Essays on Market Incentives, Effort and Risk Taking Behavior. Vito Antonio Sciaraffia Palominos

Transcription

1 Essays on Market Incentives, Effort and Risk Taking Behavior by Vito Antonio Sciaraffia Palominos A dissertation submitted in partial satisfaction of the requirements for the degree of Doctor of Philosophy in Business Administration in the Graduate Division of the University of California, Berkeley Committee in charge: Professor Steven Tadelis, Chair Professor Steven Evans Professor Noam Yuchtman Fall 2011

2

3 1 Abstract Essays on Market Incentives, Effort and Risk Taking Behavior by Vito Antonio Sciaraffia Palominos Doctor of Philosophy in Business Administration University of California, Berkeley Professor Steven Tadelis, Chair This dissertation studies the effects of compensation incentives on individual behavior. Specifically, I look at changes in effort and risk-taking behavior among financial analysts, investment managers, and teenagers. The chapters of the dissertation are organized as follows: The first chapter focuses on sell-side analysts. Evidence shows that investors heed to the predictions of sell-side analysts, who influence stock market prices. The informational content of their predictions depends on their incentives to reveal unbiased information. I hypothesize that relative performance evaluations provide analysts with incentives to distort their private information. The incentive cycle of analysts starts and ends mid-year, while their predictions are validated in the middle of this cycle. Hence, analysts with good early performance benefit later from sticking to the consensus (herding), while analysts with poor early performance can only escape their low standing by taking risk and deviating from the consensus (anti-herding). I confirm the hypotheses using data from 1990 through 2010 that include analyst recommendations and stock performance. The results suggest that asset price variation that is not due to fundamental are better understood in the context of the incentives that drive analyst behavior. The second chapter studies whether mutual fund managers have binding time constraints and whether these constraints affect their behavior and performance. I use the discontinuity in the mutual fund manager s workload generated when an additional fund is assigned to her and measure changes in trading activity, volatility, and abnormal returns. I find that managers who experience time constraints significantly reduce the turnover of the fund. I also find that an increase in the manager s time constraints increases the volatility of the fund but has no effect on the abnormal returns. These findings suggest that managers can reduce idiosyncratic risk but do not have the ability to achieve abnormal returns. The third chapter makes use of a natural experiment in education reform in Chile to show that increasing the number of school hours improves student academic performance and reduces teenage risky behaviors. Specifically, I show that an increase in the daily number of school hours of approximately 20% increased academic performance, measured by a biannual standardized national evaluation, by an average of 0.22% per year and that teenage

4 pregnancy rates dropped by an average of 1.6% per year. Furthermore, I found that outcomes are cumulative over time, and that short term exposure to the extended school hours does not have any significant effect, but it is just after a couple of years that the effect kicks in. In the long term, after 10 years of exposure to the program, I observe effects of up to a 2.8% increaseintestscoresanduptoa19%decreaseinteenagepregnancyrates. Theresultsallow me to show that the reduction in birth rates is the consequence of human capital creation and not of pure warehousing as has been suggested in previous literature. 2

5 i My Wife (the love of my life) My Daughters (the happiness of my life) My Father (to whom I owe everything)

6 ii Contents List of Figures List of Tables Acknowledgments iv v vii 1 Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations Introduction The Industry Performance & Compensation Herding and Anti-Herding Behavior Market Reactions Data, Methodology & Hypotheses Evaluation Cycle The Consensus Deviations and Risk-Taking Performance Ranking Hypotheses Empirical Strategy Summary Statistics Results Current Period Deviations from Consensus Current Period Number of Recommendations Evaluation Period Cycle Effects at the portfolio level Conclusion Limitations and Future Steps

7 Contents iii 2 Time Constraints and Performance: The Mutual Fund Industry Introduction The Event Literature Review Data and Methodology Data Source Descriptive Statistics Measures Results Decrease in Turnover Performance and Risk-taking Robustness Tests Concluding remarks Academic Performance and Teenage Pregnancies: A Natural Experiment in Education Reform Introduction Background The Full-Day Schooling Reform Conceptual Framework School Performance Teenage Pregnancies Natural Experiment and Econometric Specifications The Experiment and the Identification Strategy Measuring Performance and Teenage Pregnancy Rates Specifications Data and Statistics School Performance Teenage Pregnancies Empirical Findings School Performance Teenage Pregnancies Discussion Conclusions Bibliography 88

