The Market Response to the Earnings- Guidance Game of Financial Analysts

Transcription

1 The Market Response to the Earnings- Guidance Game of Financial Analysts Master Thesis Joseph Benjaminsen (225030) 29 November 2010, Tilburg Faculty of Economics and Business Administration Department of Finance Master: Financial Management Supervisor: Dr. P.C. de Goeij

2 Abstract This paper examines the market impact of the earnings-guidance game played by financial analysts. Analysts hereby make optimistic earnings forecasts at the beginning of the quarter and deflate these forecasts with a final revision near the end of the quarter. By setting beatable targets and playing the expectation game, analysts are pleasing firm management. This paper investigates the U.S. stock market reaction regarding this strategic deflating behavior of analysts. Using a sample covering the period from 1996 until 2006 the relation between the earnings revision and the cumulative abnormal return (CAR) in the period around the actual earnings announcement is examined. I find that there is a significant negative relation between the revision and the CAR. After a downward revision the average CAR increases with 129 basis points in the days after the actual EPS announcement. Finally, the findings in this study confirm the evidence of the rewards for meeting or beating earnings expectations (MBE), in line with previous research. 2

3 Preface This master thesis presents the results of my empirical research on earnings revisions. The process of writing this thesis increased my interest and understanding of earnings forecasts issued by financial analysts. The behavioral biases and analysts incentives that underlie the manipulation of these earnings forecasts make this subject very interesting to study. There are several persons who I want to thank for their contributions. First, I would like to thank Dr. P. C. de Goeij. He provided me the general idea for the subject as well as the dataset I used for my research. Furthermore, his insights, feedback and our long conversations were very helpful during the writing process. In addition, I would like to thank Prof. Dr. V. Ioannidou for her time and role in the exam committee. A lot of gratitude goes out to my parents. They supported and motivated me during my whole study and gave me the possibility to study at. Also I would to thank my brother, sister and girlfriend for their encouragement. Finally, I would like to thank all my friends who supported me during my years of study and made my time in Tilburg unforgettable. I am very pleased with the final result of this master thesis and sincerely hope that everybody enjoys reading it. Joseph Benjaminsen November 2010, Tilburg 3

4 Table of Contents Abstract...2 Preface Introduction Literature Review Introduction Market Efficiency Market and Analyst Efficiency Analysts' Incentives and Behavioral Biases Market Return and Meeting or Beating Expectations Research Method Introduction Methodology Control Variables Sample and Data Data Descriptives Introduction Descriptives Forecast Fault over Time Empirical Results Introduction Cumulative Average Abnormal Return Testing for Market Response Cross-Sectional Results Revision Table Surprise Table Robustness Checks Introduction Positive and Negative Revision Observations Full Sample Revision Analysis

5 6.4 Stepwise Regression Analysis Regulatory Impact Introduction Cumulative Average Abnormal Return Expectation Management Regulatory Impact Discussion Introduction Discussion Contributions and Future Research Conclusions References Appendix

6 Chapter 1 Introduction Meeting analysts earnings forecasts is very important for companies nowadays. Failing to meet or beat these expectations can lead to large downward stock price adjustments. Not surprisingly, firm management has incentives to avoid these unwanted negative earnings surprises. The following quote from CFO magazine typifies the pressure that finance executives face near the end of the quarter: As quarterly earnings announcements draw near, finance executives can safely predict at least one increase -- in pulse rates. With Wall Street's earnings targets for 1998 higher than ever and investors skittish about the course of a long-running bull market, companies that miss targets, even by small margins, face unpleasant consequences in the stock market. No wonder strategies for nudging targets downward are about as legion as cold remedies, and seldom more reliable. 1 The increased importance of meeting earnings expectations changed companies to active players who try to alter reported earnings or manage analysts expectations (McGee, 1997 and Bartov et al., 2000). According to Richardson et al. (2004): firms play an earnings-guidance game where analysts make optimistic forecasts at the start of the year and then walk down their estimates to a level the firm can beat by the end of the year. Most of the evidence from empirical studies has found systematic analyst optimism. It is only recently that empirical studies have found systematic analyst forecast pessimism relative to actual quarterly earnings (Brown, 2001). Bosquet et al. (2009) also found forecast pessimism in a large U.S. data sample with quarterly earnings forecasts. Similar to Richardson et al. (2004), the authors found empirical evidence that: financial analysts strategically inflate their initial forecast at the beginning of the forecasting period, but deflate their forecast in their final revision near the end of the forecasting period in order to please firm management. 1 See McCafferty (1997) and Kasznik and McNichols (2002). 6

