VIEWS: 27 PAGES: 67 POSTED ON: 7/21/2012 Public Domain
Can investors benefit from the predictability of EPS surprises? Master thesis in Finance Author Thesis supervisor Thomas Beckers Mr. Leonard Wolk ABSTRACT In this thesis, I show that, among the S&P500 firms, the EPS surprise and its sign (MBE) are predictable to some extent. Nevertheless, the application of quintile strategies based on the EPS surprise predictability does not lead to higher returns around the earnings announcement, suggesting that stock investors properly integrate the driving forces behind the surprises in their behavior and do not blindly follow analysts’ expectations. Based on this observation, I discuss a new framework to explain the market reaction to the earnings release. My results indeed suggest that the market formulates its own expected surprise and reacts accordingly. 1 Table of contents 1. Introduction 3 2. Literature review 5 2.1 Predicting the future performance 5 2.2. Stock price reaction drivers and the threshold mentality 7 2.3. Managerial incentives to meet EPS estimates 10 2.4. Analyst incentives 11 2.5 The earnings game: earnings and expectations management 13 2.6. Predicting MBE 15 2.7 Predicting the earnings surprise 18 3. Research design 19 3.1. Dataset 19 3.2 Data collection 21 3.3 Method 23 4. Results 33 4.1 Descriptive statistics 33 4.2 Step A 39 4.3 Step B 48 4.4 Step C 51 4.5 Discussion: what really drives the market reaction? 53 5. Conclusion 56 6. References 58 7. Appendix 62 2 1. Introduction The earnings announcement season is one of the most exciting periods for stock investors, as it gives an update about the performance of the company, which is widely used for firm valuation (Penman and Sougiannis, 1998 and Bartov et al., 2002). These earnings announcements also usually trigger large short-term price reactions, as the one-day 14% price decline in Amazon’s stock price on the 26th of October 2011 after the Q3 earnings release illustrates. The earnings disclosure is therefore highly relevant for market participants, both for the short-term and the long-term. Over the past decades, the determinants of the returns around the earnings announcement have evolved (Brown and Caylor, 2005): stock investors as well as academics shifted their focus from the growth in earnings (e.g. Ou and Penman, 1989) towards analysts’ expectations and the earnings surprise (e.g. Bartov et al., 2002). Although we would expect this shift to result in randomness in the earnings outcomes and the resulting market reactions, former academic studies suggest that the market reaction and the underlying EPS (earnings per share) surprise can be predicted. Among others, Matsumoto (2002) and Bauman and Shaw (2006) show that several factors (firm characteristics, analyst opinions) significantly influence the probability that the surprise is positive. Other papers such as Lim (2001) find a strong autocorrelation among the EPS surprises (actual EPS-analysts’ expected EPS), whereas Bernard and Thomas (1990) find that this autocorrelation can give rise to profitable strategies. In view of these results, I provide an analysis on the predictability of the EPS surprise and its sign (MBE) and its consequences for firm valuation. To achieve this goal, I use a very recent sample of more than 5,000 observations in the US. My results confirm the shift in focus toward analysts’ expectations. The other benchmarks proposed by Degeorge at al. (1999) which are a positive earnings and positive growth in earnings are not significantly affecting the abnormal return around the earnings release while the sign of the surprise (MBE) and the extent of the surprise are the key drivers of the stock reaction, as in Bartov et al. (2002). Besides, I also find that the sign and the extent of the earnings surprise are predictable, in line with the prior literature. Given that MBE and the EPS surprise positively affect the stock return around the earnings announcement (Bartov et al., 2002), I develop two models that predict their values. First, I elaborate on the profitability of the quintile trading strategy based on the predictability of MBE (meeting or beating the expectations) which has been shown to engender a “premium” 3 price reaction to the earnings announcement (Bartov et al, 2002). Second, I investigate whether the earnings surprise, which is nowadays the main driver of the stock price reaction (Brown and Caylor, 2005; Bartov et al., 2002) is predictable, and if so, if the quintile trading strategy solely based on this predicted earnings surprise is profitable. Third, I test if the predicted probability of MBE and the predicted earnings surprises obtained from the first two discussions can be used to predict the market reaction. Finally, I conduct a study similar to that of Bernard and Thomas (1990) to investigate whether abnormal returns can be gained from the study of the past surprises. Concretely, my thesis will attempt to answer the question “Can investors benefit from the predictability of EPS surprises?” In particular, I will focus on the predictability of the sign and the size of the earnings surprise by answering the subquestions “What are the drivers of MBE? (1a)” and “What are the drivers of the earnings surprises? (2a)”. I will also investigate whether the market reaction can be explained by the results gotten in subsections (1a) and (2a) to answer subquestion (3a): “Can the predicted surprise and the predicted probability of MBE explain the market reaction to earnings announcements?” Finally, I will test if “the past surprises can predict the market reaction to earnings announcements?” which is subquestion (3a’). Once I obtain an answer to these four subquestions, I examine whether a zero-cost strategy based on quintiles are profitable. Specifically, I will provide an answer to “(1b) Are the trading strategies based on the predicted probability of MBE quintiles profitable?”, “(2b) Are the trading strategies based on the predicted earnings surprise quintiles profitable?”, “(3b) Are the trading strategies based on the predicted market reaction quintiles (from 3a) profitable?” and finally “(3b’) Are the trading strategies based on the predicted market reaction quintiles (from 3a’) profitable?”. The results from all these subquestions will help me determine if indeed one can earn abnormal returns by recognizing that the EPS surprises are predictable. The discussion and the results of this thesis could be useful for the market participants. First, my thesis will discuss the relevance of the several earnings benchmarks which have been considered in the prior literature. Given that I use a sample over the time window 2009-2011, this thesis provides an update on this topic. Second, the study may reveal that there is autocorrelation between surprises (as in Lim, 2001) and analysts could learn from the results and take correcting actions, on the condition, of course, that forecast accuracy is their primary 4 goal. Some studies (e.g. Ackert and Athanassakos, 1997) suggest that this may not be the case, and this topic will be discussed later on in this thesis. Third, my thesis reviews the motivations of managers and analysts to engage in the earnings game whose result is the earnings surprise. Moreover, I provide empirical evidence on these driving forces, which “modernizes” the prior findings and I propose a new framework to explain the returns around the earnings announcements. Finally, the results of the trading strategies, and particularly the volatilities of the predicted MBE probability quintiles, may be of interest for stockholders. 2. Literature review This literature review is constructed as follows. First, I will introduce the topic of this thesis in section 2.1 by reviewing some papers whose goal was to forecast the future performance of stocks, and more specifically the returns around earnings announcements. Section 2.2 will go a step further and discuss the primary drivers of the stock price reaction to earnings announcements (the main one being the earnings surprise) and the importance of meeting the thresholds and especially the expectations benchmarks (MBE). As the probability of meeting the thresholds is influenced by the earnings game taking place between managers and analysts, the motivations of managers and analysts to engage in this earnings game are discussed in section 2.3 and 2.4 respectively. Section 2.5 covers the manipulation methods in this earnings game. This earnings game ends when the company reports its earnings figure, which may beat or meet the expectations (MBE). In section 2.6, I discuss the literature about the drivers of MBE and lastly discuss the drivers of the earnings surprise in section 2.7. 2.1 Predicting the future performance Several methods to predict the future performance of a company exist, but the belief that the market is hardly beatable tends to be as widely accepted as the GAAP rules in accounting. In fact, several models such as the CAPM or the Fama-French three-factor model have helped explain a substantial part of the variations in asset returns, but these models can only be used to explain the past variations, and not predict the future. Nevertheless, the results of Fama and French (1992) suggest that holding portfolios of firms that have a small market capitalization and a low market-to-book ratio would eventually produce abnormal returns if one only considers the systematic risk (captured by the beta of the stock). One of the reasons behind these apparent market anomalies is the asymmetric reaction of the market to earnings 5 announcements, which makes these releases of particular importance. Indeed, a study by Skinner and Sloan (2002) highlighted the fact that the premium earned on low market-to-book firms is primarily due to the overoptimistic expectations of future profitability for growth firms. As the growth firms eventually cannot keep up with the expectations, they end up missing the expectations and the market severely sanctions them after the earnings release. After accounting for this severe sanction on growth stocks that is concentrated around the earnings announcement, Skinner and Sloan (2002) find no significant difference between the returns of growth and value stocks. This central fact highlights the importance of earnings announcements and their components: the actual earnings and the expectations. The focus on the actual earnings seems surprising at first; the standard finance literature recognizes the future discounted cash flows (DCF) as the only correct method to value a firm. Nonetheless, Penman and Sougiannis (1998) found that the DCF method produces large valuation errors. In their study, they first assumed market efficiency, and took the share price as the true value of the firms. Then, they inferred the companies’ value by applying three valuation models: the discounted dividend model, the discounted cash flow model and the residual income method (RIM). They arrive at a conclusion that is consistent with the markets’ focus on earnings: the residual income method exhibits by far the lowest valuation errors. The other constituent of the earnings announcement is the earnings expectations. These expectations can be formed in several ways. First, the investors can form expectations by simply inferring expected future earnings from the past earnings while accounting for recent earnings growth; these are univariate time series forecasts (Foster and Olsen, 1984). Though this method is rather simplistic, its ease of application makes it very attractive. Second, the investors can forecast future performance by accounting for both the past performance and elements in the balance sheet such as the change in inventory, gross margin and accounts receivable (Lev and Thiagarajan, 1993; Ou and Penman, 1989; Abarbanell and Bushee, 1997). This method seems superior because this combination of momentum and fundamental analysis has produced superior forecasts that could be used to form profitable strategies (Ou and Penman, 1989). The third method, and the most widely used nowadays (Brown and Caylor, 2005) is the reliance on analysts’ predictions. In fact, an early comparison of the several expectations methods by Brown and Rozeff (1978) revealed that analyst forecasts were indeed the most accurate prediction. Thus, if we assume that investors have rational expectations, they should rely on the (more accurate) analyst forecasts and the firm valuation 6 should consequently depend on these forecasts. Indeed, Bartov, Givoly and Hayn (2002) find that the firms which meet or beat the forecasts (MBE) experience a premium return on their shares. This focus on the analysts’ forecasts may create new trading opportunities, which is the centerpiece of this thesis. These opportunities could arise for the reason that forecast errors (the actual EPS – analyst forecast) are to a certain extent predictable and exhibit autocorrelation (e.g. Lim, 2001; Bernard and Thomas, 1990). The following sections will guide the reader step by step through the components of the earnings announcement and the driving forces that govern them. 2.2. Stock price reaction drivers and the threshold mentality 2.2.1 Performance measure and prospect theory The step in a study of the reactions to earnings announcements is to identify the leading drivers of these reactions. A first idea can be retrieved from the paper of Graham, Harvey and Rajgopal (2005), who have conducted a survey in which they asked the managers to identify the benchmarks, trends and performance measures they perceived as important. Earnings were perceived as the most important reported financial performance measure, far ahead of revenues, cash flows from operations and free cash flow (Graham et al., 2005). This focus, as I already mentioned, is in complete contradiction with the standard financial theory which posits that the managers should have free cash flows as their only focus; but it is consistent with the empirical findings of Penman and Sougiannis (1998). Beside the identification of earnings as the foremost performance measure, Graham, Harvey and Rajgopal (2005) have studied the earnings benchmarks managers pay attention to. Their paper reveals that the most important benchmarks are respectively the earnings per share (EPS) of the same quarter of last year, analysts’ estimated EPS, a zero earnings and the previous quarter’s EPS. In a widely-cited study of these benchmarks, Degeorge, Patel and Zeckhauser (1999) developed the notion of “threshold mentality” to characterize the market’s reaction to earnings announcements, which itself explains managers’ motivation to meet the thresholds. In this paper, they identify the threshold of zero surprise, defined as the actual EPS minus the EPS benchmark (or investors’ EPS expectations), as the focal point of investors’ attention. This focal point can be assimilated as the reference point in the prospect theory developed by 7 Kahneman and Tversky (1979): positive surprises tend to be followed by positive reactions and negative surprises tend to be followed by negative reactions. Several papers have suggested a more rigorous prospect theoretic approach to describe the market reaction to earnings announcements, beyond the simple identification of the reference point. Freeman and Tse (1992) and Kinney, Burgstahler and Martin (2002) provide an analysis of the earnings response coefficient (ERC), which is estimated as the slope coefficient of the earnings surprise in a regression of the CAR on the earnings surprise, and observe that the ERC is not constant over the sample’s earnings surprise range. Instead, they suggest that the stock price reaction to surprises is akin to the S-shaped value function proposed by Tversky and Kahneman (1979), with the reference point as the surprise of 0, a concave function for positive surprises, a convex function for negative surprises and a steeper slope in the negative earnings surprise region. Clearly, the benchmarks matter. But which one matters most? Zero earnings, zero change in earnings or zero earnings surprise? 2.2.2 The right threshold Degeorge, Patel and Zeckhauser (1999) have proposed that the thresholds are hierarchically ordered; first, a company wants to report positive profits, then the company is willing to report profits larger than or equal to the profits of the same quarter of last year, and finally report a figure that meets or exceeds analysts’ expectations. Meeting analysts’ expectations would therefore, according to Degeorge et al. (1999), be less important than the two other thresholds aforementioned. Additional evidence on the relative importance of the thresholds is provided by Dechow, Richardson and Tuna (2003). It indicates a shift in focus around the year 1998: until 1998, managers placed most importance on achieving positive annual profits, whereas after 1998 the managers shifted their focus on meeting analysts’ expectations. A more recent paper by Brown and Caylor (2005) reaches similar conclusions: as of the mid-1990s, managers placed more importance on meeting analysts’ estimates than reaching a positive profit or reporting a higher-than-last-year profit. Brown and Caylor (2005) claim that the shift in managers’ preference is a natural response to the shift in investors’ focus from the absolute earnings figure (report positive actual profit) to the relative earnings figure (reported earnings relative to the analysts’ expectations). In fact, investors have probably put more emphasis –and thus changed the firm valuation dynamics– on analysts’ forecasts because of the increasing 8 number of firms followed by analysts as well as the increasing accuracy and the growing media consideration of their forecasts (Brown and Caylor, 2005). Bartov et al. (2002) find a significantly positive relationship between a stock’s cumulative abnormal returns and the price-scaled earnings surprise, lending support to the heightened importance of the analysts’ expectations benchmark. 