8 iv List of Figures 1.1 Sell-Side Analysts Evaluation Period Second Period Deviations from Consensus among High vs. Low Performers Changes in n. of Recommendations of High vs. Low Performers in Second Period Histogram of the Years of Experience of Analysts Histogram of the Number of Securities Followed per Analyst Histogram of the Recommendations issued per Analyst per Period Original funds detrended turnover dif in dif calculation for quarterly periods Difference in difference estimation Evolution of Schools in Full-Day Schooling Scheme Evolution of Students in Full-Day Schooling Scheme Average Number of Years of Exposure to Full-Day Schooling Scheme

9 v List of Tables 1.1 Number of Analysts, Brokerage Houses and Securities during Years of Experience of Analysts Earnings per Share Forecasts Deviations from Market Consensus Number of Recommendations per Semester Regression Results 1, 2, and Regression Results 4, 5, and Regression Results 7 and Regression Results 9 and Regression Results 11 and Regression Results 13 and Benchmarks Distribution of event dates Descriptive statistics for the original funds in the subsample Difference in difference estimation FF Alphas pre and post event Carhart Alphas pre and post event Volatilities pre and post event Difference in difference estimation Difference in difference estimation (quarterly) - Part Difference in difference estimation (quarterly) - Part SIMCE Test Scores Across Schools Teenage Birth Rates (%), , Ages 16 to 18 Across Districts Effect of Exposure to Full-Day Schooling in Test Scores Effect of Years of Exposure to Full-Day Schooling in Test Scores Effect of Periods of Exposure to Full-Day Schooling in Test Scores Effect of Years of Exposure to Full-Day Schooling in Test Scores Effect of Years of Exposure to Full-Day Schooling (Urban Schools) Effect of Years of Exposure to Full-Day Schooling (Voucher and Public) Effect of Years of Exposure to Full-Day Schooling (Small and Large) Effect of Years of Exposure to Full-Day Schooling in Teenage Birth Rates.. 79

10 LIST OF TABLES vi 3.11 Effect of Years of Exposure to Full-Day Schooling in Teenage Birth Rates Effect of Periods of Exposure to Full-Day Schooling in Teenage Birth Rates Effect of Years of Exposure to Full-Day Schooling in Teenage Birth Rates A Effect of Years of Exposure to Full-Day Schooling in Teenage Birth Rates B Effect of Years of Exposure to Full-Day Schooling in Teenage Birth Rates C 84

11 vii Acknowledgments I am particularly grateful to my dissertation committee chair Steven Tadelis for his continuous guidance and support. Noam Yuchtman deserves special thanks for his thoughtful comments and tremendous help. Steven Evans was a supportive advisor and source of ideas. In addition, I thank Ernesto Dal Bo, Lucas Davis, Alejandro Drexler, Stefano DellaVigna, Ed Egan, Pablo Hernandez, Lucy Hu, David Levine, Dmitry Livdan, Santiago Oliveros, Orie Shelef, and Matthew Watkins for their helpful comments over the years. Paul Gertler and Terrance Odean were especially supportive of my research. I also want to thank my very good friends Gonzalo Maturana, Bernardo Quiroga, and Francisco Simian for all the collaboration, long discussions, and useful suggestions that made this dissertation possible. Finally, I am thankful to the Institute of Business Innovation at the University of California at Berkeley, the Bradley Foundation and the Ryoichi Sasakawa young leaders fellowship fund for their generous financial support during the elaboration of this work.

12 1 Chapter 1 Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 1.1 Introduction Market efficiency critically depends on accurate information being timely and truthfully transmitted. This allows market participants to make optimal decisions regarding the distribution of scarce financial resources. Studying Sell-side analysts behavior is important as their role is to provide information to financial market participants who later aggregate the information they obtain and use it as an input to make investment decisions. In fact, considerable anecdotal evidence shows that sell-side analyst recommendations are closely followed by investors and taken into consideration in their investment decisions. Specifically, institutional and retail investors use equity research to inform their decisions(groysberg, Healy and Maber(2011)). Furthermore, the relation between security returns and analysts forecast revisions suggests that investors use these recommendations to extract relevant information about upcoming earnings from analyst forecasts (Clement and Tse (2005)). Given the importance of sell-side analysts recommendations, the quality of the information contained in their forecast is of great relevance as it helps in the optimal allocation of resources. It is for this reason that their behavior has been examined extensively over the last two decades (Jackson (2005)). For example, Mikhail, Walther and Willis (2007) find that both large and small traders react to financial analyst recommendations, and that unsophisticated investors are sometimes misled into making suboptimal investment decisions by financial analysts. Sell-side analysts usually work for brokerage firms, and their job is to evaluate companies in terms of their potential for future earnings and growth as well as other characteristics that might affect the stock price of those companies. On the basis of this research, they then issue estimates of future earnings per share to investors. Broadly speaking, these analysts