7 Bartov et al. (2002) found evidence of a higher stock market return for firms that manage to meet or beat earnings expectations (MBE). This paper investigates the U.S. stock market reaction regarding this strategic deflating behavior of analysts. With a sample covering the period 1996 until 2006 also used in the Bosquet et al. (2009) paper, the relation between the earnings revision and the cumulative abnormal return (CAR) in the period around the actual earnings announcement is further examined. In three empirical models and measured over seven different research intervals, statistical tests must determine how the U.S. stock market reacts to the management pleasing behavior of financial analysts. This paper is organized as follows. The next chapter gives an overview of the relevant literature. Chapter 3 will describe the research method. Chapter 4 will present the data descriptives. The empirical results are provided in chapter 5, followed by a chapter discussing the robustness checks. Chapter 7 presents the regulatory impacts and chapter 8 provides a discussion of the results. Finally, chapter 9 concludes. 7

8 Chapter 2 Literature Review 2.1 Introduction Numerous studies in the corporate finance literature analyzed earnings forecasts made by financial analysts. Ramnath et al. (2008) make a thorough review on the research regarding the role of these financial analysts in capital markets. The authors suggest seven broad areas in the financial analyst forecasting literature: (1) analysts' decision processes; (2) the nature of analyst expertise and the distributions of earnings forecasts; (3) the information content of analyst research; (4) analyst and market efficiency; (5) analysts' incentives and behavioral biases; (6) the effects of the institutional and regulatory environment (including cross-country comparisons); and (7) research design issues. The research in this paper will contribute to the literature mentioned in the above stated areas (4) and (5). Since the relation between management pleasing behavior and stock market return will be examined, both stock market efficiency and the biases underlying analysts recommendations are going to be investigated. Furthermore, the literature on stock market returns in conjunction with meeting or beating earnings expectations will be elaborated in this section. 2.2 Market Efficiency The efficient market theory (EMT) formulated by Eugene Fama in 1970 seeks to explain the behavior of capital markets. The central part of this theory is the efficient market hypothesis, which asserts that all available information is incorporated in security prices immediately. According to this hypothesis prices always fully reflect available information and it is not possible to consistently outperform the market. The efficient market hypothesis is commonly stated in three forms: weak form, semi-strong form and strong form efficiency. According to Fama (1970) each form is based on different information subsets. The different information subsets are: information on past prices, publicly available information and all information for respectively the weak, semi-strong and strong form efficiency. A brief overview, of the three forms of efficiency and their information subsets, is presented below: 8

9 1) In weak form efficiency markets it is impossible to earn extraordinary profits by finding patterns in stock price movements. This form of efficiency denies that future movements can be predicted from past movements and therefore rules out the profitability of a technical analysis. 2 2) A financial market is called semi-strong efficient if all public information is incorporated in securities prices. Neither fundamental analysis (analyzing historical and present data like annual reports, company statements and economic forecasts to predict stock prices) nor technical analysis are therefore able to generate superior returns on a consistent basis. 3) In the most extreme form of market efficiency, i.e. the strong form, prices reflect public as well as private information. On strong form efficient markets it is impossible to gain abnormal returns even with private (or inside) information since this kind of information is already transferred into stocks prices. The literature examining market efficiency is vast and increased enormously during the nineteen seventies and eighties. In his influential paper, Fama (1970) found empirical evidence supporting the weak and semi-strong form efficient hypotheses and viewed the strong form efficient markets model as a benchmark against which deviations from market efficiency can be judged. Jensen (1978) states that the semi-strong form is the generally accepted paradigm and usually meant by references in the literature regarding the efficient market hypothesis. He also points out that there is evidence inconsistent with the strong form efficient market hypothesis, but that it is surprising that such evidence is so scarce. Jensen proposes a weaker and economically more sensible version of the efficient market hypothesis saying that prices reflect information to the point where the marginal benefits of acting on information (the profits to be made) do not exceed the marginal costs (Jensen, 1978 and Fama, 1991). Furthermore, he mentions that tests of market efficiency are tests of a joint hypothesis; market efficiency and the underlying asset pricing model are tested. Therefore these tests can fail because one of the two hypotheses is false or because both parts of the joint hypothesis are false. This joint hypothesis problem can make it very difficult to interpret results. 3 In a sequel to his first paper on efficient capital markets, Fama (1991) reviewed the market efficiency literature. In line with the research following his paper in 1970, Fama renamed the categories of market efficiency. Tests for return predictability, event studies and tests for private information are the 2 See Ross, Westerfield and Jaffe (2006) and Benjaminsen (2006). 3 See for example Roll (1977) for criticisms regarding the asset pricing models. 9