2.2.3 Importance of meeting the analysts’ expected EPS benchmark Empirical studies on the forecast errors (actual EPS – Analysts’ EPS estimate) also provide support for the importance of meeting the threshold (e.g., Degeorge et al, 1999; Burgstahler and Dichev, 1997; Burgstahler and Eames, 2003). In fact, whereas one would expect a normal distribution of the errors, with mean and skewness equal to zero, academics observe an abnormally high number of slightly positive errors, an abnormally low number of slightly negative errors and a fat tail in the left part (i.e. negative errors region). The forecast error distribution of my sample, which shares these characteristics, is exhibited in figure 1. Figure 1: Histogram of the earnings surprises (or forecast errors), defined as the actual Earnings per share (EPS) minus the average of the last available analyst EPS forecasts. The vertical line represents a surprise of zero. To explain this surprising error distribution, some authors have suggested that managers engage in the earnings game, in which they influence analysts’ forecasts and the reported EPS figure in order to meet or beat the expectations (MBE). This motivation to MBE can easily be explained by the fact that MBE positively influences firm valuation (Bartov et al., 2002): 9 according to Bartov et al. (2002), over a period starting on the date of the first forecast after the previous quarterly earnings announcements and ending one day after the earnings announcement, there exists a positive abnormal return for those companies that “meet or beat expectations” (MBE), even after controlling for the earnings surprise (which also positively affects the abnormal return). Given that managers’ wealth or well-being habitually depends on the firm valuation (e.g. through stock options or human capital), one can wisely expect that they will be motivated to accomplish the actions that raise firm valuation, including meeting or beating analysts’ forecasts (Brown and Caylor, 2005). Similarly, the study of Graham, Harvey and Rajgopal (2005) indicates that a large majority of managers want to meet the benchmarks because doing so “builds credibility with the capital market and maintains the stock price”. This “valuation motivation” likely explains the abnormal distribution of forecast errors. This section discussed the relatively recent focus on the earnings surprise –henceforth defined as the actual earnings minus the analysts’ expected earnings– and the importance of meeting or beating analysts’ expectations (MBE) for equity valuation. The next section analyzes in more detail the motivations for managers to beat the estimates. 2.3. Managerial incentives to meet EPS estimates Although Brown and Caylor (2005) and Graham, Harvey and Rajgopal (2005) suggest that the managers strive to meet the benchmarks for personal economic reasons, they don’t provide empirical tests on this issue. Further research on the topic has found strong empirical evidence for two important incentives to meet the expected EPS benchmark: job safety and bonuses. First, it appears that missing the EPS analysts’ forecast may lead to a higher CEO turnover (Puffer and Weintrop, 1991). In their study, Puffer and Weintrop (1991) analyzed the relationship between the board’s EPS surprise –actual performance minus the board’s expectations of performance– and the amount of CEO turnover in a principal-agent framework. As the board’s expectations can hardly be measured for all firms, the authors used analysts’ EPS expectations as a proxy for board expectations. Embedded in a logistic model with other performance measures and control variables, the EPS surprise was found to be significantly and negatively related to CEO turnover (Puffer and Weintrop, 1991). Larger earnings surprises are thus associated with more job safety, and lower EPS surprises are 10 associated with more CEO turnover. This paper provides evidence that one of the motivations for managers to meet analysts’ forecasts could be job security. Second, there is empirical evidence that the managers meet analysts’ EPS forecasts in order to receive ex post pecuniary benefits. A paper by McVay, Nagar and Wei Tang (2006) finds a positive relationship between the likelihood of just meeting or beating the expectations (JMBE) and the subsequent managerial stock sales, suggesting that managers manipulate earnings to heighten the stock price and subsequently sell their shares. Missing quarterly earnings benchmarks has also been shown to adversely affect the subsequent CEO bonuses by Matsunaga and Park (2001). In this paper, they find that companies that missed analysts’ consensus pay a lower annual bonus to the CEO. Missing the benchmarks for several consecutive periods has an even further negative impact on the CEO bonus (Matsunaga and Park, 2001). Visibly, the earnings announcements have an impact on managers’ welfare. This impact incentivizes the managers to meet or beat the expectations (MBE), which enhances both the value of their stake in the company and their job security. This incentive also varies with the firms’ characteristics and is the topic of section 2.6. The following section 2.4 discusses the incentives of analysts in the earnings game. 2.4. Analyst incentives 2.4.1 Conflict of interest and the rational expectations hypothesis The forecasting of earnings by analysts is subject to an important conflict of interest. Indeed, analysts providing stock coverage and earnings estimates may receive payments for this coverage or have banking relationships with the companies they assess, giving rise to a dilemma: sustain the relationship by pleasing the firms’ managers (by exhibiting positive opinions and/or exhibiting pessimism for the short-term EPS forecast) or improve one’s credibility by providing accurate estimates. Dugar and Nathan (1995) confirm this concept of conflict of interest in their study of analyst optimism. They show that analysts from brokerage firms (that also provide other financial services) are more optimistic than other analysts who don’t have banking relationships with the client. In the same fashion, O’brien, Mcnichols and Lin (2005) find that banking ties indeed influence analysts’ behavior. More specifically, they find that affiliated analysts are much slower than nonaffiliated analysts to downgrade their 11 recommendation from Buy to Hold, whereas they are more rapid in upgrading their recommendation from Hold to Buy, suggesting that affiliated analysts have a tendency to negate bad news and overreact to good news about the client company. These studies suggest that analysts’ estimates and recommendations are subject to a bias that originates from the banking relationship between the analysts and the firms. This bias existence can be examined through the test of the rational expectations hypothesis. This hypothesis states that the forecast errors (1) should have a zero mean, given the available dataset and (2) should be uncorrelated to any variable (Ackert and Hunter, 1995). While early literature did not find support for this hypothesis, Ackert and Hunter (1994) propose a dynamic form of the hypothesis. In fact, they also reject the hypothesis for far-away forecasts (analysts are overoptimistic in the long-run), but argue that the analysts revise their estimates downward as time passes to end up with presumably unbiased forecasts. Hence, when the forecast error is calculated as a function of the most recent forecast, the rational expectation hypothesis is no longer rejected. All in all, it appears that (1) analysts are subject to a conflict of interest that influences the forecast bias, (2) the short-term forecasts are probably rational and (3) the long-term forecasts are overoptimistic. This overoptimism is discussed in subsection 2.4.2. 2.4.2 Overoptimism of analysts’ forecasts In order to explain the overoptimism of analysts, Ackert and Athanassakos (1997) have developed the hypothesis that overoptimism (hence, bias) is related to the uncertainty about the future earnings. In line with O’brien, Mcnichols and Lin (2005), Ackert and Athanassakos (1997) suggest that banking ties with companies lead analysts to be optimistic on average. Nevertheless, analysts must weigh this relationship with their clients against their reputation, especially when the uncertainty about the firm’s future prospects is low. Indeed, being overly optimistic about the firm’s prospects when the uncertainty is high does not hurt the analyst’s image as much as if the uncertainty is low, or equivalently, the reputation cost of forecast bias is lower for high-uncertainty than for low-uncertainty firms (Ackert and Athanassakos, 1997). In their analysis, Ackert and Athanassakos (1997) use the standard deviation of the estimations as a proxy for the uncertainty and their results actually support the view that analysts are less concerned with their reputation (i.e. they issue more optimistic forecasts) when the uncertainty about the firm, as measured by the dispersion of analysts’ forecasts, is 12 high. Das, Levine and Sivaramakrishnan (1998) obtained similar results, but provided another rationale for the analysts’ overoptimism: analysts are aware that the managers want positive opinions and that there is a tendency among corporations to block the transfer of information to those analysts that are not complacent. Given that this blocking of information transfer hurts analysts most in the case of high uncertainty (i.e. when the need for private information is more serious), the analysts will be more optimistic for high-uncertainty firms so that they keep access to private information. 2.5 The earnings game: earnings and expectations management Now that we have reviewed the incentives of the two players of the earnings game (managers and analysts) in sections 2.3 and 2.4, it is now time to analyze the two earnings manipulation methods and their consequences on the market reaction to earnings announcements. There are two ways a manager can influence the EPS surprise: earnings management and expectations management. According to Bartov, Givoly and Hayn (2002), earnings management “generally involves using accrual accounting in order to produce earnings that surpass the forecasted earnings target”. In that case, the manager influences the EPS surprise by manipulating accounting data within the limits of the GAAP rules. In contrast, “Expectations management takes place whenever management purposefully dampens analysts’ earnings forecasts to produce a positive earnings surprise upon the earnings release” (Bartov et al., 2002). Bugstahler and Dichev (1997) delivered evidence of earnings management to avoid earnings decreases and losses. In their paper, they claim that the major part of the manipulation will occur around the slightly negative earnings surprises in order to turn them into slightly positive surprises, as the marginal benefit of the manipulation is highest around this zero surprise area. This explanation for earnings management around the zero surprise is consistent with the value function from the prospect theory of Tversky and Kahneman (1979), which is steepest around the reference point. According to Burgstahler and Dichev (1997), the earnings manipulation may be conducted through the timing of sales and be reflected in changes in working capital on the balance sheet. This manipulation is consequently easier for companies that have many current assets and liabilities, as the changes become more difficult to detect. As an example of sales timing, a company may report future sales in the current earnings 13 release, treat them as receivables and hide them in the large balance sheet. Athanasakou, Strong and Walker (2009) likewise observe a tendency to use current assets as a way to meet or beat the expectations in the UK. In supplement to the timing of sales, the share repurchase programs may also be an earnings management tool according to Hribar, Jenkins and Johnson (2006). Indeed, share repurchases mechanically reduce the denominator of the earnings per share figure and can be used in some instances to shift from a negative surprise to a positive one. Given that there is evidence of earnings manipulation, we should expect the analysts to incorporate it into their forecasts. And it may actually be the case: Burgtashler and Eames (2003) find that analysts anticipate earnings management by the firms but unfortunately are unable to accurately predict which ones will manage earnings. As a result, the motivation for managers to manage earnings persists. Expectations management is another means that managers can use to achieve positive surprises. Matsumoto (2002) argues that driving down the analysts’ forecasts entails some costs, in the form of a negative stock price reaction on the date of the forecast revision. The managers thus should only use expectations management if the benefit from a positive earnings surprise (i.e. the MBE premium) is larger than the cost of the downward revision (Matsumoto, 2002). And this seems to be the case. Indeed, Chan, Karceski and Lakonishok (2003) discover that the positive reaction from the surprise outweigh the negative impact of the preceding downward revisions. Kasznik and Lev (1995) also find that prior to the earnings announcement, the proportion of bad news warnings from the management is much larger than the proportion of good news warnings, suggesting that managers frequently use downward expectations management to meet the analysts’ forecast on the earnings announcement date. As the benefits of forecast guidance outweigh its costs and the use of this guidance has been frequently observed in the past (Kasznik and Lev, 1995), it should be expected to observe many instances of expectations management in an attempt to meet or beat the expectations (MBE) in recent samples. Indeed, Matsumoto (2002) finds a significant relationship between a binary variable representing forecast guidance and the probability that a company meets or beats the expectations (MBE). Athanasakou, Strong and Walker (2009) also find evidence for this relationship in their UK sample. Given that there is strong evidence that MBE often comes from manipulation and is not the result of superior performance, one would expect that the market does not react positively to 14 MBE when it is likely achieved through any of the two manipulation methods. Bartov et al. (2002) study the relationship between the stock price reaction and MBE when there is suspicion of expectations management. They find that, indeed, the shareholders understand the tendency of managers to manage expectations and react less positively to MBE when expectations management is likely, but the premium to MBE still exists, indicating that investors do not fully incorporate this earnings manipulation into their behavior. They find the same pattern for cases of likely earnings management: investors react less positively when the MBE is likely achieved through earnings management, but the premium to MBE remains positive. As MBE generates a significantly positive premium reaction, even when it is likely achieved through manipulation, it is interesting to know what drives the motivation of the managers to engage in the earnings game and influence its result (i.e. the earnings surprise). Section 2.6 discusses the prior literature about the drivers of MBE. 2.6. Predicting the beating or meeting of analysts’ expectations (MBE) There are many potential drivers for the probability that a firm meets or beats analysts’ estimates. In the three following subsection, I distinguish between three main classes of MBE drivers: benchmark hierarchy, company features and analyst actions. 2.6.1 Hierarchy between benchmarks There is some evidence of a pecking order in meeting earnings benchmarks. Degeorge et al. (1999) suggest that the positive earnings target is more important than the change in earnings from year to year which is itself more important than the analysts’ expectations benchmark. Moreover, Graham et al. (2005) propose that the variation in earnings is more important than meeting analysts’ expectations. One would thus expect to see MBE conditional on the presence of a positive earnings figure and a positive year-to-year earnings change: MBE is more likely to occur when the other benchmarks are met. Although the relative importance of the benchmarks has somewhat changed over the last decade (Dechow et al, 2003; Brown and Caylor, 2005), there probably remains a connection between the benchmarks. Indeed, Matsumoto (2002) finds a significantly positive relationship between MBE and POSUE, a binary variable representing positive earnings changes. In addition, Athanasakou et al. (2009) 15 find a positive relationship between MBE and the positive earnings changes binary variable as well as the positive earnings binary variable. Finally, Bauman and Shaw (2006) observe a negative relationship between a loss binary variable (equal to 1 if the company makes a loss) and the probability of MBE. Taken together, these results suggest that there remains a relationship between the benchmarks. 