13 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 2 make a living predicting the future. However, sell-side analysts incentives don t always perfectly align with those of investors. In fact, sell-side analysts get paid, in the short and long term, depending on their performance relative to their peers and not on the absolute accuracy of their forecasts (i.e., recommendations). As a consequence, one might expect that these incentives affect the amount and quality of information disclosed by analysts. Strategic reasons may cause sell-side analysts may decide to not truthfully and timely transmit their private information to the market. For example, they may decide to agree with the rest of the analysts forecasts (i.e., herding) even in situations where their private information indicates otherwise, or they may decide to differ with the rest of the analysts forecasts (i.e., anti-herding) in an intent to differentiate themselves from the crowd even if their private information indicates the opposite. Given the importance of the quality of their forecasts, this strategic behavior carries negative consequences to the financial markets and several...academic studies often attribute many market ills such as excess market volatility, the dot-com bubble, and the emerging market meltdown in the 1990s to the phenomenon of herding (Jegadeesh, N. and Kim, W. (2010)). This papers studies the effects of compensation incentives on information disclosure of sell-side analysts. Specifically, within a rational model framework, I show that their recommendations are a consequence of the incentives they face. Particularly, I investigate under which circumstances will analysts herd or anti-herd and provide evidence of how this behavior depends on their past relative performance. Sell side analyst behavior has been extensively treated in the literature (Jackson, 2005), however most of this literature has focused on two distinct aspects that, to my knowledge, have not yet been directly connected. The first of these aspects is the relation between forecast accuracy and compensation, and the second is the herding or anti-herding behavior of sell-side analysts. My intention in this study is to look at the larger picture and describe the connections that exist between these aspects that have heretofore been treated separately. Several papers have looked at whether analysts herd or anti-herd. However, evidence is still confusing as to what kind of behavior they exhibit. On the one hand, Clement and Tse (2005), Gallo, Granger, and Jeon (2002), Hong, Kubik, and Solomon (2000), and Lamont (2002) all show evidence in support of herding. On the other hand, Bernhardt, and Chen and Jiang (2006), Campello and Kutsoatic (2006), and Zitzewitz (2001) all show evidence in support of anti-herding. These two set of papers, though they use similar methodologies to measure herding, use different controls and distributional assumptions. The results in this paper help to reconcile previous literature regarding why sell-side analysts herd or anti-herd. I find that herding behavior is not just a time invariant characteristic of individual analysts, but rather occurs in response to previous relative performance. This paper identifies the fact that evaluation cycles of sell-side analysts compensation are annual, but are more accurately described by mid-year to mid-year instead of January to December as has been represented by the studies described above. My paper contributes to the literature by correctly representing the evaluation cycle of sell-side analysts and showing that contradicting evidence provided by the previous literature

14 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 3 can be reconciled once compensations incentives are incorporated in the forecasts decision process of analysts. Furthermore, I use the same methodology and controls proposed in the previous literature (Gleason and Lee (2005)), and my results remain significant. The central hypothesis of this paper is that relative performance incentives generate a strong non-linearity in the compensation of sell-side analysts. Specifically, there is a cut-off toward the bottom of the distribution that analysts wish to avoid. This study examines this non-linearity to understand the deviation of analysts recommendations from the group consensus regarding Earnings per Share (EPS) forecasts. I further argue that the desire to avoid termination due to bad relative performance generates particular end-of-year patterns of deviation from consensus as a function of earlyin-the-year success. In other words, analysts close to the termination threshold will issue significantly riskier recommendations during the second period. The behavior described above can be explained by the market s incentive structure: sellside analysts who out-perform their peers do not need to take higher risks during the following period because they are participants in a relative performance contest. Consequently, they only need to replicate the market consensus forecast during the following period to remain in their relatively high rank at the end of the year. However, for underperformers the story is different: those who underperform in the first part of the evaluation cycle need to take higher risks in order to make up for their poor past performance (i.e., gamble for resurrection). In other words, poor performers have a lot to win and nothing to lose by undertaking huge risks. The hypotheses are tested and confirmed using data from 1990 to 2010 that is drawn from the Detail History dataset file of the I/B/E/S database. This dataset contains information at the recommendation-stock level and includes the issuing date and time, issuer identification (analyst and brokerage), predicted earnings per share, and realized earnings per share. Interestingly, the data shows that risk-taking behavior is not only consistent with the hypotheses outlined above, but it is also accompanied by a change in the number of recommendations issued per analyst per period; I find sell-side analysts increase the number of recommendations after bad relative performance. Such behavior may have two explanations; On the one hand, analysts change their recommendations continuously to avoid straying into the consensus. This is referred to as the differentiation effect. On the other hand, analysts may work harder, collect more information about companies, and on this basis issue recommendations that differ more markedly from the consensus. This is referred to as the effort effect. It is not the intention of this study to determine which of these two competing effects dominate, which is left for future research. To better characterize the larger scope of an analyst s behavior, I consider his decisions at the portfolio level (i.e., all the securities that are followed by the analyst). When expanding the unit of analysis from a single security to a portfolio of securities the results are robust with regard to risk taking behavior. However, the number of recommendations at the aggregate level does not vary within different relative performance groups. This may suggest that analysts have limited time and are not able (or willing) to blindly increase the