10 new practical definitions for respectively the weak, semi-strong and strong form efficient market tests. Although some anomalies are found, Fama again finds supportive evidence for semi-strong market efficiency in the preceding event-study boom. The research regarding private information tests is still scarce, but some evidence rejecting the strong form efficient market hypothesis is found. In more recent years the evidence regarding the strong form efficiency still remains mixed. Kara and Denning (1998) find elasticities of insider trading profits and reject the strong form market efficiency hypothesis for U.S. capital markets. McKinley (1999) views the U.S. stock market (NASDAQ, American Stock Exchange, and New York Stock Exchange) as an efficient market. He states that, with insider trading as exception, it is impossible to outperform the market average without luck, thereby leaving room for insiders to beat the market and bypassing the strong form market efficiency principles. Gersdorff and Bacon (2009) find support for semi-strong market efficiency in U.S. merger and acquisition announcements and also state that strong form efficiency is generally not supported within markets. The Securities and Exchange Commission (S.E.C.) has regulated insider trading in the U.S. since 1934 (Fishman and Hagerty, 1992). By preventing insider trading the U.S. government clearly indicates the possibility of excessive profits that can be made by insiders. If the strong form efficiency hypothesis actually holds, regulation would not have been necessary in the first place. Under strong form market conditions, insiders do not have the opportunity to earn abnormal profits with inside information. The U.S. insider trading regulations can be seen as proof against full market efficiency, i.e. the strong form efficient hypothesis suggesting that both public as well as private information are incorporated in securities prices immediately. In this paper the event study methodology will be used to investigate the stock price reaction of actual EPS announcements on the U.S. stock market. This approach can also be seen as a test of semi-strong form efficiency. Since stock recommendations (and the underlying behavioral biases for financial analysts) are public information, the EMT predicts that achieving abnormal returns is not possible when following stock recommendations. The financial consequences of management pleasing behavior by analysts should not go unnoticed by the general investment public and will be transferred to stock prices immediately. Making abnormal profits on a recommendation trading strategy should therefore be impossible. 10

11 2.3 Market and Analyst Efficiency Since analysts are viewed as sophisticated processors of financial information they are less likely, compared to naïve investors, to misunderstand the implications of financial information (Ramnath et al., 2008). Therefore, evidence of inefficient information processing by these financial specialists can be seen as proof of overall inefficiency by market participants. Another reason for examining analyst forecasts is that these forecasts do not suffer from benchmark issues like the joint hypothesis problem. According to Fama (1998) the expected return benchmark used in measuring abnormal return may be misspecified. Since analysts recommendations do not suffer from this benchmark problem it provides an opportunity for studying general market inefficiencies. The remainder of this section consists of a literature review that closely follows the Ramnath et al. (2008) paper. The authors reviewed analyst forecasting papers since 1992 and present a clear summary of the key insights of these papers. Easterwood and Nutt (1999) indicate that analysts underreact to negative information and overreact to positive information. 4 Therefore analysts are interpreted as being systematically optimistic in response to new information. Contrary to Easterwood and Nutt (1999), Mikhail et al. (2003) are unable to document analyst overreaction. Then again, Zhang (2006) finds results in support of an underreaction hypothesis. Clearly, the evidence is mixed regarding the question whether analysts forecasts and recommendations efficiently incorporate the information from earnings. With respect to the question whether analyst forecasts and recommendations efficiently reflect information from sources other than earnings, the literature is more consistent. Bartov and Bodnar (1994) for example, find that analyst forecast errors are correlated with changes in currency exchange rates. Elliott, Philbrick, and Weidman (1995) state that analysts systematically underweight new information, especially in the case of downward revisions. The results from Abarbanell and Bushee (1997) imply that analysts ignore available non-earnings information when making forecast revisions. Chaney, Hogan and Jeter (1999) find that analysts do not interpret the future implications of past restructuring charges appropriately, and Burgstahler and Eames (2003) also find inefficiencies. The 4 See Ramnath et al. (2008) for an overview of the key results on research related to financial analysts forecasts and stock recommendations. 11

12 authors mention that firms manage earnings to avoid losses. Although analysts are aware of this behavior, they are unable to identify which firms engage in this loss avoiding behavior. Finally, Shane and Stock (2006) find evidence that analysts' forecasts do not fully reflect the incentives of a firm to manage their earnings to mitigate taxes. The papers mentioned here indicate that not all available information is used by analysts when making forecasts and recommendations. The literature examining efficient reflection of information in analysts forecasts and recommendations seems to report quite similar results. Womack (1996) findings suggest that the market does not fully incorporate the information in sell recommendations. According to Barber et al. (2001) it is possible to earn an annual abnormal return of over 9% with a trading strategy that is based on buying (selling short) stocks with the most (least) favorable stock recommendations. Gleason and Lee (2003) find that, in particular cases, investors tend to underreact to forecast revisions. The results of Mikhail, Walther, and Willis (2004) indicate that the market recognizes superior recommendation ability. The response to both upgrades and downgrades is stronger if recommendations are made by superior analysts. But again, the authors find that the response is incomplete. Li (2005) also finds that superior analysts recommendations are not fully incorporated by the market. Furthermore, short-term price reactions to revisions are larger, as reported by Sorescu and Subrahmanyam (2006), for more reputed and more experienced analysts. Although it appears that markets are inefficient with respect to specific pieces of information, according to Barber et al. (2001) exploiting these anomalies is not profitable because of transaction costs. This evidence is in line with the market efficiency definition of Jensen (1978) which is based on marginal benefits and costs. The last area of the market and analyst efficiency literature examines whether or not earnings forecasts explain inefficiencies in stock prices with respect to publicly available information (Ramnath et al., 2008). Dechow, Hutton, and Sloan (1999) find evidence suggesting that investors do not fully adjust for predictable errors in analyst forecasts. Inconsistent with the results of La Porta (1996), Doukas, Kim, and Pantzalis (2002) find evidence against the hypothesis that analysts are pessimistic (optimistic) about value ( glamour ) stocks. Ikenberry and Ramnath (2002) state that analyst underreaction contributes to the market underreaction regarding stock split information. Finally, the paper written by Jackson and Johnson (2006) implies that momentum in returns follows fundamental news. Again, the papers clearly point out the inefficient stock price reactions with respect to public information. 12