2.6.2 Company features Besides this hierarchy, MBE may be related to company-specific variables. Abarbanell and Lehavy (2003) argue that the incentive to manage earnings depends on the firm’s stock price reaction to the earning announcement; the managers of highly earnings-sensitive companies are more incentivized to manipulate earnings to exhibit a positive surprise than low-sensitivity firms’ managers are. In fact, it has been shown at multiple occasions that some firm-specific factors affect the sensitivity of the firm to the earnings announcement. Skinner and Sloan (2002), for example, have shown that growth companies are much more sensitive to the earnings surprise than value firms are, whereas Dalhiwal, Lee and Fargher (1991) have discovered that the market responds more to earnings surprises of low-leverage firms than to earnings surprises of high-leverage firms. According to the abovementioned “sensitivity incentive” proposed by Abarbanell and Lehavy (2003) and the market reaction observation in Skinner and Sloan (2002), the market will more intensely punish the high-growth firms for reporting negative earnings surprises, leading to a stronger motivation for the high growth firms’ managers to exhibit a positive surprise. Indeed, Athanasakou et al. (2009) and Bauman and Shaw (2006) find a significantly positive relation between the growth opportunities –proxied by the market-to-book ratio– and the probability of MBE, consistent with the evocations of Skinner and Sloan (2002). In addition, Matsumoto (2002) and Mcvay (2006) find a positive relationship between the long-term growth prospects and the probability of MBE, providing further support for this “sensitivity” hypothesis. In line with Dalhiwal et al. (1991) and the “sensitivity incentive” aforementioned, Bauman and Shaw (2006) find a significantly negative relationship between the probability of MBE and leverage. The size of the company also appears to affect the earnings surprise. Das et al. (1998) find a positive, yet mostly insignificant, relationship between the forecast bias and the size of the company, while Brown (1997) observes that small firms’ mean surprises are much more 16 negative than the large firms’ mean surprises, again suggesting a positive relation between size and earnings surprises. Matsumoto (2002) used the logarithm of market value as a proxy for size, and found a significantly positive effect of this variable on the binary variable MBE. This relationship could be due to the greater media attention on large firms and their disclosed earnings (O'Brien and Bhushan, 1990). As the salience of news is positively related to market reactions (Klibanoff, Lamont and Wizman, 1998), the market may react more to large firms’ disclosures, putting more pressure on their managers to meet or beat the expectations (again consistent with the sensitivity incentive discussed by Skinner and Sloan (2002) and Abarbanell and Lehavy (2003)). The industry of the company might also affect the incentives of managers to meet the analysts’ forecasts. For example, Amir and Lev (1996) report that the relevance of earnings for equity valuation is low (if not null) in some industries (e.g. cellular industry), so MBE should be less relevant for the managers of companies in these industries. Moreover, Matsumoto (2002) finds a positive relationship between the litigation risk, which is inherent to the industry in some instances, and the probability of MBE; as large price drops can be the trigger of lawsuits, managers facing a high litigation risk are more motivated to avoid a price drop due to a negative surprise by increasing the probability of MBE. Lastly, Brown (1997) reports results that are consistent with the above-mentioned observations: there are some differences in earnings surprises between industries. 2.6.3 Analyst coverage and MBE autocorrelation The behavior and opinions of analysts could also affect MBE (Matsumoto, 2002). First, the number of analysts making a forecast and the earnings uncertainty may affect the earnings surprise and the probability of MBE. Das et al. (1998) and Ackert and Athanassakos (1997) find that the analysts are more optimistic for low predictability firms. Hence we should expect a positive relationship between the earnings surprise (or MBE) and the earnings predictability, or likewise a negative relation between the earnings surprise and the standard deviation of the forecasts (proxy for uncertainty). Regarding analyst coverage, Brown (1997) reports that the average earnings surprise is larger for firms that are covered by many analysts. The strong correlation between analyst coverage and market capitalization (Lim, 2001) may explain this relationship. 17 Second, there is evidence that the extent of forecast guidance affects the probability of MBE. Indeed, Athanasakou et al. (2009) and Matsumoto (2002) reported that downward forecast guidance leads to a higher MBE probability. The mechanism of this forecast guidance is simple: the managers voluntarily influence analysts’ expectations, making them revise their expectations downwards to a beatable level in order to display MBE (Bartov et al., 2002). Third, in a study of the relationship between forecast errors and risk factors, Abarbanell and Lehavy (2003) argue that firms with high growth expectations (e.g. with a high market-to- book ratio) have stronger incentives than other firms to meet or beat analysts’ earnings forecasts because they are more sensitive to the earnings release (Skinner and Sloan, 2002). As these growth expectations are positively correlated with analyst recommendations (Abarbanell and Lehavy, 2003), the authors have used the analyst recommendation as a proxy for the firm’s sensitivity to earnings announcements. Consistent with the expectations, they reported that firms rated a “Buy” are more likely to engage in earnings management leading to MBE than firms rated a “Hold” or “Sell”. Finally, it is likely that there is autocorrelation in the MBE occurrences: as the incentives to meet the forecast thresholds persist over time (MTB ratio, leverage), one can wisely expect to see autocorrelation between MBE. Indeed, Bartov (1992) provided results that suggest such an autocorrelation while Lim (2001) and Mendenhall (1991) found a positive autocorrelation between forecast errors. 2.7 Predicting the earnings surprise The literature about the prediction of earnings surprises is scarcer and much narrower; the majority of studies focuses on the correlation between the consecutive surprises. This autocorrelation between surprises seems to come from the fact that analysts underreact to the forecast errors (Mendenhall, 1991 and Abarbanell and Bernard, 1992). In fact, this autocorrelation is so strong that it could be used to set up profitable strategies because the market reacts to subsequent surprises as if surprise autocorrelation did not exist (Bernard and Thomas, 1990). Specifically, Bernard and Thomas (1990) found that companies with a top SUE score (EPS surprise/ cross-sectional standard deviation of estimates) for a certain quarter tend to exhibit a positive and large SUE score in the subsequent period. Furthermore, the autocorrelation is not confined within the top or bottom deciles that were used by Bernard and 18 Thomas (1990) because Lim (2001) found a significantly positive relationship between the EPS surprise and its lagged value in a linear regression. Besides this autocorrelation, the mean recommendation of analysts may affect the earnings surprise. Indeed, Abarbanell and Lehavy (2003) reported that firms rated a “Buy” have larger mean and median forecast errors than firms rated a “Hold” or “Sell”. Lastly, the extent of forecast revisions is also likely to give indications concerning the future earnings surprise. In fact, upward revisions may be associated with more positive surprises, and downward revisions with less positive surprises (Kim et al., 2001). 3. Research design 3.1. Dataset This thesis is focused on the S&P 500 companies. This choice was made for several reasons. First, the S&P 500 companies are the companies for which the data is most readily available. This study includes some specific variables such as analyst revisions, leverage (LEV), the market-to-book ratio (MTB) which are only available for the largest or most famous companies. In this regard, the S&P 500 companies represented an obvious choice. Second, analyst coverage was a crucial issue for this study. Indeed, I wanted to test the significance of forecast revisions and recommendations as explaining factors for the MBE likelihood and the EPS surprise. This analyst coverage was only large enough for the largest companies. Third, the S&P500 is one of the most widely cited and analyzed index, hence one should observe fewer market anomalies for the S&P500 index than for other, less famous indices. Consequently, if the present study finds out that any proposed trading strategy is profitable for the S&P500 companies, the likelihood that a similar trading strategy will also work out for other indices is large (while the contrary needs not be true). The dataset I use in the study includes quarterly observations for three important reasons. First, I wanted a sample large enough to produce robust results. My dataset includes 5038 firm-quarters. Similar studies about earnings surprises employ very large datasets, like Bartov et al. (2002) who uses more than 64,000 observations or Matsumoto (2002) who employs 29,460 observations. Though my sample is much smaller, it has approximately the same size as that of Athanasakou, Strong and Walker (2009) which includes 5,117 observations, suggesting that the size of the present sample is correct and can be used for research. 19 Moreover, focusing on quarterly data allows me to work with very recent data, so the likelihood that my model will work in the very near future is higher than if I had worked with annual data. Finally, working with quarterly data instead of annual data makes the model usable in more instances per year, so the total potential gains per year from the proposed trading strategies are larger with quarterly data. In this thesis, I analyze the most recent 11 periods for the 500 S&P500 companies, leading to a full sample of 5,500 firm-quarters observations. I recorded the company releases as of February 2009 and until October 2011. Among those 500 companies, 42 were dropped out of the sample for several reasons. First, I removed those companies whose EPS surprises were not available from I/B/E/S for one or more quarters, as the earnings surprise is one of the focus of the paper. Second, I removed the companies which had less than four forecast revisions for at least five periods. Indeed, firm-quarter observations that have less than four revisions are puzzling the results, because their percentage of upward or downward revisions (UP% and DOWN%) usually lie at the extremes (0% or 100%), leading to incoherent coefficients in the regressions. Moreover, I removed some financial companies such as AIG, because these firms mostly produced outliers for all the variables analyzed. As financial companies likely have different incentives to manage earnings, Matsumoto (2002) removed all the financial companies from her sample. Nevertheless, doing so would have reduced my sample size by 748 observations, which I judged as too important. As a result, I kept the financial companies in my sample. Lastly, some companies were dropped because of conflicting event dates. Eventually, the final sample comprises 5,038 firm-quarter observations spread over 11 quarters between 2009 and 2011. One important issue about the methodology of this thesis is the pooling of the observations that happened at different moments in time into one dataset. By doing this, I neglect the time- varying impact of macroeconomic factors on the mean surprise. Matsumoto (2002) found a significantly positive relationship between the growth in industrial production and the probability to meet or beat the expectations and O’Brien (1994) observed a positive relation between growth in industrial production (and other macroeconomic factors) and analysts' forecast errors. These observations suggest that I should use panel data. Nevertheless, the great majority of papers that study the drivers of MBE pool the observations together (Bartov, 1992, Matsumoto, 2002; Athanasakou, Strong and Walker (2009) and Bauman and Shaw (2006)). Moreover, the relationship between the surprises and the macroeconomic factors is a contemporary relationship: papers such as Matsumoto (2002) and Bauman and Shaw (2006) 20 are backward-looking and include an “industrial production growth” variable in their regression only as a control variable. This control variable happens to be concurrent to the surprise observation and therefore is not compatible with the forecasting nature of the models I want to construct. Nevertheless, I will mitigate the issue of the time effect on the earnings surprise by including quarter dummies in the models. The last issue about the sample is its time frame. Indeed, the period between 2008 and 2011 was rather turbulent in terms of market returns and macroeconomic news, meaning that the driving forces in my regressions may not hold in other time periods. Although I cannot counteract this argument, I justify the time frame decision by the fact that I wanted to use a sample whose characteristics and dynamics would most probably persist in the near future (years 2012 and 2013). This naturally led me to choose the most recent sample of quarterly data. As will be discussed in section 3.3, I split the sample into two time periods. The motivation behind this split is specifically to address this time frame concern: I will determine if the relations that I find in the 2009-2010 period also hold for the (less turbulent) first three quarters of 2011. 3.2 Data collection The models in my thesis will involve four main types of data: accounting measures, analyst estimates and recommendations, EPS surprises and cumulative abnormal returns, whose temporal relative positions are depicted in Figure 2. 21 T q- s Tq T q+1- s T q+1 CAR f,q (CRSP) SUR f,q (I/B/E/S) REC,EST (I/B/E/S) Acc. Measures (Compustat) T q -1 Tq T q+1 Figure 2: Time frame. All the accounting measures are lagged by one quarter and all the analyst recommendations and estimates are anterior to the earnings announcement by “s”, which lies between 0 and 1. Starting from the bottom of the figure, one can see that all the accounting measures are lagged by one quarter in comparison to the surprise and the corresponding market reaction (the dependent variables). This lag was compulsory because the surprise and the accounting measures are generally released at the same time, and because these accounting numbers must be anterior to the announcement in the forecasting model I develop. All the accounting measures used in the model have been subject to this lag manipulation and are therefore anterior to the earnings release. The accounting variables were retrieved from Compustat. A stage higher on figure 2, we can see that the recommendations and estimates are also anterior to the EPS announcement, but by less than the accounting measures. I retrieved the last available recommendation before the earnings announcement (MEANREC) from I/B/E/S and the last analyst forecasts (FEPS) before the announcement to calculate the earnings surprise, according to the methodology of Bauman and Shaw (2006). The variable FEPS is usually recorded between one week and one month prior to the EPS announcement. Finally, the actual earnings per share (EPS) figure was retrieved from I/B/E/S and the returns on the S&P500 index and the returns on every stock around the earnings announcements were 22 retrieved from CRSP. These returns were then combined to calculate the cumulative abnormal returns (CARs). This combination is discussed in the following section. 3.3 Method I will begin with answering subquestions 1a, 2a, 3a and 3a’ by running regressions (1), (2), (3a) and (3a’) –that are to be discussed– on a test sample that includes all observations in years 2009 and 2010 to verify that one can predict MBE, the surprise and the corresponding market reaction. This test sample thus encompasses 3664 observations (458 firms reporting their earnings over 8 quarters). Then, I will investigate the profitability of the four trading strategies (subquestions 1b, 2b, 3b and 3b’) on this test sample. Then, I will test if the relationships hold in the holdout sample, comprising 1374 observations (458 firms over 3 quarters). Figure 3 exhibits the time framing of the studies I will conduct. 2009 Q1234 2010 Q1234 2011 Q123 4 4 4 Test sample/ Model creation: 3664 Obs. Holdout sample: 1374 Obs. Figure 3: The regressions and tests of srategies will first be conducted on the test sample that comprises 8 quarters. Then, the relationships will be tested on the holdout sample. Consequently, the results section of my thesis will go through the following steps: Step A. Answer subquestions 1a, 2a and 3a and 3a’ and identify the true drivers in the test sample, Step B. Answer subquestions 1b, 2b and 3b and identify the profitable strategies in the test sample, Step C. Verify that the relationships found in A and B persist in the holdout sample. 