15 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 4 number of recommendations issued, thus providing some support to the effort-effect over the differentiation-effect. However, this issue must be studied further. To my knowledge, the approach that is most similar to that of this paper is found in Clarke and Subramanian (2006). They studied the effect of compensation and employment risk on sell-side analysts forecasting behavior. The authors propose a u-shaped relation between sell-side analysts forecast boldness and their prior performance. However, there are several differences between their paper and mine. Their basic unit of analysis is analyst/stock for only a subsample of firms, and they only take into account the last recommendation per period. Furthermore, they assume that all analysts recommendations are issued simultaneously. Their paper is important because it develops some initial intuition of what is going on in the market of sell-side analyst forecasts. However their analysis does not take into consideration the complete temporal progression dynamic of the analysts behavior. Also, they only consider the analyst/stock unit and do not consider all dimensions of the analysts decision process in their study (i.e., they only consider the analysts and not at portfolio of securities recommendations produced by the analysts). Their results also differ from ours because they find that top performers issue bolder forecasts than the rest of the analysts. My paper shows that by not controlling for the previous period s risk-taking behavior at the analyst level, the conclusions about who herds and who does not are incomplete. Indeed, Clarke and Subramanian (2006) assume that all analysts have the identical risk-taking preferences, and their performance measure is cumulative over the entire analysts career, which I show is not the case. The rest of the paper is organized as follows. In Section II, I provide a brief description of the industry and this study is situated relative to the previous literature. In Section III the data is described, the methodology is explained, and the main hypotheses are presented. In Section IV summary statistics are shown. Section V shows the results, and Section VI concludes.

16 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations The Industry The sell-side analysts industry is divided into 10 categories (e.g. Technology, Media, and Healthcare). Within each of these categories, one can find several subindustries. For example, within Technology there are Internet, Semiconductors, and Software, among others. Within each of these categories, sell-side analysts issue earnings per share forecasts with a clear target date which is generally the end of the year. This particularity of the industry makes it easy to evaluate the performance of analysts forecasts as all of them will see the results of their predictions contrasted to the actual price realization simultaneously. This said, sell-side analysts usually specialize in a subindustry and are evaluated relative to their peers in the same subindustry. Salaries for these analysts vary significantly, ranging from low 100K to over 1M for those considered top performers.therefore, in this industry there are huge payoffs for being relatively good Performance & Compensation Compensation schemes for sell-side analysts typically consist of a base salary plus a performance bonus. This compensation package is negotiated annually and depends on the analyst s relative performance. Hong, Kubik and Solomon (2000) find that financial analysts with good performance are acclaimed by specialized media and are pursued by other brokerage houses. Moreover, Hong and Kubik (2003) elaborate on the career concerns of financial analysts and demonstrate that brokerage house directors take into account the accuracy of analyst forecasts in their consideration of compensation and career opportunities. They claim that analyst forecast accuracy is an important factor for investors, therefore brokerage houses strive to employ the best financial analysts to enhance their reputation and secure more business with investors. According to the same authors, in the absence of explicit compensation data one can observe that analyst compensation at top-tier brokerages is higher than at low-tier brokerages. Therefore the relative differences in compensation may be inferred by looking at the analyst s brokerage house. They also assert that the movement of analysts up or down the brokerage hierarchy serves to indicate positive or negative compensation shock. According to the same authors, poor relative performers are less likely to move up in the brokerage hierarchy and improve their total compensation. Groysberg et al. (2011) complement these findings by showing that poor relative performer analysts are more likely to move to low tier banks or disappear altogether from the sell-side analyst population. The authors indicate that: Fired analyst-year observations had larger forecast errors than other analysts who covered the same stocks, and extremely adverse forecasting outcomes are associated with increased probability of dismissal.