13 2.4 Analysts' Incentives and Behavioral Biases Financial analysts, although considered to be sophisticated investors, are still subject to different biases (De Bondt and Thaler, 1990). Sell-side analysts, often employed by brokerage firms, assimilate and process publicly available information, and acquire private information (Bosquet et al, 2009). By issuing recommendations and earnings forecasts they disseminate new information and are therefore an important source of information to the stock market in the valuation of firms (Schipper, 1991). Although the efficient market hypothesis postulates that all market agents are rational, it is nevertheless documented that analysts earnings forecasts systemically deviate from the rational decision process (De Bondt and Thaler, 1990; Abarbanell, 1991; Brown, 1997; Easterwood and Nutt, 1999). There are different explanations for these forecast inefficiencies. According to Bosquet et al. (2009) these systematic deviations from rationality in the decision making process of financial analysts can be assigned to a behavioral bias (e.g. overconfidence), or conflicts of interest, also referred to as strategic incentives (Friesen and Weller, 2006). Ramnath et al. (2008) mention that there are two different categories of biased forecasts. Biased forecasts driven by judgment errors are distinct from forecasts that are influenced by economic incentives. The former is non-motive driven, while the latter is motive driven. These incentives underlying the decision process of analysts can be related to career concerns, analyst s compensation, the underwriting and trading incentives of their employers and how communication with company management is established (Bosquet et al., 2009 and Ramnath et al., 2008). There is much literature analyzing incentives and biases. 5 McNichols and O'Brien (1997) find that analysts cover firms, about which they have optimistic views, implying a selection bias in coverage decisions. In line with this, Hayes (1998) reports that incentives for gathering information are strongest for stocks that are expected to perform well, so forecasts are likely to be more accurate for such well performing stocks. Finally, the results of Hong et al. (2000a) indicate that forecast accuracy is directly related to the likelihood of promotion. This seems especially true for less experienced analysts. Many studies have found systematic analyst optimism. Francis and Philbrick (1993) report that earnings forecasts are more optimistic for sell and hold stocks than for buy stocks. This result suggests that analysts try to maintain relationships with managers when recommendations are negative. Chen and 5 Again, this literature review closely follows Ramnath et al. (2008). 13

14 Matsumoto (2006) find that managers provide more information to analysts who issue more favorable recommendations (Bosquet et al., 2009). Furthermore, Dugar and Nathan (1995) find that earnings forecasts and recommendations are relatively optimistic when issued by analysts of brokerage firms that also provide investment banking services. On the contrary, the results of Ljungqvist, Marston, and Wilhelm (2006) indicate that optimistic recommendations do not appear to increase underwriting business. The results of Das, Levine, and Sivaramakrishnan (1998) indicate that analysts make relatively optimistic forecasts when earnings are least predictable. This suggests that analysts believe that by issuing optimistic forecasts, they obtain better information from managers. The paper written by Lim (2001) states that forecast bias varies predictably as a function of firm size, analyst coverage, company specific uncertainty and brokerage size, suggesting that analysts may rationally bias forecasts to improve management access and accuracy. According to Bosquet et al. (2009) it is only recently that empirical studies have found systematic analyst forecast pessimism relative to actual quarterly earnings (Brown, 2001). One of the reasons which might explain this shift in analysts behavior is the adoption of several reforms by the SEC in 2002 to pursue the objectivity of the financial analyst s research (Bosquet et al., 2009). Furthermore, Markov and Tan (2006) indicate that analysts have incentives to systematically underpredict earnings in order to keep earnings at a beatable level and to ensure a positive earnings surprise (Bosquet et al., 2009). The results of both Lin and McNichols (1998) and Michaely and Womack (1999) indicate that investors respond differently to recommendations made by affiliated analysts, implying that investors consider analysts incentives. Furthermore, the findings of Barber, Lehavy and Trueman (2007) indicate that the market can unravel optimism in recommendations made by investment bank related analysts. But then again, Hayes and Levine (2000) find that adjusting for bias makes forecasts more accurate and less biased, but no more correlated with contemporaneous returns. Their results suggest that either the market does not adjust for bias or the adjustment captured by the researchers is not the same as the market's adjustment. By exaggerating information, analysts increase trading volume to increase their commission fees. 6 Chen and Jiang (2006) find that analysts overweight private information when issuing forecasts that are more favorable than the consensus. These reported deviations from efficient weighting corresponds to 6 See Chen and Jiang (2006) and Groysberg et al. (2008). 14