3.3.1 Step A 23 In this section, I discuss how I concretely incorporate the prior findings that were discussed in section 2.5 in a model whose goal is to predict MBE. This model will be called regression (1). MBE Q aPOSEE Q S bPOSUE Q S cMTB Q 1 dLEV Q 1 eSIZE Q 1 fSTDEV _ PQ S (1) gNBESTQ S hUP % Q S iDOWN % Q S jMEANREC Q S kMBEQ 1 lMBE Q 2 mMBE Q 3 nQ 2 oQ3 pQ 4 qQ5 rQ 6 sQ 7 y This equation’s goal is the prediction of the probability of meeting or beating of the estimates (MBE). As this MBE variable is a binary variable, I model this equation in a logistic regression (LOGIT), as in Bauman and Shaw (2006). The MBE term obtains a value of 1 if the earnings surprise is nonnegative (SUR>= 0), and 0 if the surprise is negative. The earnings surprise is itself calculated as the difference between the actual earnings per share (EPS) and the latest mean EPS estimate from I/B/E/S, anterior to the earnings announcement. On the other side of the equation, one can find the independent variables, which are the factors which have been found to significantly influence MBE in the prior literature, as discussed in section 2.6. First, I included two binary variables POSEE and POSUE. The variable POSEE takes a value of 1 if the expected EPS is nonnegative and 0 otherwise, and POSUE catches a value of 1 if the expected EPS is larger than the EPS of the same quarter of last year as in Matsumoto (2002). The inclusion of these two variables in a regression having MBE as a dependent variable seems righteous given the suggestion of pecking order between the benchmarks (Degeorge at al., 1999) and the prior evidence of Athanasakou et al. (2009), Bauman and Shaw (2006) and Matsumoto (2002). The market-to-book ratio has also been shown to significantly and positively influence the probability of MBE (Matsumoto, 2002 and McVay et al., 2006). Following Matsumoto (2002), I include the market-to-book ratio variable, called MTB in the regression, and expect a positive coefficient for this variable in line with the “sensitivity incentive” reported by Abarbanell and Lehavy (2003), which was examined in section 2.6. Moreover, Bauman and Shaw (2006) find a significantly negative relationship between the probability of MBE and firm leverage. For this reason, I incorporate the leverage, labeled LEV, in the regression explaining MBE. The last firm feature that is included in the regression is SIZE. In fact, size was found to be significantly and positively related to the probability of meeting or beating the estimates by Matsumoto (2002). One of the reasons for the relationship between MBE and size could be the larger media attention on large firms and their disclosed earnings. As the salience of news is positively related to market reactions (Klibanoff, Lamont and Wizman, 24 1998), large firms may be more motivated to meet the forecasts. Another explanation for this positive relation between MBE and size may be provided by Kasznik and Lev (1995). They find that larger firms are more likely to warn investors of future inferior performance (similar to downward forecast guidance), and these warnings increase the probability of MBE (Matsumoto, 2002). I will follow the methodology of Matsumoto’s (2002) paper and will include the logarithm of the market value, called SIZE, in the regression. In equation 1, the market-to-book ratio (MTB) is defined as Total market value (MKVALTQ in Compustat) divided by the book value of equity (SEQQ in Compustat) as in Skinner and Sloan (2002), and the leverage is calculated as Total long-term debt (DLTTQ)/Book value of equity (SEQQ in Compustat). Moreover, the logarithm of market capitalization (MKVALTQ) is used as a proxy of size as in Matsumoto (2002) and Bhushan (1989). The next variable in the regression is the price-scaled standard deviation of analysts’ estimates. The reason for the inclusion of this variable is its function of proxy for the uncertainty. Ackert and Athanassakos (1997) proposed that analysts are more optimistic for high-uncertainty predictions, suggesting a negative relationship between the uncertainty and the surprise and consequently between the uncertainty and MBE. In equation (1), I use the same proxy for uncertainty as Ackert and Athanassakos (1997): the standard deviation of the forecasts, which was retrieved from I/B/E/S. As this standard deviation depends on the expected earnings and the price of the company’s shares, I use the price-scaled standard deviation of the estimates, STDEV_P as in Lin et al. (2006). This price corresponds to the price of the stock on the day of the earnings announcement in the previous quarter, and was retrieved from Compustat. Moreover, Brown (1997) reported that the extent of analyst coverage may affect the earnings surprise. In fact, this variable could have a significant effect on the probability of MBE, because it may also be used as a proxy of size, or market attention. Indeed, the market capitalization (proxy for firm size) is positively related to analyst coverage (O'Brien and Bhushan, 1990), and this relationship is very strong: Lim (2001) finds a correlation of 0.8 between the natural logarithm of the market capitalization and the natural logarithm of anlyst coverage. Given that the market tends to react more to salient news (Klibanoff, Lamont and Wizman, 1998) and that analyst coverage is related to the media attention, we should expect those “heavily followed” firms to be motivated to avoid negative surprises. I will investigate this issue by including the number of analyst forecasts in my regression through the term “NBEST”. NBEST is simply the number of forecasts used to calculate the average forecast, and is retrieved from I/B/E/S. 25 Additionally, Athanasakou et al (2009) and Matsumoto (2002) found a significantly positive relationship between the downward forecast guidance and the probability of MBE. As this forecast guidance likely results in forecast revisions by the analysts, the number of downward revisions may be used as a proxy for downward guidance (upward and downward). Consequently, I include the number of upward and downward revisions over the three months preceding the earnings announcement to proxy the extent of guidance. Instead of using the absolute number of upward and downward revisions, I scale the revisions by the number of analysts forecasts (NBEST) to control for the extent of the coverage. These two terms will be referred to as UP% and DOWN%. These variables can be seen as the number of downward and upward estimates per forecast. If we assume that no analyst revises his expectations more than once over the three months before the announcement (though this is unlikely), then DOWN% and UP% can be considered as the percentage of analysts that revised their forecast down and up. DOWN% and UP% probably do a better job at capturing the trend in analysts’ expectations because the absolute numbers of revisions are very sensitive to the number of estimates, NBEST. Moreover, Abarbanbell and Lehavy (2003) reported that “buy” rated firms (firms with a mean recommendation lower than 2 on a Value Line scale) are more likely to engage in earnings management than “hold” (mean recommendation between 2 and 3) or “sell” (recommendation >3) rated firms. In regression (1), I take the mean recommendation variable, MEANREC, which is the variable on which Abarbanell and Lehavy (2003) based their Buy, Hold and Sell categories. I expect a negative relationship between MEANREC (1= strong buy, 5=strong sell) and MBE. MEANREC was retrieved from I/B/E/S and represents the most recent (but anterior to the EPS announcement) mean analyst recommendation on a scale from 1 to 5. In addition, as Lim (2001) and Mendenhall (1991) have found positive autocorrelation in the forecast errors, I decided to include three lagged MBE terms, MBE Q 1 , MBE Q 2 and MBE Q 3 which are binary variables which take a value of 1 if the firm has met or beaten the expectations in quarter -1, -2 and -3 respectively, and 0 if it was not the case. Finally, I include quarter dummies to control for the change in MBE frequency over time. The base level is the first quarter of 2009, and seven quarter dummies are used in the model, given that the regression is run on the test sample that comprises eight quarters (figure 3). 26 I am now to discuss the model I use to predict the earnings surprise. This model will help me answer subquestion 2a “What are the drivers of the earnings surprises?” To identify the drivers of the earnings surprise, I regress linearly the surprise against several variables, as in equation (2): SURQ aSURQ1 bSURQ2 cSURQ3 cMEANREC S dUP% Q S eDOWN % QS Q (2) fQ2 gQ3 hQ4 iQ5 jQ6 kQ7 y The variable SUR is a variable representing the earnings per share (EPS) surprise. This EPS surprise is defined as the actual (reported) EPS minus the most recent average of analysts’ expected EPS. The actual and mean expected EPS were all retrieved from the I/B/E/S database. Furthermore, the surprise SUR was scaled by the price of the stock in quarter Q-1 in order to control for the “price effect” on EPS as in Abarbanell and Lehavy (2003) and Abarbanell and Bushee (1997) and in line with the evidence of Christie (1987). Even after the scaling of the surprise by the price, the sample of surprises (5038 firm-quarters) still included many extreme values. Indeed, of the 5038 observations, 103 units or 2.04% of the sample reported a price-scaled surprise outside of the “mean +/- 3 standard deviations region” whereas the normal distribution would predict 0.28% or 14 exceptions. These outliers are crucial, because they heavily influence the coefficients and the significance results. Consequently, I follow the methodology of Lim (2001) who also performed a linear regression, and remove the observations that are in the two 2.5% extreme tails. This manipulation will be relevant for regressions (2), (3a) and (3a’) because they involve linear regressions that are dependent on the surprises. On the other hand, the manipulation is not performed for regression (1): as equation (1) requires the coding of the surprises into a binary variable, the issue of outliers in the dependent variable vanishes and the manipulation is no longer needed (Bartov, 1992). Equation (2) includes three lags of the surprise: SURQ 1 , SURQ 2 and SURQ 3 . While Lim (2001) only included one lag in a regression to explain the forecast bias, Bernard and Thomas (1990) documented that the one, two and three-quarter-lagged standardized unexpected earnings (SUE scores) are positively and significantly related to the actual SUE score, indicating a tendency for surprises to persist over time, partly due to analyst underreaction (Mendenhall, 1991 and Abarbanell and Bernard, 1992). I thus followed the results of Bernard and Thomas (1990) and included the one-, two- and three-quarter lagged surprises to predict the current earnings surprise. 27 I insert the variable MEANREC which represents the mean recommendation of analysts shortly before the earnings announcement, expecting a negative relationship between the surprise and the recommendation. Indeed, Abarbanell and Lehavy (2003) have found that firms with a low mean recommendation (< 2 on the Value Line scale) exhibit larger mean and median surprise than firms with a larger mean recommendation. I also include the number of upward and downward revisions over the three months preceding the earnings announcement to control for the trend in analyst expectations, following Kim, Lim and Shaw (2001) who advise to use a revision-adjusted mean forecast, i.e. the mean forecast plus a multiple of the forecast revisions. According to Kim et al. (2001), upward revisions are associated with more positive surprises, and downward revisions with less positive surprises. The inclusion of the two terms UP (number of upward revisions) and DOWN (number of downward revisions) is meant to fill this gap. As the number of revisions is dependent on the extent of the coverage, I insert in the equation the number of upward and downward revisions divided by the number of analysts following the stock to control for the size of the company. These two terms will be referred to as UP% and DOWN%. Hence, I expect a positive coefficient for UP%. The sign of the DOWN% coefficient is more uncertain. Indeed, one the one side, Kim et al. (2001) argue for a negative sign but on the other hand downward analyst revisions could be due to forecast guidance by the managers, which has been shown to positively affect the surprise by Matsumoto (2002), meaning a positive coefficient for the term DOWN%. Lastly, I incorporate quarter dummies to control for the changes in the mean surprise over time. The base level is the same as that of regression (1): the first quarter of 2009. The coefficients in these two regressions (without the quarter dummies whose coefficients cannot be known ex ante) will then be used to produce predicted values for the dependent variables SUR and MBE: E(SUR) and E(MBE). These expected (forecasted) values are then inputted in regression (3a). This third regression is a multiple regression that relates the market reaction to the earnings announcements and the predicted values obtained from equations (1) and (2): (3a) CAR Q aE ( SUR Q ) bE ( MBE Q ) cQ 2 dQ3 eQ 4 fQ 5 gQ6 hQ7 iQ8 y 28 This equation is based on the cumulative abnormal return (CAR) of the stock on the time window (-2;0) like Bernard and Thomas (1990). This CAR is the cumulative market-adjusted return and is calculated as follows: for each firm-quarter observation, I subtract the S&P500 daily return from the daily stock return in order to get the abnormal return for every day around the earnings announcement, and then add up these daily returns to obtain the cumulative abnormal return over the three-day time window. As the use of market-adjusted returns instead of risk-adjusted returns for event studies does not significantly alter the significance results (Brown and Warner, 1985) and given that the calculation of risk-adjusted returns is much more time consuming, I think my decision to use market-adjusted returns was wise. The right-hand side of this equation only includes estimates of the concurrent surprise and probability of MBE. E(SUR) is the estimate of the surprise at time Q got from equation (2) and E(MBE) is the estimate of the probability that the firm will meet or beat the estimates (MBE) at time Q got from equation (1). These two terms are estimates (gotten from equations (1) and (2)) of the earnings surprise and MBE, which have been proven to significantly influence the market reaction to an earnings announcement (Bartov et al., 2002). Surprisingly, there is evidence that the short-term market reaction following an earnings announcement can be predicted by using the past standardized unexpected earnings (SUE score, defined as the earnings surprise reported by the firm divided by the cross-sectional standard deviation of the earnings surprises), and can produce abnormal returns (Bhushan, 1992 and Bernard and Thomas, 1990), probably because of the autocorrelation in the surprises. This autocorrelation is mainly positive, and lasts for about 3 quarters (Bernard and Thomas, 1990). Regression (3a’) accounts for this autocorrelation and attempts to explain the market reaction to the earnings announcements with the past three price-scaled earnings surprises, SURQ 1 , SURQ 2 and SURQ 3 which are expected to have a positive coefficient in the regression as the study of Bernard and Thomas (1990) suggests. I expect to obtain a low R square for this model given that Bernard and Thomas (1990) arrived at a predictive model that explains between 0.6% and 1.0% of the variation in the CARs. (3a’) CAR Q aSUR Q 1 bSUR Q 2 cSUR Q 3 dQ2 eQ3 fQ 4 gQ5 hQ6 iQ7 jQ 8 y 29 Table 1 below gathers all the variables that are included in the regressions. I also mention the regression in which the variable is present (Reg.), the database from which I retrieved the data (origin), the papers that motivated the inclusion of these variables in the regression (motivation) and the papers that inspired the methodology. Table 1 Features of the variables included in the regressions Variable Reg. Origin Motivation Methodology Bauman and Shaw (2006) Lim (2001) Dummy indicating a positive or null MBE 1 I/B/E/S Athanasakou et al. (2009) earnings surprise (Matsumoto, 2002) Matsumoto (2002) Dummy indicating a positive expected EPS POSEE 1 I/B/E/S Degeorge et al. (1999) (Degeorge et al., 1999) Degeorge et al. (1999) Dummy indicating a positive expected POSUE 1 I/B/E/S Matsumoto (2002) change in EPS (Degeorge et al., 1999) Matsumoto (2002) Market value/Book value of equity MTB 1 COMPUSTAT McVay (2006) (Matsumoto, 2002) Skinner and Sloan (2002) Total long-term debt (DLTTQ in LEV 1 COMPUSTAT Bauman and Shaw (2006) Compustat)/Book value of equity (SEQQ). Logarithm of market capitalization SIZE 1 COMPUSTAT Matsumoto (2002) (MKVALTQ) (Matsumoto, 2002 and Bhushan, 1989) Standard deviation of estimates/ price on Ackert and Athanassakos the previous quarter’s earnings STDEV_P 1 I/B/E/S (1997) announcement (Lin, 2006) Brown (1997) NBEST 1 I/B/E/S Number of earnings forecasts (Lim, 2001) Lim (2001) Number of upward revisions over the three 1: Athanasakou et al. (2009), months preceding the earnings release UP% 1/2 I/B/E/S Matsumoto (2002) scaled by the number of estimates 2: Kim, Lim and Shaw (2001) (UP/NBEST) Number of downward revisions over the 1: Athanasakou et al. (2009), three months preceding the earnings release DOWN% 1/2 I/B/E/S Matsumoto (2002) scaled by the number of estimates 2: Kim, Lim and Shaw (2001) (DOWN/NBEST) MEANREC 1/2 I/B/E/S Abarbanell and Lehavy Mean recommendation before the 30 (2003) announcement (Abarbanell and Lehavy, 2003) Lim (2001) Dummy indicating a positive one-quarter- MBEL1 1 I/B/E/S Mendenhall (1991) lagged EPS (Matsumoto, 2002) Lim (2001) Dummy indicating a positive two-quarter- MBEL2 1 I/B/E/S Mendenhall (1991) lagged EPS (Matsumoto, 2002) Lim (2001) Dummy indicating a positive three-quarter- MBEL3 1 I/B/E/S Mendenhall (1991) lagged EPS (Matsumoto, 2002) Earnings surprise/ price on the last Lim (2001) quarter’s earnings announcement SUR 2 I/B/E/S Bernard and Thomas (1990) Abarbanell and Bushee (1997) Christie (1987) Abarbanell and Bushee (1997) SURL1 2/3a’ I/B/E/S Lim (2001) Christie (1987) Abarbanell and Bushee (1997) SURL2 2/3a’ I/B/E/S Bernard and Thomas (1990) Christie (1987) Abarbanell and Bushee (1997) SURL3 2/3a’ I/B/E/S Bernard and Thomas (1990) Christie (1987) Market-adjusted cumulative abnormal CAR 3a/3a’ CRSP Bernard and Thomas (1990) return (Brown and Warner, 1985) Predicted value for the surprise from E(SUR) 3a Regression 2 Bartov (2002) regression 2 Predicted probability of MBE from E(MBE) 3a Regression 1 Bartov (2002) regression 3 3.3.2 Step B The papers which study the profitability of trading strategies generally use quintiles and deciles to create subsamples whose average abnormal returns are subsequently analyzed (e.g. Jegadeesh and Titman, 1993; De Bondt and Thaler, 1985). In a study of the market reaction to the earnings announcements, Bernard and Thomas (1990) formed deciles based on the past SUE scores to describe the differences in abnormal returns between subgroups, while Collins and Hribar (2000) used the quintiles of the past SUE scores. The quintiles and deciles methods produced roughly similar results: the top decile and quintile with respect to the past SUE score produce significantly positive abnormal returns while the bottom decile and quintile produce significantly negative abnormal returns. However, the quintile method has one important advantage: it produces less variable results (Collins and Hribar, 2000), and this is particularly relevant for my study which involves a relatively small sample. Given that the 31 returns observed using both methods are similar and that the quintile method produces more robust results, I decided to adopt the quintile formation that Collins and Hribar (2000) have used. The first proposed strategy pertains to the expected probability of MBE: after having generated the predicted probabilities of MBE from regression (1), I rank the observations in an ascending order and isolate the bottom and top quintiles. Afterwards, I compute the cumulative average abnormal returns (CAAR) of those subsamples and compare them. The strategy I will test consists of buying the top quintile and selling the bottom quintile, which is a zero-cost strategy as in Jegadeesh and Titman (1993). Though no prior paper investigates the profitability of such a trading strategy, the fact that (a) MBE leads to a premium (Bartov et al. 2002) and (b) MBE is predictable (see section 2.6) renders this analysis decent. Specifically, if the market participants are not aware of this MBE predictability, we should observe positive abnormal returns for the top quintile with respect to the predicted MBE and negative abnormal returns for the bottom quintile. In contrast, if the stock traders are aware of this predictability and incorporate it in their reaction to the earnings announcement, we should observe no statistical difference between the quintiles (consistent with the semi-strong form of market efficiency). The results of this strategy will provide an answer to subquestion 1b. Second I will investigate the profitability of the zero-cost strategy that buys the expected surprise top quintile and sells the bottom quintile, using the same method as for the first strategy. As with the first strategy, the returns may not differ in case the investors account for the surprise drivers in equation (2). The results of this strategy will answer subquestion 2b. The third and fourth trading strategies go a step further and involve buying the expected CAR top quintile and selling the bottom expected CAR quintile. Indeed, as will be shown, the cumulative abnormal returns around an earnings announcement may be predictable to some extent, and trading strategies harnessing this predictability of market reaction might be profitable. These quintiles will be formed by using equations (3a) and (3a’) sequentially. Subquestions 3b and 3b’ will then be answered. 