17 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations Herding and Anti-Herding Behavior Herding behavior has been widely studied in the literature. Recent literature s definition of herding behavior (e.g. Gleason and Lee (2005)) classify herding as a behavior in which an analyst revises his forecast toward the market consensus, and they define anti-herding (i.e., boldness) as a behavior in which forecast revisions move away from both the previous forecast and the consensus. Their classification concurs with the currently prevailing approach. Groysberg et al. (2007) assert that sell-side analysts who diverge from the market consensus face a higher risk. Specifically, they show that when analysts recommendations prove to be wrong and diverge from the consensus, they incur reputation penalties, while those who are wrong without diverging the consensus are unlikely to suffer such penalties. Analyst characteristics and their relationship to forecast accuracy have been documented by Clement and Tse (2005). They show that certain analyst characteristics such as past forecast accuracy, employer brokerage house, years of experience, and number securities followed, influence sell-side analyst forecast accuracy. In his investigation of herding behavior in sell-side analysts, Olsen (1996) concludes that the human proclivity toward consensus combined with the internal dynamics and culture of the earnings forecasting analysts leads to herding. Additionally, the author states that this sell-side analyst herding behavior creates noise in the market and ultimately affects stock returns. Scharfstein and Stein (1990) construct a rational-choice model and reach the conclusion that there are several settings in which herding behavior is possible. Specifically, the authors conclude, Herd behavior can arise in a variety of contexts, as a consequence of rational attempts by managers to enhance their reputations as decision makers. Hence, Scharfstein and Stein (1990) show that analysts herd in order to enhance their reputation. Trueman (1994), shows that analysts tend to issue recommendations that are closer to the market consensus if by doing so they stand to improve their reputations. They conclude that more inexperienced and less recognized analysts will herd more. Hong et al. (2000) show that experienced sell-side analysts behave differently than inexperienced ones and that the market reacts differently to their forecasts. Specifically, experienced analysts tend to issue bolder forecasts, and inexperienced analysts tend to lose their jobs more often when issuing ex-post inaccurate bold forecasts. Clement and Tse (2005) extend this idea by investigation what other characteristics of analysts may affect their behavior. They find that previous forecasting accuracy, the prestige of the analyst s brokerage house, the analyst s frequency of recommendations, and the breadth of industries covered by the analysts increase the boldness of his forecasts. They propose the idea that herding is simply a way for analysts to mimic the market consensus, and that these forecasts contain little or no new information. However, the authors do not seek to understand the motivations of this behavior.

18 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations Market Reactions Information disclosed by sell-side analysts to investors has a significant impact on the market and could increase the price volatility of stocks. Specifically, Stickel(1995) and Womack (1996) show a significant positive price reaction to security recommendation upgrades and a significant negative reaction to downgrades. Gleason and Lee (2000) show that market reactions with regard to security returns responses are weaker for forecasts that move toward the market consensus in comparison with those that deviate from it. This raises the question as to whether the market is somewhat aware of the incentives that sell-side analysts face and these different levels of market reactions to forecasts is consequence of a best response function.

19 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations Data, Methodology & Hypotheses The data for this study comes from the Detail History dataset of the I/B/E/S database from the years 1990 to This dataset provides information on recommendations of sellside analysts at the security level and includes the issuing date and time, issuer identification (analyst and brokerage), predicted earnings per share, and realized earnings per share. The data include 3,098 sell-side analysts employed by 441 brokerage houses, following 927 securities and totaling more than 55,000 forecasts. This dataset, which to my knowledge is the longest panel ever used for studying herding behavior, allows me to test different specifications and use a vast set of controls Evaluation Cycle Figure 1.1: Sell-Side Analysts Evaluation Period Figure 1 shows a comparison between the usually hypothesized evaluation cycle used in the literature and the actual evaluation cycle which I use in this paper. Figure 1.1 shows how the actual evaluation period compares to the usually hypothesized evaluation period used in the literature. As the figure shows the first evaluation period is actually the second semester of every calendar year, whereas the second evaluation period is the first semester of the following calendar year. The present study will employ the actual evaluation period, which better represents many of the events that occur in the financial analyst forecast industry. For example, new analysts join brokerage firms shortly after graduation (i.e., in July or August), and bonuses, promotions and terminations are generally decided at the end of the year after performance evaluations are completed. This performance evaluations are based, among other things, the

