990) or Bikhchandani et al. (1992)). The Journal of Finance We investigate the herding phenomenon using a simple model of stock analysts, patterned after the model in Scharfstein and Stein (1990). Each analyst in our model is one of two types, smart or dumb, although the type is unobservable to all. Smart analysts receive informative private signals about the stock market's expected return, dumb analysts receive uninformative signals. The smart analysts' signals are positively cross-correlated, implying that smart analysts following their private information have a tendency to act similarly. Consequently, in certain circumstances, an analyst can "look smart" by herding. The analysts in the model act sequentially. The theoretical part of the paper investigates several factors that provide incentives for the second-mover to discard her private information and instead mimic the action of the first-mover. The analysts use Bayes' rule to determine their optimal actions and so prior public information is an important input in their decision-making processes, as is the precision of their private information (which we interpret as ability). The amount of correlation across informative private signals is also instrumental because it affects the degree to which analysts can look smart by herding. Finally, given that analysts maximize expected posterior reputation, their prior reputations also influence their optimal decisions. After documenting the existence of parameter regions associated with "herding" and "deviating" equilibria, comparative statics are used to show that the incentive for the second-mover to discard her private information and instead mimic the market leader 1. increases with her initial reputation 2. decreases with her ability 3. increases in the strength of prior public information that is consistent with the leader's action 4. increases with the level of correlation across informative signals. Though these factors are obviously interrelated, it is instructive to isolate the individual contribution of each to herding behavior, rather than blurring the distinction among them, as is often done.2 The intuition behind the reputation implication is that analysts with high reputation (and salary) herd to protect their current status and level of pay? For example, Institutional Investor's All-American Research Team is made up of high reputation analysts. Stickel (1990, 1992) shows that All-Americans give more accurate earnings forecasts and "follow the crowd" less often than non-All-Americans. Based on these findings, it appears that having a high reputation reduces the incentive to herd. In contrast, our model indicates that, to preserve status and salary, high reputation All-Americans have greater incentive to herd than non-All-Americans of equal ability. This implication may seem to contradict Stickel's (1990) finding that All-Americans "follow the crowd" less; however, his results reflect the net effect of reputation, ability, and other factors. We can isolate the effect of reputation on herding only by controlling for the other factors. This is consistent with the implication in Prendergast and Stole (1996) that "youngsters" exaggerate private information to look knowledgeable, while "old-timers" make more conservative decisions. However, their prediction arises because old-timers do not want to deviate too far from their own past decisions, while our model predicts that agents herd on a leader's current decision to remain part of the crowd. 239 Herding among Investment Newsletters We test the implications of the theoretical model with a sample of investment newsletter asset allocation recommendations. A typical newsletter contains four to eight pages of analysis of current economic trends, combined with the newsletter editor's interpretation of how the trends affect various investment strategies. Though the mode and frequency of information transfer varies widely, the typical newsletter is published monthly and mailed to subscribers for an annual fee of approximately $200; some letters also have a telephone, Internet, or fax updating service. The best known investment newsletter is the Value Line Investment Survey. Our sample consists of the market timing advice (i.e., recommendations about what portion of an investor's wealth should be invested in the stock market, cash, etc.) offered by 237 newsletter strategies over the period 1980 to 1992. Using these data, we identify the attributes of newsletters that herd on the advice of Value Line. Our strongest empirical finding is that herding decreases with the precision of private information, which lends support to the broad class of cascade and herding models. We also find evidence supporting the predictions that the incidence of mimicking Value Line increases with newsletter reputation, when a proxy for private information is highly correlated across analysts, and when prior information is strong. The herding literature can be subdivided in the following manner, although these categories are neither exhaustive nor mutually exclusive: (1)informational cascades, (2) reputational herding, (3) investigative herding, and (4) empirical herding. (For a general review of the herding literature, see Devenow and Welch (1996).) The first two types of herding occur when individuals choose to ignore or downplay their private information and instead jump on the bandwagon by mimicking the actions of individuals who acted previously. Informational cascades occur when the existing aggregate information