3.3.3 Step C 32 In this step, I will verify that the relationships found in steps A and B persist in the holdout sample (see figure 3). For example, I will run once again regression (1) to investigate whether the variables in the regressions remain significant. Moreover, if any of the proposed strategy is profitable, I will inspect its persistence in the holdout sample. 4. Results 4.1 Descriptive statistics In this section, I will consider the characteristics of all observations, whether they are in the test or the holdout sample. As can be seen in table 2, the firms in the sample are large companies. The mean market capitalization is $21.4 billion and the median is $9.78 billion, indicating a skewed distribution with some very large market capitalizations. The average market-to-book ratio is 2.95, and the average leverage of the companies, computed as long- term debt divided by book value of equity, is as high as 78%. This leverage level looks rather high, but it is driven by some observations with extreme values as the median is much lower at 50%. This skewness is partly explained by the presence of financial companies which are by nature more levered than the rest, but remains even if we only consider nonfinancial firms. The features of the non-financial and financial companies can be observed in the last two rows. It seems indeed that the financial companies tend to be systematically more levered than the non-financial companies. In addition, the standard deviation of the distribution and the mean leverage are greater for financial companies, which indicates that the leverage of some financial companies is very large. 33 Table 2 Firm-quarter characteristics The table provides descriptive statistics on firms’ characteristics. These characteristics are the market capitalization (market value in $ million), the market-to-book ratio (MTB) and the leverage. A distinction is also made between financial and non-financial companies. Mean St. dev. Q1 Median Q3 Market Value ($ Ml) 21402 36266 5274 9777 20134 MTB 2,014 3,451 1,353 2,194 3,59 Leverage 0,779 1,037 0,225 0,497 0,968 Leverage financials 0,872 1,269 0,245 0,512 1,028 Leverage non-financials 0,764 0,993 0,217 0,500 0,955 Another important aspect of the dataset is analysts’ opinions. These analyst beliefs are of key importance for the study of the earnings announcements because they influence the earnings surprise (e.g. Kim, Lim and Shaw, 2001), a fundamental issue for firm valuation (Bartov et al., 2002). Table 3 summarizes the data concerning analysts’ recommendations, upward and downward revisions and estimates. Table 3 Descriptive statistics on Analysts’ opinions The table exhibits descriptive statistics on analysts’ opinions. The table includes the number of estimates for a particular firm-quarter (NBEST), the number of upward and downward revisions over the three months preceding the earnings release divided by the number of estimates (%UP and %DOWN), the mean analyst recommendation shortly before the announcement (MEANREC) and the percentage of analysts that recommend to buy and sell the stock (BUY% and SELL%) before the earnings release. Variables Mean St. dev. Q1 Median Q3 NBEST 16,81 7,214 12,00 16,00 21,00 %UP 0,521 0,421 0,200 0,444 0,778 %DOWN 0,557 0,436 0,222 0,476 0,818 MEANREC 2,331 0,378 2,050 2,300 2,570 BUY% 50.5% 21,1% 35,7% 51,7% 66,7% SELL% 6,12% 8,26% 0% 4,17% 9,14% We can see from table 3 that the average number of EPS estimates for a particular firm- quarter is close to 17. Thus, on average, slightly less than 17 analysts make a prediction for a particular firm-quarter EPS announcement. The distributions of upward and downward revisions are relatively akin (mean, quartiles, standard deviation), yet %DOWN appears 34 consistently larger than UP%. This indicates that analysts are slightly more prone to revise their forecasts downward than upward. The mean recommendation, which can take a value from 1 to 5, is on average 2.33 and indicates a clear positive bias. The 75th percentile is 2.57, still far away from the supposed 3-neutral point, which means that more than 75 percent of the firm-quarters benefit from a positive recommendation shortly before the announcement. One can retrieve an even more accurate idea of the analysts’ opinion by analyzing the BUY, HOLD and SELL advice. BUY% and SELL% represent the percentage of analysts that respectively advise buying and selling the stock of the firm. Table 3 shows that the average percentage of “BUY” advice exceeds 50% of the total recommendations while the average and median of SELL% are only 6.12% and 4.17%. This means that in 50% of the cases, less than 4.17% of the analysts recommend selling the stock. These observations go hand in hand with analyst optimism and the conflict of interest (O’brien, Mcnichols and Lin, 2005) I have discussed in section 2.4. These analyst opinions are usually coupled with estimates of the future earnings which are used for the earnings surprise computation. Table 4 exhibits the descriptive statistics on the variables related to the expected earnings and the earnings surprise. Table 4 Descriptive statistics on Analysts’ EPS estimates and the EPS surprise The table provides descriptive statistics on analysts’ EPS estimates and the EPS surprise. The first row reports the statistics for the “unscaled” surprise (USUR) defined as the actual EPS minus the last available EPS mean forecast. Moreover, the table exhibits the price-scaled surprise SUR (calculated as the EPS surprise divided by the price of the stock on the day of the announcement of the past quarter’s earnings) and the data about MBE, which is a binary variable that equals one if the surprise is nonnegative and 0 otherwise. The variable POSEE is a binary variable which takes a value of one if the mean of the analyst estimates is positive or zero, and 0 if it is negative. Finally, POSUE is a binary variable with a value of 1 if the mean analyst expectation is larger than or equal to the EPS in the same quarter of last year and 0 otherwise. Variables Mean St. Dev. Q1 Median Q3 USUR 0,0504 0,259 0,0000 0,0300 0,0800 SUR 0,0013 0,003 0,0000 0,0008 0,0022 MBE 0,8067 0,394 1,0000 1,0000 1,0000 POSEE 0,9440 0,229 1,0000 1,0000 1,0000 POSUE 0,6443 0,478 0,0000 1,0000 1,0000 As can be seen from the first two rows, the average as well as the median EPS surprise is positive. As the 25th percentile is 0, one can readily conclude that at least 75% of the companies in the sample report a nonnegative earnings surprise. In fact, we see from the third 35 row that the average MBE is 0.8067, meaning that 80.67% of the companies in the sample beat or meet the estimates (MBE). Moreover, positive earnings (POSEE) are expected 94.4% of the time and growth in earnings (versus the same quarter of last year – POSUE) is expected in 64.43% of the instances. These earnings figures trigger market reactions whose features are described in the following table. Table 5 summarizes the cumulative abnormal returns of the stocks around the earnings announcements for various time windows. Table 5 Descriptive statistics on the CARs The table provides descriptive statistics on the cumulative abnormal return. The CARs are the cumulative abnormal returns and are calculated as the sum of the abnormal returns over the time window. The abnormal return is gotten by subtracting the S&P500 return from the firm’s stock return. CAR CAAR St. Dev. Q1 Median Q3 CAR (-5 ; 4) ,0063 ,0738 -,0351 ,0055 ,0441 CAR (-2 ; 1) ,0024 ,0628 -,0307 ,0014 ,0347 CAR (-5 ; 0) ,0065 ,0644 -,0274 ,0042 ,0375 CAR (-2 ; 0) ,0033 ,0576 -,0275 ,0014 ,0326 CAR (0 ; 0) ,0010 ,0548 -,0276 -,0010 ,0278 CAR (0 ; 1) ,0001 ,0602 -,0308 -,0002 ,0301 CAR (0 ; 4) ,0008 ,0654 -,0344 ,0004 ,0355 The values in the last three rows suggest that the length of the time window following the announcement is not a crucial issue. Indeed, the last three windows have positive, yet fairly small, CAARs and the medians do not substantially differ. Moreover, we can observe (as expected) that the standard deviation of the returns increases along with the length of the time window, but that the major part of the fluctuations are concentrated on the announcement day. In contrast, the inclusion of the trading days prior to the announcement in the time window produces significantly different results. Indeed, the time windows that include some (2 or 5) days before the announcement exhibit larger average cumulative abnormal returns than the time windows strictly posterior to the announcement. For example, the mean CAR for a time window of (-5; 4) is 0.63% while it is only 0.08% for the time window (0; 4). Though the reasons for such a relation are hard to discover, the superior performance of stocks prior to the announcement could be explained by the fact that the equally-weighted CAAR documented in the above table does not take account of the size of the company, while the benchmark (the 36 S&P500 index) is a value-weighted average: it may be the case that small firms earn a large premium, while the larger firms get a null or negative return prior to the announcement. Table 6 exhibits the cumulative average abnormal return over the five days preceding the announcement for each of the size quartiles. Table 6 Descriptive statistics on CAR(-5,1) for 4 size quartiles The table provides descriptive statistics on the CAR (-5,-1). The sample is divided into 4 size quartiles, quartile 1 being the quartile that includes the 25% smallest companies (by market capitalization). The mean, the standard deviation, the standard error and the 95% confidence interval of the CAR are reported. Size Mean St. Dev. St. Error 95% Lower Bound 95% Upper Bound Quartile 1 0,0094 0,0457 0,00129 0,0069 0,0120 Quartile 2 0,0079 0,0332 0,00094 0,0061 0,0098 Quartile 3 0,0027 0,0332 0,00093 0,0009 0,0046 Quartile 4 0,0019 0,0287 0,00081 0,0004 0,0035 Total sample 0,0055 0,0358 0,00050 0,0045 0,0065 We indeed see that the 25% smallest firms earn an average of 0.94% market-adjusted return over the five trading days prior to the announcement, while the other quartiles do worse. Nevertheless, the 25% largest firms still exhibit a slightly positive CAAR over the five days, which suggests that the firms in the sample have had a tendency to outperform the market before the announcement, and consequently, the firm-quarters that were removed from the sample for data-availability issues probably underperformed. Besides, an important conclusion one can draw from table 5 is that size has an important effect on the return around the earnings announcement; in an untabulated ANOVA test, the difference in CAAR between the four quartiles is significant. Moreover, the distribution of CARs in the first quartile has a bell-shape, with a median of 0.72% which suggests that these results are not driven by positive outliers but represent instead a real trend. I will end this section with a short discussion of the drivers of the two-day market reaction. The papers I have discussed in section 2.2 suggest that nowadays, the main driver of the market reaction to the earnings announcement is the earnings surprise (Dechow, Richardson and Tuna, 2003, Brown and Caylor, 2005 and Bartov et al., 2002). Table 7 presents the results of the regression between the cumulative abnormal return of the stocks over the time window (0,1) and the four explaining variables that were proposed in the prior literature and discussed 37 in section 2.2; MBE is the dummy variable that equals 1 if the company beats the estimates, and 0 if not, SUR is the price-scaled surprise, GROWTH is the growth in earnings (defined as the difference between the actual EPS at time t and the actual EPS at time t-4, scaled by the price of the stock) and POS EARN is a dummy variable that takes a value of one if the earnings is positive. Table 7 Drivers of the market reaction to earnings announcements The table exhibits the results of the regression: CAR Q aSUR Q bMBE Q cGROWTH Q dPOSEARN Q y CAR is the cumulative abnormal return over the (0,1) time window. SUR is the price-scaled surprise observed at time 0, MBE is a binary variable equal to 1 if the surprise is nonnegative and 0 otherwise, GROWTH is a dummy variable that takes a value of one if the company reports growth in earnings compared to the same quarter of last year and 0 otherwise and finally POSEARN equals 1 when the company reports a positive earnings per share (EPS) and 0 otherwise. All these variables are observed at the same time. CAR (0,1) Coefficient Std. Error t-test p-value Constant (y) -,010 ,005 -2,159 ,031 MBE ,016 ,002 6,498 ,000 SUR 3,418 ,290 11,780 ,000 GROWTH ,014 ,027 ,525 ,600 POS EARN -,008 ,005 -1,637 ,102 The growth in earnings (GROWTH) does not significantly influence the stock return following the EPS announcement. Moreover, firms that exhibit a positive earnings figure have a 0.8% lower market-adjusted return than firms with a negative EPS, though the difference is marginally significant (p-value = 10.2%). As expected, MBE produces a significantly positive premium of 1.6%, relatively similar to the 2.3% premium found in Bartov (2002). The price- scaled surprise (SUR) has a positive coefficient of 3.418. This coefficient means that, all else the same, an increase of 0.01 in the price-scaled surprise corresponds to an increase of 3.42% in the two-day return following the announcement. The results of this table appear to confirm the hypothesis that the focus has shifted towards analysts’ expectations. Moreover, this result validates the emphasis of this paper on the predictability of the earnings surprise (and its sign: MBE) as a means to forecast the market- adjusted returns around earnings announcements. I will conclude this section with a figure plotting the cumulative market-adjusted return (CAR) over the 2-day window (0,1) against the price-scaled surprise. Figure 3 illustrates the relationship between the return and the surprise, but also confirms one of the key points of the relationship between the surprise and the market 38 reaction: the S-shape function. The red line is a LOESS function that provides the best fit for local subsets, and it is in all aspects similar to the prospect theoretic value function: it is steeper in the negative region, it has a reference point at 0 surprise (or very slightly larger than 0) and it is steepest around this reference point. C A R (0; 1) SUR Figure 3: Scatter plot of the CAR (0,1) against the price-scaled earnings surprise. 4.2 Step A In this section, I will elaborate on the extent to which the surprise, MBE and CAR can be forecasted in the test sample. 