15 related cost/benefit considerations, which in turn suggest that incentives, rather than cognitive biases, play a prominent role in issuing stock recommendations. Contrary to Chen and Jiang, Maines and Hand (1996) find that psychological biases may be responsible for market and analyst inefficiency with respect to earnings news. The results of Loffler (1998) are somewhere in the middle. He finds evidence that psychological biases related to underreaction and overconfidence explain the empirical evidence of inefficiency better than rational, game-theoretic models. These inefficiencies, however, do not seem to have important economic consequences. Friesen and Weller (2006) find evidence of overconfidence and cognitive dissonance of individual analysts. Thereby suggesting that behavioral biases play a role in the decision making process of financial analysts. Finally, the results of Bosquet et al. (2009) show a consistent overweighting of private information. The authors also find, due to possible conflicts of interest, a trade-off between management pleasing behavior (or compensational benefits) and forecast accuracy. Since benefits outweigh costs, financial analysts are willing to commit to pleasing management. 2.5 Market Return and Meeting or Beating Expectations According to Bosquet et al. (2009) management is nowadays for a large part compensated with stock options. This type of compensation induces an increased interest in stock prices and accordingly management prefers beatable targets before an earnings announcement. Meeting or beating earnings expectations (MBE) has as sizable effect on stock prices due to the asymmetric response to earnings surprises. The results of Skinner and Sloan (2002) clearly indicate that the average response to negative earnings surprises is significantly larger than the average response to positive surprises. 7 Of course firm management is aware of a possible large downward stock price adjustment following a negative earnings announcement. It is therefore not surprising that management has incentives to manage analysts expectations and/or reported earnings (Skinner and Sloan, 2002). In this sense it can be very beneficial for firms to smooth their earnings. Hereby making earnings forecasts by financial analysts more easily predictable and avoiding these unwanted negative earnings surprises. Bartov et al. (2002) find that there exists a market premium for firms meeting or beating quarterly earnings expectations. These firms enjoy an average quarterly return that is almost 3% higher than firms that fail to do so (Bartov et al., 2002). 7 Bosquet et al. (2009) also mention this outcome in their paper. 15

16 Both Lopez and Rees (2001) and Kasznik and McNichols (2002) also find premiums for firms that consistently meet earnings expectations. According to Richardson et al. (2004): firms play an earningsguidance game where analysts make optimistic forecasts at the start of the year and then walk down their estimates to a level the firm can beat by the end of the year. The authors also find that this behavior of walking down estimates is associated with managerial incentives to sell stock after earnings announcements. The results of Brown and Caylor (2005) show that managers focus shifted more towards meeting earnings expectations in the mid-1990s, since rewards became more noticeable. Bosquet et al. (2009) find similar results as Richardson et al. (2004). They examined a large U.S. data sample with quarterly earnings forecasts which cover the period 1996 until 2006 and found empirical evidence that: financial analysts strategically inflate their initial forecast at the beginning of the forecasting period, but deflate their forecast in their final revision near the end of the forecasting period in order to please management. Mirciov (2008) examined firms that closely meet earnings forecasts. He finds that firms have better future returns as they closely match the last available mean analyst forecast. Although the evidence is clear about the existence of a market premium for meeting or beating earnings forecasts made by financial analysts, market efficiency theory denies the profitability of a recommendation trading strategy and/or trading on earnings announcements. According to this theory, the market should recognize it when earnings forecasts are deflated. It is therefore expected that the market should not respond heavily on the announcement date with respect to the earnings surprise that is created by the deflation process of financial analysts. If there exists an abnormal return on the days surrounding the announcement date, this could indicate that the market does not fully incorporate (or take notice of) the strategic deflation behavior of financial analysts. The next chapter will elaborate on the method used to examine the U.S. quarterly earnings data and the event study methodology that will be used to examine the announcement effects on the U.S. stock market. The results of this analysis will show whether there is a possibility to gain an abnormal return in the period surrounding the actual earnings announcement. 16

17 Chapter 3 Research Method 3.1 Introduction This chapter will explain the research method used to investigate the abnormal stock market returns surrounding the actual earnings announcement date of the firm. Furthermore, this chapter describes how the relationship between these abnormal returns and the walk down of earnings estimates is tested. The next paragraph will discuss the methodology and the hypotheses used. In the third paragraph the control variables will be described. The final paragraph will discuss the sample selection and data collection. 3.2 Methodology The empirical research in this master thesis will start with an event study. In conducting an event study the choice of a benchmark model is a critical step for examining abnormal stock market returns. According to De Jong en de Goeij (2009) abnormal returns (AR) are defined as the return (R) minus a benchmark or normal return (NR) for firm i on time t: AR it = R it - NR it (1) There are several models available that can be used as a benchmark model. In this thesis the betaadjusted market model will be used as benchmark. This model accounts for differences in the beta of each stock and accordingly abnormal returns are defined as residuals of the market model 8 : R it = α i + β i R mt + ε it (2) The abnormal returns are defined as residuals (or prediction errors) of the following market model: NR it α i β i R mt (3) 8 The equations are taken from the De Jong en De Goeij (2009) paper. 17