becomes so overwhelming that an individual's single piece of private information is not strong enough to reverse the decision of the crowd. Therefore, the individual chooses to mimic the action of the crowd, rather than act on his private information. If this scenario holds for one individual, then it likely also holds for anyone acting after this person. This domino-like effect is often referred to as a cascade. Research by Welch (1992), Bikhchandani et al. (1992), Banerjee (1992), Lee (1993), Smith and Sorensen (1994), Khanna and Slezak (1998), Banerjee and Fudenberg (1995), and Brandenburger and Polak (1996) investigates cascades. Like cascades, reputational herding takes place when an agent chooses to ignore her private information and mimic the action of another agent who has acted previously. However, reputational herding models have an additional layer of mimicking resulting from positive reputational externalities that can be obtained by acting as part of a group or choosing a certain project. Our theoretical model falls in the reputational herding category. Other reputational herding models include Scharfstein and Stein (1990), Trueman (1994), Zwiebel (1995), Huddart (1996), and Prendergast and Stole (1996). Because these papers deal with issues similar to those investigated by our paper, they are discussed in detail in later sections. 240 The Journal of Finance Investigative herding occurs when an analyst chooses to investigate a piece of information she believes others also will examine. The analyst would like to be the first to discover the information but can only profit from an investment if other investors follow suit and push the price of the asset in the direction anticipated by the first analyst. Otherwise, the first analyst may be stuck holding an asset that she cannot profitably sell. Papers by Brennan (1990), Froot, Scharfstein, and Stein (1992), Dow and Gorton (1994), Hirshleifer, Subrahmanyam, and Titman (1994), and Golec (1997) fall into this group. To an outsider, it can be difficult to differentiate whether an observed "herd" occurs for reasons put forth by models in any of the above categories. Indeed, there is a group of papers that investigate empirical clustering without directly testing the implications of the herding models. Clustering has been observed by Lakonishok et al. (1991), Peles (1993), Grinblatt, Titman, and Wermers (1995), Wermers (1999), Falkenstein (1996), Nofsinger and Sias (1996), and Wylie (1996) among pension funds, mutual funds, and institutional investors when a disproportionate share of investors engage in buying, or at other times selling, the same stock. Among other things, these papers suggest that clustering can result from momentum-following (also called "positive feedback investment," e.g., buying past winners) or perhaps from repeating the predominant buy or sell pattern from the previous period. We control for momentum-following but do not find that it contributes to clustering in our sample. The existing empirical literature largely tests whether "too many" investors appear to make the same choice; our paper attempts to more directly test the implications of the theoretical herding models.4 Another distinction is that our paper tests whether individual analysts take the same action as a "market leader" (Value Line) who sequentially precedes them, rather than examining "clusters" of analysts as is done in most other empirical herding papers. Papers by Lamont (1995) and Ehrbeck and Waldmann (1996) are similar in spirit to ours. Lamont finds that a forecaster's age is positively related to the absolute first difference between his forecast and the group mean. Lamont interprets this as evidence that as a forecaster ages, evaluators develop "tighter priors" about the forecaster's ability, and hence the forecaster has less incentive to herd with the group. We investigate how the log of age affects herding but do not find a statistically significant relation. Ehrbeck and Waldmann find that the empirical patterns across T-bill forecasts are not supportive of simple reputational herding models, but instead seem to support behavioral hypotheses. In contrast, our results are consistent with explanations of herd behavior put forth by theoretical models. The rest of the paper proceeds as follows. Section I develops the reputational herding model. Section I1 contrasts the model with cascades and other herding models, derives empirical implications, and tests the implications with investment newsletter data. Section I11 concludes and offers some thoughts on testing theoretical herding models. * Golec (1997) provides empirical evidence consistent with investigative herding. Herding among Investment Newsletters I. The Reputational Herding Model Consider an economy in which two risk neutral agents, A and B, evaluate an investment. The investment can have either a high (XH) or low (X,) payoff, with the prior probability of the high payoff being a. All of the analysts receive information about the investment payoff in the form of a privately observed signal, which they use to update a.A high signal (s,) provides information in favor of the high investment payoff, a low signal (s,) does the same for a low payoff. A or B superscripts are used to indicate which agent receives the signal. Analysts are of two types, either smart or dumb. Smart analysts receive informative private signals regarding the investment payoff, dumb analysts receive purely random signals. The information structure is symmetric in that