20 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 9 result of publicly available analysts rankings (e.g., Institutional Investor Magazines Ranks) that are usually constructed add released after the end of the fiscal year. All of the above factors provide strong support for the idea that the evaluation year begins in July and ends the following June. If analysts yearly relative performance affects their utility one would observe that analysts strategically change their behavior in the second part of the evaluation cycle depending on their prior (first period) relative performance. Therefore, this study focuses on level changes in the analysts risk-taking behavior. Using the same reasoning, this study considers the number of recommendations per analyst per period. More specifically, second period behavior both in risk-taking and in the number of recommendations issued is considered in response to their first period relative performance. Figure 1.2: Second Period Deviations from Consensus among High vs. Low Performers I define low performers as the half of the total analysts with the lowest accuracy, measured as the absolute difference between the forecast and the realized value of the stock. In the same way, I define high performers as the half of the total analysts with highest accuracy. As shown in the Kernel density in Figure 1.2, high performers tend to have lower second period deviations from the market consensus compared to their low performing peers. Furthermore, within the low performer group there appears to be a sub-group characterized by even greater deviations from the market consensus. This difference in deviations serves as preliminary evidence that past relative performance may alter behavior. Figure 1.3 shows that low performers increase the number of their recommendations in the second period compared to high performers, a fact that provides further evidence that

21 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 10 changes in behavior are multidimensional (i.e., involving frequency of recommendations and deviation from the consensus). Figure 1.3: Changes in n. of Recommendations of High vs. Low Performers in Second Period The Consensus I start with the construction of the continuous time consensus through a calculation of themean 1 foralltheoutstandinganalystrecommendations(i.e., earningspershareforecasts) for each security at each target period. This means that every time a new recommendation is issued, the continuous time consensus immediately changes in order to correctly capture all of the information that is available to the analyst at the time his decision is made. If an analyst updates one of his older recommendations through a reissue, the newer recommendation overwrites the one that preceded it; consequently the older recommendation ceases to be part of the consensus. Therefore, there is never more than one active recommendation per security per analyst in the consensus. Financial markets are dynamic and use information that is continuously updated, and the continuous time consensus measure is designed to reflect this reality. Consequently, this measure is better than measures that are updated at discrete intervals. This methodology 1 An alternative would be to use the median as a robustness test, because median may be thought as more about relative performance than mean is. However, this extension is left for future research.

22 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 11 is used to calculate the consensus for every security in the sample, while carefully removing inactive and outdated recommendations as well as duplicates. Equation 1 illustrates the construction of the consensus. The consensus for security k in period t is the mean of EPS forecasts for that same security issued by all analysts, from i=1 to n, in period t. consensus k,t = 1 n n eps estimate k,i,t (1.1) i= Deviations and Risk-Taking I use the absolute deviation 2 from an existing consensus as a measure of risk-taking. This means that the further an analyst s recommendation departs from the consensus, the higher the risk he assumes in issuing it. Equation 2 illustrates the procedure used to calculate the deviations per analyst/security pair in each period. The difference between each security recommendation, j, for each analyst, i, in period, t, and the existing consensus is first calculated. Next, for each analyst all these deviations for security, k are summed up and then divided by the number of recommendations, J, issued by this analyst in period t. This procedure yields a measure of mean deviations from the consensus for each analyst/security/period triplet. deviation k,i,t = 1 J J eps estimate k,i,t consensus k,i,t (1.2) j=1 A measure of aggregated deviations at the portfolio level is also constructed to capture the behavior of analysts, taking into account all the securities they follow in each industry. Equation 3 shows the procedure. The deviations are averaged for all securities, k, in analyst i s portfolio to arrive at a measure at the analyst/period level. portfolio deviations i,t = 1 K K deviations k,i,t (1.3) k= Performance I use the absolute deviations from the realized EPS value as a measure of performance. This means that the further away a recommendation is from the realized EPS value, the lower the accuracy of this recommendation (equation 4). It is important to emphasize that, by construction, the higher the performance value is, the poorer the actual performance. performance i,k,t = eps estimate i,k,t eps actual i,k,t (1.4) 2 I could alternatively use the square of the deviation. However, this is just a matter of re-scaling and the results remain the same.

23 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 12 After all deviations have been calculated, an aggregate measure of these deviations is constructed. For each period I calculate the mean of the deviations per security per analyst and match it to the corresponding analyst Ranking I group analyst aggregated performances by security and assign a relative ranking to each, ranging from 1 (lowest absolute deviation / best performance) to n (highest absolute deviation / poorest performance). This procedure to construct the ranking is standard in the literature. Furthermore, Leone and Wu (2007) find a strong positive relation between widely used industry rankings (e.g. Institutional Investor Ranking) and sell-side analyst performance. Finally, the ranking is divided into five different groups(quintiles): Excellent Performers, Good Performers, Average Performers, Mediocre Performers, and Poor Performers. The reason for the use of a semi-parametric approach is to further explore the effects of the previous period performance because an analysis based on a single variable may not permit the determination of the effect s true value. For example, the coefficient associated with a previous period performance may lose significance because of the existence of non-linearity in the responses of different groups of analysts Hypotheses In this paper I study how sell-side analysts adjust their current period risk-taking behavior and the number of recommendations they issue as a response to their previous period s performance relative to their peers. Therefore, I hypothesize that relative performance evaluations provide analysts with incentives to distort their private information. More specifically, I claim that analyst behavior is a non-linear response across the performance distribution of analysts Empirical Strategy Consistent with previous literature, the specifications in this study include fixed-effects for analyst, brokerage firm, security, and year (unless otherwise specified). Therefore, the findings should not be biased by the omission of specific, important sell-side analyst or institutional characteristics. The main specification is: deviation i,k,t = α+β deviation i,k,(t 1) +γ 1 brokerage+γ 2 security +γ 3 analyst +γ 4 experience+γ 5 year+δ 2 rank q2 +δ 3 rank q3 +δ 4 rank q4 +δ 5 rank q5 (1.5)