4.2.1 Predicting MBE Table 8 below reports the coefficients obtained from the LOGIT regression (1). The number of observations included in the analysis is 3497 instead of 3664 in the test sample because 167 39 firm-quarters had missing values for one or more variables. The model is valid, as the Chi- square test produces a value of more than 350 (p<0.0001). Table 8 LOGIT regression: predicting MBE Table 8 exhibits the results of the LOGIT regression of MBE against various factors, as in regression (1) MBE Q aPOSEE Q S bPOSUE Q S cMTB Q 1 dLEV Q 1 eSIZE Q 1 fSTDEV _ PQ S gNBESTQ S hUP % Q S iDOWN % Q S jMEANREC Q S kMBEQ 1 lMBE Q 2 mMBE Q 3 nQ 2 oQ3 pQ 4 qQ5 rQ 6 sQ 7 y MBE is a dummy variable equal to one if the surprise is positive or null, the variable POSEE is a binary variable which takes a value of one if the mean of the analyst estimates is positive or zero and 0 if it is negative, POSUE is a binary variable with a value of 1 if the mean analyst expectation is larger or equal to the EPS in the same quarter of last year and 0 otherwise. MTB is the market-to-book ratio of the firm in the previous quarter, LEV is the firm’s one-quarter lagged leverage and SIZE is the logarithm of the market capitalization on the day of the previous earnings announcement. STDEV_P is the price-scaled standard deviation of the EPS estimates, NBEST is the number of EPS estimates recorded in I/B/E/S, UP% and DOWN% are the number of upward and downward revisions over the three months preceding the earnings release scaled by the number of estimates (NBEST). MEANREC is the mean analyst recommendation before the announcement, MBEL1, MBEL2 and MBEL3 are the one-, two- and three-quarter lagged MBE. Q2 is a dummy variable equal to 1 if the release was made in the second quarter of 2009, Q3 equals one if the release was made in the third quarter of 2009, Q4 has a value of 1 if the earnings were released in the fourth quarter of 2009 and so forth. MBE Coefficient S.E. Wald p-value POSEE*** ,448 ,172 6,774 ,009 POSUE*** ,896 ,108 68,213 ,000 MTB -,002 ,005 ,152 ,697 LEV -,003 ,012 ,062 ,804 SIZE ,067 ,119 ,313 ,576 STDEV_P -,353 1,156 ,093 ,760 NBEST*** ,029 ,008 13,081 ,000 UP%* ,241 ,124 3,820 ,051 DOWN%*** -,809 ,116 48,795 ,000 MEANREC -,132 ,130 1,023 ,312 MBEL1*** ,505 ,110 20,982 ,000 MBEL2* ,214 ,111 3,698 ,054 MBEL3*** ,490 ,108 20,465 ,000 Q2 ,176 ,179 ,972 ,324 Q3** ,484 ,197 6,001 ,014 Q4 -,293 ,186 2,490 ,115 Q5 -,258 ,199 1,676 ,195 Q6*** -,606 ,190 10,137 ,001 Q7*** -,597 ,186 10,285 ,001 Q8*** -,703 ,187 14,129 ,000 Constant (y) -,062 ,627 ,010 ,921 * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level 40 As can be seen in this table, the hierarchy documented by Degeorge et al. (1999) seems to hold; firms that are expected to have positive earnings (POSUE=1) or display earnings growth (POSUE=1) are more likely to meet or beat the expectations than firms that are expected to report losses or an earnings decrease. The coefficients of these variables are positive and significantly different from 0. On the contrary, I find no relation between the probability of MBE and firm characteristics. The market-to-book ratio is insignificant (in contrast to the findings of Athanasakou et al. (2009) and Bauman and Shaw (2006)). Moreover, the coefficient of MTB is negative whereas the two other studies found a significantly positive one. The variables SIZE (logarithm of the market capitalization) and LEV (leverage) are not significant either, in conflict with Matsumoto (2002) and Bauman and Shaw (2006) respectively. The insignificance of SIZE is due to its large standard error, as Matsumoto (2002) obtains a coefficient of 0.03 (0.067 here) which is significant at the 1% level (the probable reason is that Matsumoto (2002) included more than 22,000 observations in her LOGIT regression and I have around 3,500). Nevertheless, the number of estimates NBEST, which is significantly correlated with SIZE in the sample (pearson correlation of 0.37, p <0.01), significantly and positively influences the probability of MBE. In that sense, the number of estimates may be more capable than the market capitalization to capture the sensitivity effect of media coverage. My findings concerning the actions and opinions of analysts prior to the announcement are more contrasted. One the one side, my results do not corroborate the findings of Ackert and Athanassakos (1997): I find no significant relationship between the uncertainty (here proxied by the price-scaled standard deviation of the estimates) and the probability of MBE. Moreover, I do not find the relationship suggested by Abarbanell and Lehavy (2003) between the mean recommendation by analysts and MBE. This disagreement may be caused by the fact that I included the mean recommendation in the regression instead of the three categories (Buy, Hold and Sell) used in their study. Accordingly, I reran the logistic regression following their methodology (unreported), but found similar results: the mean recommendation cannot predict MBE. The significance of the UP% and DOWN% coefficients indicates that the intensity of the revisions can predict the future forecast error: UP% is positively related to the 41 probability of MBE while DOWN% is negatively and very significantly related to MBE. These results support the theory of Kim et al. (2001). Additionally there is indeed a strong autocorrelation in MBE. The one period-lagged and the three-period lagged MBE binary variables very significantly and positively influence the probability of MBE while the two period-lagged MBE variable is significant at the 10% level. The coefficient of the first lag of MBE is 0.505 while Athanakasou et al (2009) find a moderately significant coefficient of 0.32. The difference in the coefficients may be explained by the nature of the sample (Athanakasou used a UK sample). Finally, the dummy Q3 is significantly positive, which means that the frequency of MBE during this quarter (third quarter of 2009) is larger than in the first quarter of 2009 which is the base level. Q6, Q7 and Q8 have a significantly negative coefficient. The other quarter variables are not significant. Unfortunately, the independent variables in this regression are likely to be correlated. For example, I expect a strongly negative correlation between UP% and DOWN% because they represent opposite actions by the analysts. There is also probably correlation between the one- , two- and three-quarter lagged MBE variables. Table 9 displays the correlation coefficients for all the independent variables considered. Table 9 Correlation between variables in regression (1) This table shows the Pearson correlation coefficients between the variables that significantly influence MBE. POSEE POSUE NBEST UP% DOWN% MBEL1 MBEL2 MBEL3 POSEE Pearson c 1 POSUE Pearson c ,208** 1 NBEST Pearson c ,033 ,040 1 UP% Pearson c ,042 ,151** ,106** 1 ** ** ** DOWN% Pearson c -,173 -,203 -,127 -,408** 1 ** ** ** ** MBEL1 Pearson c ,201 ,149 ,087 ,199 -,308** 1 ** ** ** ** MBEL2 Pearson c ,195 ,172 ,092 ,014 -,135 ,220** 1 ** ** ** * ** MBEL3 Pearson c ,188 ,158 ,084 -,007 -,060 ,146 ,218** 1 **. Correlation is significant at the 0.01 level *. Correlation is significant at the 0.05 level As can be seen from this table, there is indeed correlation between the variables. The strongest correlation is, as expected, between UP% and DOWN%: -0.408. Moreover, we see that the extent of upward and downward revisions depend on the past MBEs: there are more upward 42 revisions and less downward revisions when a company met the earnings expectations in the previous quarter (MBEL1). Additionally, the number of estimates is positively related to the three MBE lags. Eventually, the analysts are more likely to expect earnings growth (POSUE) and positive earnings (POSEE) if the company met the earnings expectations in the past. As the extent of the correlation between UP% and DOWN% may be problematic, I report in the appendix the results of the LOGIT regression (1) once with only UP% (Table A in the appendix) and once with only DOWN% (Table B in the appendix). As can be seen from Table A, the withdrawal of DOWN% from the regression does not significantly influence the coefficient of the other variables, but the coefficient of UP% almost doubles from 0.256 to 0.504 and becomes very significant (p<0.001). Table B shows that the removal of UP% has a small effect on the coefficient of DOWN%, which moves from -0.733 to -0.870 while the other coefficients remain rather invariable. As the sign and the significance of UP% and DOWN% remain similar when one of the variables is removed, I assessed the correlation between the two variables in the LOGIT regression as not very serious, and decided to keep the model with both variables. Lastly, table 10 demonstrates the predictive ability of the LOGIT model. This table compares the actual value of MBE (rows) and the value of MBE predicted by the LOGIT model, given a cutoff point of 0.8 (the proportion of MBE in the sample as reported in table 4; if the expected probability of MBE exceeds this point, then MBE is predicted to equal 1 whereas it is predicted to be 0 if the probability is below this cutoff point). The model is 89.2% of the time right when it predicts MBE, while one would be 80% of the time right without the model, and the model is 34.4% of the time right when it predict that MBE=0 whereas one would be 20% of the time right without the model. Clearly, this LOGIT model has some predictive power. Table 10 Predictive ability of the LOGIT model Table 10 shows the predictive accuracy of the LOGIT model studied in table 9. The two rows correspond to the prediction of the LOGIT model with respect to a cutoff point of 0.8. If the LOGIT model predicts a probability of MBE that is larger than 0.8 for a certain observation, this firm-quarter will be predicted to have MBE (classified in the row MBE=1). If this probability is lower than 0.8, then it is recorded in the MBE=0 row. Table 11 compares the predictions of the LOGIT model with the observations from the sample. MBE predicted by LOGIT 0 1 43 Observed MBE 0 388 258 1 740 2132 Percentage correct 34.4% 89.2% 4.2.2 Predicting the EPS surprise In this section, I discuss the extent to which the EPS surprise is predictable using regression (2). Table 11 exhibits the results of the regression. Table 11 OLS regression: predicting the EPS surprise Table 11 exhibits the results of the following OLS regression (regression (2), discussed on page 27): SUR Q aSUR Q 1 bSUR Q 2 cSUR Q 3 cMEANREC Q S dUP % Q S eDOWN % Q S fQ 2 gQ3 hQ4 iQ5 jQ 6 kQ7 lQ8 y The surprise (SUR) is calculated as the EPS surprise divided by the price of the stock on the day of the announcement of the past quarter’s earnings. SURL1, SURL2 and SURL3 are the one-, two- and three-quarter lagged surprises. MEANREC is the mean analyst recommendation before the announcement. UP% and DOWN% refer to the number of downward revisions over the three months preceding the earnings release scaled by the number of estimates (NBEST). The seven quarter dummies are also included in this regression. SUR Coefficient S.E. t-test p-value Constant (y) ,000 ,000 ,865 ,387 SURL1*** ,103 ,013 8,100 ,000 SURL2 ,018 ,013 1,420 ,156 SURL3*** ,031 ,010 2,947 ,003 MEANREC*** ,0005 ,000 4,398 ,000 UP%*** ,0006 ,000 4,832 ,000 DOWN%*** -,0004 ,000 -3,208 ,001 Q2 ,000 ,000 ,954 ,340 Q3 ,000 ,000 1,183 ,237 Q4*** ,0005 ,000 -2,782 ,005 Q5 ,000 ,000 -1,642 ,101 Q6*** -,0005 ,000 -2,732 ,006 Q7*** -,0004 ,000 -2,593 ,010 Q8*** -,0006 ,000 -3,637 ,000 * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level The adjusted R² is as high as 0.067, meaning that the model can predict more than 6% of the variations in the EPS surprise. The one-period lagged surprise is positively and significantly related to the current period’s surprise as in Lim (2001). The two-period lagged surprise is not 44 significant and the three-lagged period surprise has a significantly positive effect on the surprise. The coefficients of UP% and DOWN% have the expected sign, consistent with Kim et al. (2001). Specifically, upward revisions are associated with larger subsequent surprises and downward revisions predict lower surprises. The mean recommendation has a coefficient sign that is opposite to the findings of Abarbanell and Lehavy (2003): companies that have a high value line score (similar to a sell recommendation) tend to have larger surprises than firms with a low Value Line score (buy recommendation). Finally, the surprises tend to be large in the fourth quarter of 2009 (dummy Q4) compared to the first quarter of 2009. One can also observe the significantly negative coefficients of the quarter dummies Q6, Q7 and Q8. 4.2.3 Predicting the market reaction Once the predicted probabilities of MBE and the predicted surprise are generated from regressions (1) and (2), I run regression (3a). I chose the time window (-2,0) as in Bernard and Thomas (1990) for the calculation of the cumulative abnormal return. The results of the regression are reported in table 12. Table 12 OLS regression: predicting the market reaction to earnings releases Table 12 exhibits the results of the OLS regression (3a): CAR Q aE ( SUR Q ) bE ( MBE Q ) cQ 2 dQ3 eQ 4 fQ 5 gQ6 hQ7 iQ8 y The cumulative abnormal return (CAR) is the return calculated for the time window (-2,0). It is obtained from adjusting the securities’ returns by the S&P500 return. The predicted surprise (E(SUR)) is the surprise that is predicted by regression (2) . The predicted probability of MBE (E(MBE)) is the predicted probability of MBE by regression (1). The seven quarter dummies are also included in this regression. CAR Coefficient Std. Error t-test p-value Constant (y)*** ,019 ,007 2,663 ,008 E(SUR) -1,658 1,917 -,865 ,387 E(MBE) ,013 ,010 1,301 ,193 Q2*** -,024 ,004 -5,943 ,000 Q3*** -,024 ,004 -5,961 ,000 Q4*** -,021 ,004 -5,143 ,000 Q5*** -,024 ,004 -5,705 ,000 Q6*** -,030 ,004 -7,168 ,000 Q7*** -,029 ,004 -7,088 ,000 Q8*** -,026 ,004 -6,282 ,000 * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level 45 The R² of the model is 2.2%. Nevertheless, the results suggest that the market reaction to the earnings announcements using the predicted surprise and the predicted probability of MBE is hardly predictable; the predicted surprise and the predicted MBE probability that were generated from regressions (2) and (1) are not significant. E(SUR) and E(MBE) are positively and significantly correlated with a Pearson correlation coefficient of 0.273. All the quarter dummies are negatively and significant with roughly equal coefficients, which indicates that the sample of firms experienced considerably high abnormal returns around the earnings releases during the first quarter of 2009 (base level), but remained approximately flat afterwards. The results from this regression may give a first insight into the exactness of traders’ expectations. If we assume that the regression (2) is the best model to form expectations for surprises, then the results suggest that the stock traders appropriately incorporate the driving forces behind the surprise into their behavior around the earnings release. The coefficient of the expected probability of MBE is positive but insignificant, again insinuating that the investors react in an efficient manner to the EPS surprise (here, its sign). This may signify that the market reacts to the expected surprise before the announcement and price the information accordingly. The results of the regression between the cumulative abnormal return over the five trading days before the announcement and E(SUR) and E(MBE) seem to validate such a hypothesis. Table C in the appendix reports the results of this regression, which show that the expected surprise is significantly and positively related to the cumulative abnormal return over the time window (-5;-1). Investors may indeed anticipate the surprise, and partly shortly before the release. Moreover, I find that the top expected surprise quintile exhibits a CAAR (- 5,-1) of 0.8% (median of 0.8%) while the bottom quintile earns a CAAR of 0.45% (0.51%). Those results will be discussed in more detail in section 4.3.2. According to Bernard and Thomas (1990), the market reaction can be predicted by using the past EPS surprises. I will examine this issue by examining the coefficients obtained from regression (3a’) which are displayed in table 13. Table 13 OLS regression: predicting the market reaction to earnings releases using the past surprises 46 Table 13 exhibits the results of the OLS regression (3a’) CAR Q aSUR Q 1 bSUR Q 2 cSUR Q 3 dQ2 eQ3 fQ 4 gQ5 hQ6 iQ7 jQ 8 y The cumulative abnormal return (CAR) is the cumulative abnormal return calculated for the time window (- 2;0). It is obtained from adjusting the securities’ returns by the S&P500 return. The three lags of the price- scaled surprises (SURL1, SURL2 and SURL3) are included in the regression along with the seven dummies. CAR Coefficient Std. Error t-test p-value Constant (y)*** ,027 ,003 9,400 ,000 SURL1 -,080 ,270 -,294 ,768 SURL2** -,563 ,268 -2,102 ,036 SURL3 -,068 ,209 -,327 ,744 Q2*** -,023 ,004 -5,693 ,000 Q3*** -,023 ,004 -5,809 ,000 Q4*** -,019 ,004 -4,940 ,000 Q5*** -,021 ,004 -5,287 ,000 Q6*** -,027 ,004 -7,032 ,000 Q7*** -,027 ,004 -6,843 ,000 Q8*** -,023 ,004 -6,011 ,000 The results clearly indicate that the past surprises are not able to predict the future market reaction to the earnings releases, apart from the two-period lagged surprise which has a negative coefficient. Given that the one-period lagged is insignificant and is the one supposed to have the strongest effect (Bernard and Thomas, 1990), it is hard to believe that the two- period lagged surprise has a real influence on the market reaction. The coefficient and the significance level may just be the product of chance. All in all, the relationship proposed by Bernard and Thomas (1990) seems to not hold anymore in my recent sample. 