18 In this model α and β are ordinary least squares (OLS) estimates of the regression coefficients. For every firm-quarter observation α and β are estimated. Since it takes too much time to perform almost nine thousand regressions manually, an excel macro 9 is written to obtain the coefficients for the benchmark model. The Center for Research in Security Prices (CRSP) provides index returns for the U.S. stock market; in this research the equally weighted index returns will be used. The estimation of the benchmark return is performed over an estimation period [T 1, T 2 ], which precedes the event period [t 1, t 2 ]. The following figure will give a graphical overview of the time line. Figure 1: Time line around an event 10 T 2 of the estimation window is determined by subtracting ten workdays from the end of the quarter date, T 1 is determined by subtracting fifty-five working days from the last day of the quarter. The result is an estimation window over forty-six working days, whereby approximately two weeks at the beginning of the quarter and precisely two weeks at the end of the quarter are excluded from the estimation window. These weeks are excluded because actual earnings announcements from the previous quarter and last revisions for the current quarter are frequently announced in respectively the first and last weeks of a quarter. Although the coefficients for both α and β can be very volatile, the large number of observations will hopefully cancel out these volatility effects. Furthermore, the estimation windows for the different firm-quarter observations do not overlap each other with this short estimation window construction. 9 See Macro 1 on the last two pages of the Appendix for the Visual Basic code for the excel macro. 10 Figure is taken from the De Jong en De Goeij (2009) paper. 18

19 The event window is determined in a similar manner. The beginning of the event window, t 1, is calculated by subtracting ten working days from the event date (the actual EPS announcement). The end of the period, t 2, is determined by adding ten working days to the event date. The result is an event window over four weeks with twenty-one working days. In the days surrounding the actual earnings announcement of the firm, the cumulative abnormal returns (CAR) will be examined. The effect of the explanatory variables will be examined over several CAR: [t 1, t 2 ] intervals. The event window interval is the CAR: [-10, 10] interval, which starts ten working days before the event date and ends ten working days after the event date and comprises 21 working days in total. This interval will capture the cumulative abnormal returns in the whole event window period. To gain more insight about the effects of the actual earnings release, the event window interval is divided in three sub-intervals. In a similar manner the CAR: [-10, -1] interval measures the effects before the announcement date, the CAR: [0, 1] interval measures the effect surrounding the EPS announcement and the CAR: [2, 10] interval measures the effects after the announcement. Besides these four intervals, two other intervals are examined in the event window as well. The CAR: [0] interval is constructed to examine the effects on the event date solely and the CAR: [0, 10] interval measures the effects from the event date until the end of the event window. The [0, 10] interval is considered as the main research interval since the effects of the actual EPS announcement are best captured in this timeframe. Finally, the CAR: [Start Q, 1] interval measures the CAR over an interval that runs from the first day of the quarter till one day after the actual EPS date. The CAR: [Start Q, 1] interval is the only interval which is measured over a variable number of days, since the EPS announcement is made public a variable number of days after the end of the quarter. When conducting the analysis in this thesis, all seven intervals, as described above, are taken into account. Besides a graphical overview of the cumulative average abnormal returns (CAAR) and the average abnormal returns (AAR), statistical tests will determine whether the calculated cumulative abnormal returns (CAR) over the different intervals are significantly different from zero. The hypotheses for this event study are: H 0 : E(CAR i,q ) = 0 Hypothesis 1 H 1 : E(CAR i,q ) 0 Hypothesis 1a 19

20 where i is the firm index and Q is a quarter notation. The null hypothesis states that the cumulative abnormal returns are not significantly different from zero. When the null hypothesis is rejected, the cumulative abnormal returns are significantly different from zero, in line with the alternative hypothesis Simple Benchmark Model In the second part of this empirical research the walk down of earnings estimates is going to be investigated. Analysts make optimistic forecasts at the beginning of a quarter and deflate their forecasts in a final revision near the end of quarter to set beatable earnings targets. The degree of deflation will be used as a proxy for the management pleasing behavior of financial analysts. The difference between the final revision and the initial forecast, labeled Revision, will be measured throughout the sample. Equation (4) describes how the variable Revision is calculated: Revision i,q = Last Revision i,q - First Forecast i,q (4) After the calculation of the variable revision (labeled REV in the regression equations) the following regression equation will be estimated using the ordinary least squares (OLS) method: CAR i,q = α + β 1 REV i,q + ε i,q (5) where i is the firm index and Q is the quarter notation. The dependant variable measures the cumulative abnormal return (CAR). The coefficient estimate of β 1 is the parameter which quantifies the effect of the independent (or explanatory) variable on the CAR. Finally, ε, the error or disturbance term, captures all other factors which influence the dependent variable, other than the independent variable REV. The hypotheses for the regression analysis are: H 0 : β 1 = 0 Hypothesis 2 H 1 : β 1 0 Hypothesis 2a The null hypothesis states that the effect, which the variable Revision has on the CAR, is zero. The alternative hypothesis indicates a significant relation between management pleasing behavior and the 20