24 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 13 Specification 1 shows the main regression to be used in this paper as well as the use of dummies for each ranking quintile.

25 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations Summary Statistics This section provides an overlook of the main variables used for this study. The basic unit of observation used corresponds to earnings per share forecasts from the years 1990 to 2010, and the total number of observations is 143,378. However, after eliminating from the sample the securities with low analyst coverage the sample has approximately 55,000 observations 3. Table 1.1: Number of Analysts, Brokerage Houses and Securities during Year Analysts Brokerage Houses Securities Total Table 1.1 shows the evolution of the number of sell-side analysts covering US firms from the year 1990 to The number of financial analysts increased by 61% during this period. Also, notice that there are a total of 3,098 unique financial analysts in the sample. Similarly, the number of Brokerage Houses increased by 55% for the same period, thus keeping the average number of analyst per brokerage house, 6, relatively constant when comparing the 3 This elimination is brought about by either of these two conditions: the existence of fewer than 10 recommendations per security per period, or having fewer than 5 analysts actively following that security during that period.

26 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 15 year 1990 to the year On the other hand, the number of securities covered by financial analysts increased by 52%. Notice that all three variables remain relatively in the same proportion while comparing the beginning and the end of the series. However, by carefully looking at the evolution of the series, one can observe that the number of brokerage houses and securities have increased relatively constantly over the 20 year period, while the number of analysts significantly increased in the late nineties and then decreased during the second half of the following decade. This is consistent with the financial market dynamics because analysts are much easier to hire and terminate with business cycles, while the creation and closure of brokerage houses is not. Experience Table 1.2: Years of Experience of Analysts 10th %ile 25th %ile Mean 75th %ile 90th %ile Std Dev Years Table 1.2 shows the number of years of experience of analysts. The distribution is not symmetric, and the experience ranges from 0 to 20 years. The average distribution is approximately 6 years with a standard deviation of 4 years. Figure 1.4 provides additional information; the histogram shows how most of the analysts are grouped around 3-4 years of experience, and the distribution is highly skewed to the right. These statistics are consistent with other studies of the industry. Table 1.3 shows the deviations from consensus of the earnings per share forecasts of analysts. Observe that that second semesters are characterized by higher standard deviations in the forecasts and that the distribution flattens in the first semesters. Table 1.3: Earnings per Share Forecasts Deviations from Market Consensus Deviations from Consensus (dollars) 10th %ile 25th %ile Mean 75th %ile 90th %ile Std Dev Complete Sample if semester = if semester = Sub Samples if semester = if semester = if semester = if semester = As for the number of recommendations, observe that in the second periods there is a slightly higher number of recommendations issued. This comes accompanied by an increase

27 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 16 Figure 1.4: Histogram of the Years of Experience of Analysts in the standard deviation of recommendations per period per analysts. The next section further explores how the figures from table 1.3 and table 1.4 are consistent with the proposed two period evaluation cycle described in section III. Figure 1.5 shows the number of securities followed per analysts. Observe that most of the analysts follow less than 20 securities, while analysts following only 1 security represent roughly 15% of the population. Figure 1.6 shows the comparison between the numbers of recommendations issued per analyst in each of the two periods. Observe from this comparison (which is a visual extension of Table 1.4) that second periods are characterized by an increase in the number of recommendations and an extension of the right tail of the distribution.