4.2.4 Summary of the results I have found that MBE can be predicted using several factors. Indeed, the expected Earnings (POSEE), the expected growth in earnings (POSUE), the analyst coverage (NBEST), the extent of upward and downward revisions (UP% and DOWN%) as well as the MBE history are powerful predictors of the probability of MBE. The results of the LOGIT regression may be useful, given that they provide some predictive ability: without the model, the best guess of an investor would be that all observations meet or beat the expectations (MBE=1) and he would be right about 80% of the time whereas with the model, the investor can improve the 47 percentage of exactness to almost 90% (when the model predicts that the probability of MBE is larger than 80%). The EPS surprise can also be anticipated. Together with seven quarter dummies, UP%, DOWN%, MEANREC and the three surprise lags can explain 6.7% of the variations in the earnings per share surprise. The one-quarter lagged surprise is particularly strong with a coefficient of 0.1 and a t-test of more than 8. However, the results from the two above-mentioned studies have a very limited utility: the predicted probability of MBE and the predicted EPS surprises cannot explain the returns around the earnings release at all. This may signify that the investors properly integrate the autocorrelation in surprises and the trends in analysts’ expectations to form their reaction to the earnings release. The final test is based on the findings of Bernard and Thomas (1990) who found that the returns around the earnings disclosures can be predicted by the past earnings surprises thanks to the positive autocorrelation between the forecast errors. The results of regression (3a’) appear to challenge this finding: only the two-lagged surprise is significant, and it has a negative coefficient (while the expected sign was positive). 4.3 Step B 4.3.1 Strategy based on MBE As I already elaborated on, I will use the quintiles method (Collins and Hribar, 2000) to form the trading strategy. In this strategy, I propose to buy the top expected MBE probability quintile and sell the bottom one. Table 14 illustrates the cumulative average abnormal return (CAAR) for the lowest and highest quintiles with respect to the predicted probability MBE for the time window (-2,0) (as in Bernard and Thomas, 1990). Table 14 Trading strategies based on the predicted probability of MBE Table 14 exhibits the CAR features on the quintiles formed on the basis of the expected probability of MBE that is produced from regression (1). The mean CAR is called CAAR (cumulative average abnormal return). The table also features the data about the full sample. CAAR St. Deviation Minimum Maximum Q1 Median Q3 48 Quintile 1 ,0046 ,0723 -,35 ,31 -,0335 ,0000 ,0390 Quintile 5 ,0047 ,0552 -,25 ,26 -,0245 ,0026 ,0348 Full sample ,0056 ,0586 -,35 ,36 -,0260 ,0026 ,0354 The CAARs in the lowest and largest quintiles are both positive but smaller than in the full sample. The CAARs for the two subsamples are remarkably close to each other (0.46% and 0.47%), which means that the trading strategy buying the top quintile and selling the bottom one will definitely not be profitable. The results of this table suggest that, in line with the findings in table 12, the predictability of MBE does not help earn abnormal returns. Consequently, we can suppose that the market anticipates the probability of MBE correctly. Still, paying attention to the quintiles may be highly relevant: the first quintile’s CAR is much more volatile than in the fifth quintile: the volatility in the bottom quintile is 7.24% while it is only 5.55% in the top one, which is even lower than the volatility of the whole sample (5.87%). The range in the fifth quintile is also substantially smaller than in the two other samples. Accordingly, the risk-averse investors may use regression (1) to invest in the companies that are more likely to meet or beat the expectations (e.g. quintile 5) because they carry less risk during the earnings announcement period. 4.3.2 Strategy based on the predicted EPS surprise Here, I test if the 20% of observations with the highest expected EPS surprise outperform the bottom 20%. The results are displayed in table 15. Table 15 Trading strategies based on the predicted value of EPS surprise Table 15 displays the CAR features on the quintiles formed on the basis of the predicted EPS surprise that is produced from regression (2). The mean CAR is called CAAR (cumulative average abnormal return). CAAR St. Deviation Minimum Maximum Q1 Median Q3 Quintile 1 ,0055 ,05747 -,33 ,27 -,0232 ,0028 ,0321 Quintile 5 ,0034 ,06083 -,24 ,26 -,0335 -,0001 ,0372 Full sample ,0056 ,05866 -,35 ,36 -,0260 ,0026 ,0354 The cumulative average abnormal return and the median CAR of the fifth quintile are smaller than in the first quintile and in the whole sample. This observation could be due to an 49 “asymmetric reaction” to the surprises for the fifth quintile: if investors account for the autocorrelation in surprises, they expect the firm-quarters that have exhibited positive surprises in the past to keep the same pace of positive surprises (as the coefficients in regression (2) suggest). Consequently, the investors would shift their focus from the analyst expectations (surprise of 0 on which MBE is based) towards the EPS surprise expectation (obtained from a model similar to regression (2))! In the event where a firm with a history of positive past surprises discloses a surprise lower than that predicted by the model (even if it is positive), the market may be disappointed. Worse: if the surprise is negative, the market would severely punish the firm. Conversely, if the company once again achieves a surprise better than expected, the market would strongly reward the firm. This hypothetical explanation may explain the lower CAAR, the higher standard deviation and the larger interquartile range observed for quintile 5. From table C in the appendix, we have seen that the return over the five days before the earnings release are positively related to the expected surprise, suggesting that the investors anticipate the surprise. As the relationship is significant, let us analyze the CAAR over the time window (-5;-1) for the first and the fifth expected EPS surprise quintile. Table 16 shows the results. Table 16 Trading strategies based on the SUR predicted from regression (2) Table 16 displays the CAR (-5,-1) features on the quintiles formed on the basis of the predicted surprise that is produced from regression (2). The mean CAR is called CAAR (cumulative average abnormal return). CAAR St dev. Q1 Median Q3 Quintile 1 ,0045 ,0367 -,0157 ,0051 ,0228 Quintile 5 ,0083 ,0358 -,0092 ,0081 ,0268 Full sample ,0071 ,0372 -,0119 ,0064 ,0244 We indeed see that the 20% observations with the highest expected surprise earn a cumulative abnormal return of 0.83% on average over the five trading days preceding the earnings release, while the 20% firms with the lowest expected surprise earn a 0.45% CAAR. The difference in CAAR between the first and the fifth quintiles is significant at the 10% level (p=0.058). This indeed suggests that the investors anticipate the surprise, and explain why the expected surprise cannot explain the returns that follow the earnings announcement. 50 Finally, I display in table 17 the CAR distribution of the bottom and top expected CAR quintiles. These expected CAR values were calculated using the coefficients of E(SUR) and E(MBE) from table 12, without the quarter dummy variables. Table 17 Trading strategies based on the CAR predicted from regression (3a) Table 17 displays the CAR features on the quintiles formed on the basis of the predicted CAR that is produced from regression (2). The mean CAR is called CAAR (cumulative average abnormal return). CAAR St. Deviation Minimum Maximum Q1 Median Q3 Quintile 1 ,0043 ,0646 -,24 ,27 -,0340 -,0009 ,0367 Quintile 5 ,0029 ,0492 -,24 ,22 -,0215 ,0032 ,0283 Full sample ,0056 ,0536 -,35 ,36 -,0260 ,0026 ,0354 Once more, it is not possible to develop a profitable zero-cost strategy, as the returns are roughly akin. However, we find again the relatively low volatility in returns for the fifth quintile, probably because of the inclusion of the predicted MBE variable which helps choose the stocks that are less volatile around the earnings announcements (see table 14). I will not test the profitability of the trading strategy based on regression (3a’) because the lagged surprise variables could not predict the market reaction. 4.4 Step C: Comparison of the results: holdout sample versus test sample There are two main findings in this thesis. First, MBE and the surprises are predictable. The main drivers of both variables are their lagged values (autocorrelation) and the trends in analysts’ expectations (UP% and DOWN%). Second, the strategies based on quintiles of predicted MBE, surprise and market reaction are not profitable, but there are large discrepancies in volatility of the returns between quintiles, especially in the quintiles with respect to the predicted probability of MBE. Accordingly, this section will investigate whether (1) the drivers of MBE and the surprise remain significant in the holdout sample and (2) if the top predicted MBE quintile keeps its relatively low volatility in the holdout sample. Let us first compare the coefficients of the variables in table 8 with those that are obtained from the running of the LOGIT regression on the holdout sample. 51 Table 18 Comparison of coefficients: LOGIT This table shows the coefficients and the significance levels for the variables used in regression (1) for the test sample (2009-2010) and the holdout sample (2011 Q1,Q2 and Q3). POSEE POSUE UP% DOWN% NBEST MBEL1 MBEL2 MBEL3 Test sample 0.505*** 0.884*** 0.256** -0.733*** 0.029*** 0.503*** 0.207* 0.484*** Holdout sample 0.623* 1.370*** -0.13 -0.532*** 0.020** 0.361** 0.470*** 0.400** * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level As can be seen from the table above, the relationship that existed between MBE and its drivers over the period 2009-2010 tends to persist over time, as the third row of the table shows. In fact, all the variables except UP% keep the same coefficient sign, and remain significant at the 10% level at least. This observation suggests that the relationships found in this thesis are likely to continue over the near term and are not only due to the distinctiveness of the test sample’s time period. Secondly, I report the coefficients of regression (2) tested on the test and the holdout sample in table 19. All the variables indeed remain significant and keep the same sign in the holdout sample. Table 19 Comparison of coefficients: OLS regression This table shows the coefficients and the significance levels for the variables used in regression (2) for the test sample (2009-2010) and the holdout sample (2011 Q1,Q2 and Q3). SURL1 SURL2 SURL3 MEANREC UP% DOWN% Test sample 0.103*** 0.013 0.031*** 0.0005*** 0.0006*** -0.0004*** Holdout sample 0.126*** 0.070** 0.053** 0.0003** 0.0007*** -0.0004*** * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level Finally, the volatility of the returns around the earnings announcement in the fifth quintile with respect to the predicted probability of MBE remains significantly low in the holdout sample. In the test sample, the volatility in the top quintile was 0.3% lower than in the whole sample and 1.7% lower than in the bottom quintile (table 14); in the holdout sample, the volatility is 0.4% lower than in the whole sample and 1.1% lower than in the first quintile (unreported table). This suggests that running regression (1) and selecting stock accordingly 52 indeed may enable the stock trader to pick stocks with lower volatility but equal expected returns around the earnings announcement. 4.5 Discussion: what really drives the market reaction? In section 4.3, I have noticed the inability of the surprise-predicting models to produce abnormal returns after the earnings release. The expected surprise thus cannot predict the return following the release: the stock operators are aware of the drivers of the EPS and incorporate them in their behavior to formulate their reaction to the earnings. The results in table C in the Appendix and table 16 seem to corroborate this hypothesis by showing that the expected surprise is positively related to the abnormal return over the five trading days preceding the release. I thus propose a novel scheme to explain the market reaction to the earnings announcement: investors price the stock depending on the expected surprise they obtain from a predicting model (similar to regression (1) and (2)) before the announcement and then, after the release, they react to the difference between the actual surprise and the expected surprise (excess surprise), and not according to the 0 surprise level. This hypothesis could explain (1) the fact that some positive surprises are associated with negative abnormal returns, (2) that negative surprises are sometimes associated with a positive market reaction, (3) the abnormally high return shortly before the announcement for those stocks that are expected to exhibit a positive surprise and (4) the observation that the one-day CAAR (0) for companies with a slightly positive surprise is negative (see table D in the Appendix). To test the validity of such a framework, I develop a new variable called MBES which stands for “meet or beat the expected surprise”. It obtains a value of 1 whenever the company meets or beats the expected surprise obtained from regression (2). Moreover, I create the variable DSUR which is equal to the difference between the actual surprise and the surprise which is expected from regression (2). The hypothesis that I suggested states that the market reacts to MBES instead of MBE (whose reference surprise is 0) and to DSUR instead of SUR. In that view, I regress linearly the market reaction against the two sets of variables and compare the R² of e ach model. The first table relates the (-2,0) abnormal return with MBE and SUR, as in Bartov et al. (2002). 53 Table 20 Drivers of the market reaction to earnings announcements: standard model The table exhibits the results of the regression: CARQ aSURQ bMBEQ y CAR is the cumulative abnormal return over the (-2,0) time window. SUR is the price-scaled surprise observed at time 0, MBE is a binary variable equal to 1 if the surprise is nonnegative and 0 otherwise. CAR(-2,0) Coefficient Std. Error t-test p-value Constant (y) -,010 ,002 -4,068 ,000 MBE ,009 ,003 2,842 ,005 SUR 5,514 ,455 12,121 ,000 In this first regression, the portion of the market reaction over the time window (-2;0) that is explained (R²) is 7.6%. The two variables are significantly positive, in line with Bartov et al. (2002) and the results I have observed before. The table that follows includes the variables MBES and DSUR that I proposed. Table 21 Drivers of the market reaction to earnings announcements: excess surprise model The table exhibits the results of the regression: CARQ aDSUR Q bMBES Q y CAR is the cumulative abnormal return over the (-2,0) time window. DSUR is the difference between the price- scaled surprise and the price-scaled surprise estimated from regression (2), MBES is a binary variable equal to 1 if DSUR is positive and 0 otherwise. CAR(-2,0) Coefficient Std. Error t-test p-value Constant (y) ,000 ,001 ,139 ,890 MBES ,012 ,003 4,546 ,000 DSUR 5,338 ,581 9,192 ,000 In this regression, the R² is as high as 9.9%. This suggests that indeed, the “excess surprise” model probably better explains the market reaction than the standard model (MBE + absolute surprise) studied in the past literature. When I include the binary variable MBE into this regression (not reported), I obtain a significantly positive coefficient (0.008) for this variable. Consequently, meeting or beating the estimates still produces a premium return of 0.8%, while the premium to meeting or beating the estimated surprise (MBES) stays at 1.2%. If I use a LOGIT model to explain the sign of the cumulative abnormal return, the excess surprise model I propose offers again a better fit: the model with MBES and DSUR has a Nagelkerke R² of 10.9% while the model with SUR and MBE has a Nagelkerke of 7.0%. As a final test for the model I propose, I compare the capability of MBE and MBES to explain the sign of the CAR over the time window (-2,0). As can be seen in the following table, there is a 54 relationship between the sign of the CAR (-2,0) and both MBE and MBES. Yet, if we sum up the percentages of the cases where MBE and MBES have the same value as CAR (-2,0), MBES explains correctly the sign of the return in 60.13% of the cases while MBE does it in 57.80% of the cases. It seems that MBES can better explain the negative returns, while MBE is better at explaining the positive return. Consequently, the usage of both MBE and MBES to explain the market reaction is probably superior to the model using only one of the variables. I reach the same conclusion in a LOGIT model (unreported): both MBE and MBES are significantly and positively related to the probability that CAR (-2,0) is positive. Table 22 Drivers of the market reaction to earnings announcements (2) This table shows the frequency distribution between the sign of the CAR (-2,0) and MBE and MBES. The “accuracy” row is the sum of the percentages where MBES or MBE have the binary variable “POS RETURN". MBES MBE 0 1 0 1 1354 410 449 1315 0 POS. 36.95% 11.19% 12.25% 35.89% RETURN 1051 849 231 1669 1 28.68% 23.18% 6.30% 45.55% Accuracy 60.13% 57.80% All in all, it seems that the market reaction to earnings announcements is not only related to the surprise and MBE, whose reference surprise is 0, but also to the excess surprise (actual surprise versus the expected surprise) and its sign (MBES). This expected surprise can be obtained from regressions such as regression (2) in this thesis, in which the investors recognize the true drivers of the surprise, including surprise autocorrelation. The model that includes the excess surprise does a better job at explaining the market reaction than the regression with the absolute surprise: the portion of the CAR variations that is explained indeed rises by 2.3%, from 7.6% with the absolute surprise model (standard model) to 9.9% with the excess surprise model. The inability of the expected surprise to generate a profitable strategy further corroborates the idea that the predicted surprise is already integrated in traders’ mind, and is consistent with the semi-strong form of market efficiency. Moreover, it seems that they anticipate the expected surprise over the few days before the earnings release. 