21 cumulative abnormal returns surrounding the actual earnings announcement date. If β 1 > 0 this will indicate a positive relation between the variable Revision and the CAR; a change in the variable Revision is associated with a change in the variable CAR in the same direction. If β 1 < 0 there exists a negative relation between the variables Revision and CAR; a decrease in the variable Revision will, in this case, lead to an increase in the cumulative abnormal return. The coefficient of β 1 is expected to be negative and significant. Therefore, negative revision observations are expected to have an increase in the CAR as result and positive revision observations are expected to end up with a decrease in the CAR. The following figure will give a graphical overview of the earnings-guidance game of financial analysts. In this figure two observations with a negative revision value are displayed. It is however possible that the final revision is on a higher level than the initial forecast. The value of the revision will then be positive instead of negative. Figure 2: Time Line for Forecasts and Event Date As can be seen in Figure 2, the initial forecast in quarter N is higher than the initial forecast in quarter N+1. Furthermore, the final revision is on a lower EPS level for quarter N. The magnitude of deflation, or 21

22 the degree of management pleasing behavior, is much higher in the first quarter since the difference between the final revision and the initial forecast is much larger. The impact of these differences will be tested in order to determine their significance in conjunction with the abnormal returns surrounding the event date Extended Model After testing whether or not the variable Revision has a significant impact on the CAR, an extended model is used to further investigate the dynamics of the management pleasing behavior. Whereas Bartov et al. (2002) examined firm-quarter observations spanning the period from January 1983 to December 1997 and focused mainly on the forecast error, this paper examines a more recent data sample and focuses on the revision during the quarter. Similar to Bartov et al. (2002), several explanatory variables are added to the simple regression model 11. The regression equation of this extended model will be: CAR i,q = α + β 1 REV i,q + β 2 SURP i,q + β 3 ACTUAL i,q + β 4 DMBE i,q + (6) β 5 DBEAT i,q + β 6 DBEAT SURP i,q + ε i,q where i is the firm index and Q is the quarter notation. The revision, REV, measures the difference between the last revision and the initial forecast. SURP is the earnings surprise and is calculated as the difference between the actual EPS and the last revision. ACTUAL is the actual EPS. DMBE and DBEAT are dummy variables that receive the value of 1 if, respectively, SURP 0 and SURP > 0. Otherwise, these two dummy variables receive the value of zero. For more information about the definitions of the variables see also Figure 6 in the Appendix. Again, the coefficient estimate of β 1 will be examined with the use of regression results. The hypotheses for this regression are similar to those of the simple benchmark model. Although several explanatory variables are added to the regression equation, the coefficient of β 1 is still expected to be negative and significant. Furthermore β 2 is expected to be positive, since a large positive surprise will lead to a higher CAR according to the literature. 11 Bartov et al. (2002) use for instance the following regression equation: CAR i,q = β 0 + β 1 DMBE i,q + β 2 DBEAT i,q + β 3 ERROR i,q + β 4 SURP i,q + β 5 DBEAT SURP i,q + ε i,q 22

23 3.3 Control Variables Since there are many variables that affect the cumulative abnormal return, it is important to include these explanatory variables in the regression analysis. This paragraph discusses the most important control variables used in the literature and the expected relations between these independent variables and the dependent variable. Firm size According to Richardson et al. (2001) past studies have indicated that large firms have less optimistic forecasts. This could lead to a lower degree of deflation and therefore could also have an impact on the CAR. The size of the firm will be measured by the logarithm of the market capitalization, the share price times the number of shares outstanding. The relation between firm size and CAR is expected to be negative since a lower magnitude of deflation is likely to cause a lower CAR. Growth Brown (2001) finds that reported earnings are managed by managers from growth firms to avoid negative earnings surprises. Matsumoto (1998) also reports that growth firms manage earnings expectations to avoid these negative earnings surprises. To capture these effects the book-to-market ratio will be added as independent variable as a measure for growth- or glamour stocks. A positive relation is expected to exist between growth and CAR due to the manipulation of expected- as well as reported earnings. Profitability Burgstahler and Eames (1999) find that the forecast bias is also related to whether firms make profits or losses (Richardson et al., 2001). Richardson et al. (2001) also mention that analysts turn out to be pessimistic for firms reporting profits. Therefore profitability is also included in the analysis. Two variables are added to measure profitability. The first one is the return on equity (ROE), calculated by dividing a company's net income by its total shareholders equity. A second measure of profitability, the price/earnings ratio, is also added to the model and calculated as the market value per share divided by the earnings per share (EPS). The expected relation between the profitability and the CAR is ambiguous. 23