28 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 17 Table 1.4: Number of Recommendations per Semester Number of Recommendations per Semester 10th %ile 25th %ile Mean 75th %ile 90th %ile Std Dev Complete Sample if semester = if semester = Sub Samples if semester = if semester = if semester = if semester = Figure 1.5: Histogram of the Number of Securities Followed per Analyst

29 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 18 Figure 1.6: Histogram of the Recommendations issued per Analyst per Period 1.4 Results This section is divided into four parts: first, I present the results corresponding to the previous period performance effect on the subsequent period s deviations from consensus; second, I present the results corresponding to the previous period performance effect on the subsequent period s number of recommendations; third, I verify that the evaluation cycle really is played in two periods at a time and that the analysts behavior resets to their baseline levels at the beginning of the next cycle ;fourth, I explore whether these results change when considering the analysts aggregate portfolio decisions Current Period Deviations from Consensus Specifications 1, 2, and 3 in Table 1.5 capture the effect of the previous period s performance on the current period risk-taking behavior for the entire period between 1990 and 2010, for the total sample of 55,311 observations. From specification (1) in Table 1.5 one can observe that: previous period deviations from consensus positively affect the deviations from the consensus in the current period. The coefficient from this variable is significant at the 1% level. This result seems logical as intuition suggests that each sell-side analyst has different risk preferences and therefore different base risk levels that are captured by the unconditional first-period behavior. Second, consistent with the hypotheses in Section III, the previous period ranking positively affects the deviations from the consensus in the current period. The coefficient from the regression, which is significant at the 1% level, shows that the higher the ranking (i.e., worst previous period performance), the higher the deviations from consensus in the current period. Furthermore, the effect is non-linear and significantly higher when one compares the coefficients across performance groups. Third, another important factor is the experience of the sell-side

30 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 19 Table 1.5: Regression Results 1, 2, and 3 (1) (2) (3) Previous semester deviations 0.405*** 0.503*** 0.503*** (10.63) (21.98) (21.27) Previous semester deviations squared *** *** (-2.91) (-2.91) Previous semester performance: Excellent Performers (omitted) (omitted) (omitted) Good Performers (0.58) (0.90) (0.90) Average Performers 0.028** 0.028** 0.028** (2.22) (2.28) (2.25) Mediocre Performers 0.045*** 0.047*** 0.047*** (3.58) (3.91) (3.90) Poor Performers 0.075*** 0.082*** 0.082*** (4.69) (5.33) (5.34) Experience 0.052*** 0.046*** 0.046*** (13.87) (17.55) (17.36) Number of securities followed ** (2.40) Constant 0.256* (1.71) (1.38) (1.12) Semester-Year Fixed Effects YES YES YES Analyst Fixed Effects YES YES YES Security Fixed Effects YES YES YES Broker Fixed Effects YES YES YES r N 55,311 55,311 55,311 Statistically significant *at the 10% level ** at the 5% level *** at the 1% level Coefficients with t-statistics robustly estimated allowing for cluster -intragroup- correlation.

31 Chapter 1. Compensation Incentives and Risk Taking Behavior: Evidence from Sell-Side Analysts Recommendations 20 analyst. Observe that the experience coefficient, which is positive and highly significant, suggests that analysts that have been in this industry for a longer time tend to deviate more from the consensus. There are several possible explanations to this phenomenon: it may be that higher risk pays off and that these analysts survive longer in the set, or, that more experienced analysts cultivate a reputation that gives them a cushion in case of mistakes, so they tolerate higher risk. To further explore the central hypotheses of this paper the square of the previous period risk included in specification (2) in Table 1.5. This gives more flexibility to the underlying form of the model and does not change the results compared to the previous Specification (1). After performing a test to compare the coefficients of these two regressions (i.e., (1) and (2)) it is not possible to reject the alternative hypothesis that these two are different at any reasonable significance level. This result is important because it demonstrates that even after imposing a diverse set of controls, the main variable of interest remains significant both economically and statistically. Specification (3) in table 1.5 includes an additional control (number of securities followed per analyst). This takes into account the existence of limited attention together with capacity constraints. The results show that the higher the number of securities followed by an analyst, the more he deviates from the consensus. This may be due to several reasons, one is that analysts may gamble more when they have a portfolio of securities. A logical extension is to study the behavior of analysts at the portfolio level. This is addressed in the third part of this section Current Period Number of Recommendations In a slightly different approach, the specifications in table 1.6 show how the number of recommendations per period varies depending on the previous period performance. As expected, the number of recommendations increases when the past performance is bad. This may suggest that bad performer analysts exert more effort in the second period in order to make up for their previous bad accuracy. For consistency with the previous specifications, additional controls include the number of firms followed in the current period and number of firms followed in the previous period, and the results remain the same. Also, notice that the experience of the analyst plays an important role in the number of recommendations issued per period. The more experienced the analyst is, the more recommendations he issues. This may be due to the institutional knowledge he acquires through the years that allows him to revise his recommendations with less effort or cost. Finally, as expected, the number of recommendations in the second period of the evaluation cycle is well predicted by the number of recommendations in the first period Evaluation Period Cycle If the past always affects the future and things do not reset every year, one should expect to see that further previous periods continue to affect future behavior of analysts.