24 3.3.1 Extended Model with Control Variables With the inclusion of the control variables mentioned above the regression equation will be: CAR i,q = α + β 1 REV i,q + β 2 SURP i,q + β 3 ACTUAL i,q + β 4 DMBE i,q + (7) β 5 DBEAT i,q + β 6 DBEAT SURP i,q + β 7 LOG_MV i,q + β 8 B/M i,q + β 9 ROE i,q + β 10 P/E i,q + ε i,q where i is the firm index and Q is the quarter notation. The first six variables are as defined in equation (6), the last four variables are calculated as described above. As with the two previous models, the coefficient estimate of β 1 will be examined. The hypotheses for this regression are similar to those of the simple benchmark model and the coefficient of β 1 is still expected to be negative and significant. Furthermore the expected relations between the control variables and the cumulative abnormal returns will be examined. In addition to the three regressions, two tables are constructed. The first table examines how the CAR varies due to changes in the magnitude and sign of the variable Revision. The second table is constructed to provide more insights about the sign of the earnings surprise, controlled for the sign of the variable Revision. 3.4 Sample and Data This thesis uses approximately the same dataset as the Bosquet et al. (2009) paper. The quarterly earnings forecasts are obtained from the Institutional Broker Estimate System (I/B/E/S) database and cover the period 1996 until the end of The I/B/E/S database is restricted to highly covered U.S. companies with a fiscal year end in December (Bosquet et al., 2009). The data is stripped from errors and potential companies in difficulties are also erased. Since the magnitude of revision is examined in this thesis, it is necessary to compare first forecasts with last revisions. Therefore two datasets are created. The first dataset contains only first forecasts and the second dataset contains only last revisions. 24

25 The initial First Forecast sample contains quarterly earnings forecasts issued on companies. The initial Last Revision sample contains earnings revisions on 837 companies. Since McNichols and O Brian (1997) find results that indicate that analysts drop stocks with unfavorable future prospects and to ensure a sample of forecasts with maximum coverage and processing of information, negative forecast are dropped (Bosquet et al., 2009). By deleting the negative forecasts, the First Forecast sample is reduced by 11%. The final First Forecast sample contains quarterly earnings forecasts issued on companies. Similarly, the Last Revision sample is reduced by 12%. The final Last Revision sample contains earnings revisions on 792 companies. After matching both samples by company ticker and taking averages of multiple forecast observations 12 for the same firm-quarter, a combined sample with average initial forecasts as well as average last revisions is created. This combined sample contains firm-quarter observations issued on 767 companies. Daily stock price information and the equally weighted index returns are obtained from The Center for Research in Security Prices (CRSP). This database maintains a comprehensive collection of security price, return, and volume data for the NYSE, AMEX and NASDAQ stock markets. Daily stock price information is necessary to compute the estimates of both α and β in the benchmark model. Since this information is not available for all companies, the combined sample with firmquarter observations issued on 767 companies is reduced to a final sample of firm-quarter observation issued on 568 companies. With the inclusion of control variables additional data is needed. Data on market capitalization, the book-to-market ratio, the return on equity and the price/earnings ratio, to control for firm effects is downloaded from Datastream Multiple analysts can make earnings estimates for the same firm for the same quarter. These estimates are averaged to end up with one average observation, for the initial forecast as well as the last revision, per firm, per quarter. 13 For Thomson Reuters Datastream, see 25

26 Chapter 4 Data Descriptives 4.1 Introduction This chapter gives an overview of the data descriptives. In the next paragraph the descriptive statistics and a correlation matrix will be presented. In the third paragraph the average forecast fault over time will be examined for both the First Forecast sample as well as the Last Revision sample. 4.2 Descriptives The descriptive statistics, see Table 3 in the Appendix, provide some interesting results. The mean revision is a negative number. This indicates that the last revision is on average on a lower level than the first forecast and that financial analyst deflate their earnings forecasts on average. Also, the average surprise has a positive value. Since the surprise is calculated as the actual EPS minus the last revision, this result indicate that the last revision is on average on a lower level than the actual EPS. These results suggest that financial analysts manage earnings forecasts to create a positive earnings surprise, or at least try to avoid an unwanted negative earnings surprise. As can be seen in Table 3, not for all firm-quarter observations data for the control variables was available. The number of observations, indicated with an N, is therefore lower than the total number of observations for all four control variables. The correlation matrix, which can be found in Table 4 in the Appendix, shows very low correlations between the variables overall. The higher correlation values can easily be explained. The highest value of 0,84 is the correlation between the variables DMBE and DBEAT. This is not surprising since both variables are derived from the variable Surprise. Also the correlation between the variables SURP and DBEAT SURP, which has a value of 0,77, is high due to the common variable SURP. All other correlation values are under the value of 0,5. 26