55 To my knowledge, this relation between the market reaction and the excess surprise has not yet been discussed in the literature. Lastly, I tested regressions (1) and (2) on the excess surprise to investigate whether the excess surprise is predictable, and the results are provided in Table E and F in the appendix. Indeed, the excess surprise can be predicted to some extent. For example STDEV_P, the price-scaled standard deviation, can predict the excess surprise and it sign. Then, I redid the quintile strategy proposed in subquestions 1b and 2b to study the usefulness of these predictabilities. The results are reported in table G of the appendix. The top quintiles for both the expected probability of MBES and the expected excess surprise have a CAAR superior to 1.3% and the bottom quintiles exhibit a negative CAAR. A zero-cost strategy buying the top quintile and selling the bottom could thus generate significant abnormal returns, thanks to the predictability of the excess surprise. 5. Conclusion In line with the trends in the literature, I find that the EPS surprise and its sign are the most important drivers of the market reaction to the release of results. Specifically, firms that meet or beat the expectations are estimated to receive a 1.8% positive premium on their stock price, while the earnings surprise is strongly and positively related to the abnormal return. I also find that these surprises and the probability of MBE are predictable, consistent with the prior papers that studied the issue. The probability that the firm meets or beats the forecasts of analysts is an increasing function of its past MBE history (three MBE lags), the number of forecasts, the degree of upward revisions and the analysts’ expected sign of earnings (POSEE) and growth in earnings (POSUE). It is negatively related to the extent of downward revisions prior to the release. In addition, I find strong autocorrelation between the surprises (coefficient of 0.1 for the first-order autocorrelation) and a relation between the forecast revisions and the surprise which is conform to the theory of Kim et al. (2001): upward revisions tend to be followed by more positive surprises and the downward revisions are disposed to predict lower surprises. These relations are not due to the uniqueness of the sample’s time period, given that I obtain similar coefficients and significance levels in a holdout sample. I have shown that the sign and the extent of the EPS surprise are predictable. Nonetheless, my results suggest that this predictability cannot help anticipate the market reaction to the 56 earnings release. Indeed, the surprise that is expected from my model and the predicted sign of the surprise (MBE) are unable to explain the market reaction after the earnings announcement. Moreover, I find that the returns to the earnings announcements cannot be anticipated by using the past EPS surprises, as Bernard and Thomas (1990) proposed. Additionally and in line with all these observations, I found that one cannot earn abnormal returns by selecting the extreme expected surprise quintiles, in which the investor buys the top expected surprise quintile and sells the bottom one. The other quintile strategies based on the expected sign of the surprise and the expected abnormal return provided similarly disappointing results. Still, I find that the 20% observations whose predicted probability of MBE is highest display much less volatile returns than the rest of the sample for the same expected return. This could be of particular interest for the risk-averse investors as well as for the “sensation seekers”. Accordingly, I suggest that my sample leans towards the semi-strong form of market efficiency: on average, investors understand the driving forces behind the EPS surprises and its sign and form expectations consequently; the expected surprise is already priced in. I also find that the cumulative abnormal returns over the five trading days preceding the earnings announcement are positively related to the expected surprise generated from my model, which may mean that the market anticipates the surprise, and partly does it shortly before the announcement. As the expected surprise is already “priced in”, I propose that the market responds to the EPS surprise relative to the expected surprise (the expected surprise being for example generated from a OLS regression as the one I used) and its sign and not to the absolute EPS surprise. The focus would thus shift from the absolute value of the surprise towards the surprise in excess of the expected surprise (here called the excess surprise). I find that the model including the excess surprise and its sign explains more variations in the CAR (-2,0) than does the model with the absolute surprise and its sign (MBE). Moreover, I discover that the excess surprise is better than the absolute surprise at explaining the sign of the market reaction after the announcement. Finally, I find that investors may benefit from the predictability of the excess surprise by selecting the extreme expected excess surprise quintiles, in sharp contrast with my findings about the usefulness of the absolute surprise predictability. I would suggest that the future research as well as the financial press acknowledge these new dynamics. Specifically, one should keep in mind that meeting or beating the analyst estimates remains relevant, but the market also reacts to its own expectations. 57 6. References Abarbanell, J. and Bernard, V., 1992. “ Tests of analysts’ overreaction/underreaction to earnings information as an explanation for anomalous stock price behavior”, Journal of Finance, 47: 1181-1207. Abarbanell, J. and Bushee, B., 1997. “Fundamental analysis, Future earnings and Stock prices”, Journal of Accounting Research, 35: 1-24 Abarbanell, J. and Lehavy, R., 2003. “Can stock recommendations predict earnings management and analysts’ earnings forecast errors?”, Journal of Accounting Research, 41: 1– 31. Ackert, L. and Athanassakos, G., 1997. “Prior uncertainty, analyst bias and subsequent abnormal returns”, Journal of Financial Research, 20: 263-273 Ackert, L. and Hunter, W, 1994. “Rational expectations and the dynamic adjustment of security analysts’ forecasts to new information”, Journal of Financial Research, 17: 387-401. Ackert, L. and Hunter, W., 1995. “Rational expectations and security analysts’ forecast earnings forecasts”, The Financial Review, 30: 427-443. Amir, E., and B. Lev., 1996. “Value-relevance of nonfinancial information: The wireless communica-tions industry”, Journal of Accounting and Economics, 22: 3-30. Athanasakou, E., Strong, C. and Walker, M., 2009. “Earnings management or forecast guidance to meet analyst expectations?”, Accounting and Business Research, 39: 3-35 Bartov, E., 1992. “Patterns in Unexpected Earnings as an Explanation for Post-Announcement Drift”, The Accounting Review, 67: 610-622 Bartov, E., Givoly, D. and Hayn, C., 2002. “The rewards to meeting or beating earnings expectations”, Journal of Accounting and Economics, 33: 173–204. Bauman, M. and K. Shaw, 2006. "Stock option compensation and the likelihood of meeting analysts' quarterly earnings targets." Review of Quantitative Finance and Accounting, 26 : 301-319. Bernard, V.L., and Thomas, J.K., 1990. “Evidence that stock prices do not fully reflect the implications of current earnings for future earnings”. Journal of Accounting and Economics, 13: 305–40. Bhushan, R., 1989. “Firm characteristics and analyst following”. Journal of Accounting and Economics, 11: 255-274. Brown, L. and M. Rozeff (1978), “The superiority of analyst forecasts as measures of expectations: Evidence from earnings”. Journal of Finance 33, 1-16. Brown, S. and Warner, J., 1985. “Using Daily Stock Returns, the Case of Event Studies”, Journal of Financial Economics, 14: 3-31. 58 Brown, L.D., 1997. “Analyst Forecasting Errors: Additional Evidence”, Financial Analysts Journal, 53: 81–88. Brown, L.D. and Caylor, M.L., 2005. “A temporal analysis of earnings management thresholds: propensities and valuation consequences”, The Accounting Review, 80: 423–440. Burgstahler, D. and Dichev, I., 1997. “Earnings management to avoid earnings decreases and losses”, Journal of Accounting and Economics, 24: 99–126. Burgstahler, D. and M. Eames, 2003. “Earnings Management to Avoid Losses and Small Decreases: Are Analysts Fooled?”, Contemporary Accounting Research, 20: 253–94. Chan, K., Karceski, J. and J. Lakonishok, 2003. “The Level and Persistence of Growth Rates”, Journal of Finance, 58: 634-684. Christie, A, 1987. “On cross-sectional analysis in accounting research”, Journal of Accounting and Economics, 9: 231–258. Collins, D. and Hribar, P., 2000. “Earnings-based and accrual-based market anomalies: one effect or two?”, Journal of Accounting and Economics, 29 : 101-123 Das, S., C. B. Levine, and K. Sivaramakrishnan, 1998. “Earnings predictability and bias in analysts' earnings forecasts”, The Accounting Review, 73: 277-294. De Bondt, W.F.M., Thaler, R.H., 1985. “Does the Stock Market Overreact?”, The Journal of Finance, 40: 793-805. Dechow, P., Richardson, S. and Tuna, I., 2003. “Why are Earnings Kinky? An Examination of the Earnings Management Explanation”, Review of Accounting Studies, 8: 355–384. Degeorge, F., Patel, J. and Zeckhauser, R., 1999. “Earnings management to exceed thresholds”, Journal of Business, 72: 1–33. Dalhiwal, D., Lee, K., and Fargher, N., 1991. “The association between unexpected earnings and abnormal security returns in the presence of financial leverage”, Contemporary Accounting Research, 10: 20-41 Dugar, A. and Nathan, S, 1995. The effect of investment banking relationships on financial analysts' earnings forecasts and investment recommendations. Contemporary Accounting Research, 12: 131-160. Fama, E., and K. French. 1992. “The cross-section of expected stock returns”, Joumal of Finance, 47: 427-465. Foster, G., Olsen, C. and Shevlin, T., 1984. “Earnings releases, anomalies, and the behavior of security returns”, The Accounting Review, 59: 574-603. 59 Freeman, R., and Tse, S., 1992. “A nonlinear model of security price response to unexpected earnings”, Journal of Accounting Research, 30: 185-209. Graham, J.R., Harvey, C.R. and Rajgopal, S., 2005. “The economic implications of corporate financial reporting”, Journal of Accounting and Economics, 40: 3–73. Hribar, P., Jenkins, N.T., Johnson, W.B., 2006. “Stock repurchases as an earnings management Device”, Journal of Accounting and Economics, 41: 3–27 Jegadeesh, J and S. Titman, 1993. “Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency”, The Journal of Finance, 48: 65-91. Kahneman, D., and Tversky, A., 1979. “Prospect theory: An analysis of decision under risk”, Econometrica, 47: 263–91. Kasznik, R., Lev, B., 1995. “To warn or not warn: management disclosures in the face of an earnings surprise”, The Accounting Review, 70: 113–134. Kim, O., S. Lim and K. Shaw, 2001. “The Inefficiency of the Mean Forecast as a Summary Forecast of Earnings”, Journal of Accounting Research, 39: 329–336 Kinney, W., D. Burgstahler, and R. Martin, 2002. “Earnings Surprise “Materiality” as Measured by Stock Returns”, Journal of Accounting Research, 40: 1297-1329. Klibanoff, P., Lamont O. and T.A. Wizman, 1998. “Investor Reaction to Salient News in Closed-End Country Funds”, The Journal of Finance, 53: 673-699. Lev, B. and Thiagarajan, S., 1993. “Fundamental Information Analysis”, Journal of Accounting Research, 31: 190-215 Lim T., 2001. “Rationality and Analysts' Forecast Bias”, The Journal of Finance, 56: 369-385 Lin, S., Radhakrishnan, S. and Su, L., 2006, “Earnings management and guidance for meeting or beating analysts’ earnings forecasts”, Working Paper, University of Texas. Mendenhall, R., 1991. “Evidence of possible underweighting of earnings-related information”, Journal of Accounting Research, 29: 170-180. O’Brien, P.C. and Bhushan, R., 1990. “Analyst following and institutional ownership”, Journal of Accounting Research, 28: 55-76. O’Brien, P.C, 1994. “Corporate Earnings Forecasts ane the Macroeconomy”, Working Paper, University of Waterloo. O’Brien , P., McNichols , M. and Lin, H., 2005. “Analyst Impartiality and Investment Banking Relationships”, Journal of Accounting Research, 43: 623-650. Ou, J. A., and S. H. Penman, 1989. “Financial Statement Analysis and the Prediction of Stock Returns”, Journal of Accounting and Economics 11: 295– 329. 60 Matsumoto, D.A., 2002. “Management’s incentives to avoid negative earnings surprises”, Accounting Review, 77: 483–514. Matsunaga, S., Park, C., 2001. “The effect of missing a quarterly earnings benchmark on the CEO’s annual Bonus”, The Accounting Review, 76: 313–332. McVay, S., V. Nagar, and V. W. Tang. 2006. “Trading incentives to meet the analyst Forecast”, Review of Accounting Studies, 11: 575–598. Penman, S. and Sougianis, T., 1998. “A Comparison of Dividend, Cash Flow, and Earning Approaches to Equity Valuation”, Contemporary Accounting Research, 15: 343-383 Puffer, S. M., and J. B. Weintrop, 1991. “Corporate performance and CEO turnover: The role of performance expectations”, Administrative Science Quarterly, 36: 1-19. Skinner, D.J. and Sloan, R.G., 2002. “Earnings surprises, growth expectations and stock returns or don’t let an earnings torpedo sink your portfolio”, Review of Accounting Studies, 7: 289–312. 61 7. Appendix Table A LOGIT regression: predicting MBE without DOWN% Table A exhibits the results of the LOGIT regression(1) without DOWN% MBEQ aPOSEE Q S bPOSUE Q S cMTBQ 1 dLEVQ 1 eSIZE Q 1 fSTDEV _ PQ S gNBESTQ S hUP %Q S jMEANREC Q S kMBEQ 1 lMBEQ 2 mMBE Q 3 nQ 2 oQ3 pQ 4 qQ5 rQ 6 sQ 7 y MBE is a dummy variable equal to one if the surprise is positive or null, the variable POSEE is a binary variable which takes a value of one if the mean of the analyst estimates is positive or zero and 0 if it is negative, POSUE is a binary variable with a value of 1 if the mean analyst expectation is larger or equal to the EPS in the same quarter of last year and 0 otherwise. MTB is the market-to-book ratio of the firm in the previous quarter, LEV is the firm’s one-quarter lagged leverage and SIZE is the logarithm of the market capitalization on the day of the previous earnings announcement. STDEV_P is the price-scaled standard deviation of the EPS estimate, NBEST is the number of EPS estimates recorded in I/B/E/S, UP% is the number of upward revisions over the three months preceding the earnings release scaled by the number of estimates (NBEST). MEANREC is the mean analyst recommendation before the announcement, MBEL1, MBEL2 and MBEL3 are the one-, two- and three-quarter lagged MBE. Q2 is a dummy variable equal to 1 if the release was made in the second quarter of 2009, Q3 equals one if the release was made in the third quarter of 2009, Q4 has a value of 1 if the earnings were released in the fourth quarter of 2009 and so forth. Coefficient S.E. Wald Sig. POSEE ,557 ,171 10,662 ,001 POSUE ,975 ,108 81,368 ,000 MTB -,002 ,005 ,120 ,730 LEV -,006 ,013 ,242 ,623 SIZE -,076 ,117 ,425 ,514 STDEV_P -,826 1,293 ,408 ,523 NBEST ,033 ,008 17,690 ,000 UP% ,504 ,132 14,625 ,000 MEANREC -,187 ,129 2,110 ,146 MBEL1 ,643 ,108 35,546 ,000 MBEL2 ,233 ,110 4,472 ,034 MBEL3 ,475 ,108 19,473 ,000 Q2 ,202 ,177 1,306 ,253 Q3 ,595 ,196 9,248 ,002 Q4 -,096 ,183 ,273 ,601 Q5 -,098 ,197 ,250 ,617 Q6 -,540 ,190 8,101 ,004 Q7 -,563 ,185 9,286 ,002 Q8 -,611 ,185 10,914 ,001 Constant (y) -,386 ,621 ,386 ,535 62 Table A LOGIT regression: predicting MBE without DOWN% Table A exhibits the results of the LOGIT regression(1) without DOWN% MBEQ aPOSEE Q S bPOSUE Q S cMTBQ 1 dLEVQ 1 eSIZE Q 1 fSTDEV _ PQ S gNBESTQ S hUP %Q S jMEANREC Q S kMBEQ 1 lMBEQ 2 mMBE Q 3 nQ 2 oQ3 pQ 4 qQ5 rQ 6 sQ 7 y MBE is a dummy variable equal to one if the surprise is positive or null, the variable POSEE is a binary variable which takes a value of one if the mean of the analyst estimates is positive or zero and 0 if it is negative, POSUE is a binary variable with a value of 1 if the mean analyst expectation is larger or equal to the EPS in the same quarter of last year and 0 otherwise. MTB is the market-to-book ratio of the firm in the previous quarter, LEV is the firm’s one-quarter lagged leverage and SIZE is the logarithm of the market capitalization on the day of the previous earnings announcement. STDEV_P is the price-scaled standard deviation of the EPS estimate, NBEST is the number of EPS estimates recorded in I/B/E/S, UP% is the number of upward revisions over the three months preceding the earnings release scaled by the number of estimates (NBEST). MEANREC is the mean analyst recommendation before the announcement, MBEL1, MBEL2 and MBEL3 are the one-, two- and three-quarter lagged MBE. Q2 is a dummy variable equal to 1 if the release was made in the second quarter of 2009, Q3 equals one if the release was made in the third quarter of 2009, Q4 has a value of 1 if the earnings were released in the fourth quarter of 2009 and so forth. Coefficient S.E. Wald Sig. POSEE ,557 ,171 10,662 ,001 POSUE ,975 ,108 81,368 ,000 MTB -,002 ,005 ,120 ,730 LEV -,006 ,013 ,242 ,623 SIZE -,076 ,117 ,425 ,514 STDEV_P -,826 1,293 ,408 ,523 NBEST ,033 ,008 17,690 ,000 UP% ,504 ,132 14,625 ,000 MEANREC -,187 ,129 2,110 ,146 MBEL1 ,643 ,108 35,546 ,000 MBEL2 ,233 ,110 4,472 ,034 MBEL3 ,475 ,108 19,473 ,000 Q2 ,202 ,177 1,306 ,253 Q3 ,595 ,196 9,248 ,002 Q4 -,096 ,183 ,273 ,601 Q5 -,098 ,197 ,250 ,617 Q6 -,540 ,190 8,101 ,004 Q7 -,563 ,185 9,286 ,002 Q8 -,611 ,185 10,914 ,001 Constant (y) -,386 ,621 ,386 ,535 * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level 63 Table B LOGIT regression: predicting MBE without UP% Table A exhibits the results of the LOGIT regression(1) without UP% MBEQ aPOSEE Q S bPOSUE Q S cMTBQ 1 dLEVQ 1 eSIZE Q 1 fSTDEV _ P S Q gNBESTQ S hDOWN %Q S jMEANREC Q S kMBEQ 1 lMBEQ 2 mMBE Q 3 nQ 2 oQ3 pQ 4 qQ5 rQ 6 sQ 7 y MBE is a dummy variable equal to one if the surprise is positive or null, the variable POSEE is a binary variable which takes a value of one if the mean of the analyst estimates is positive or zero and 0 if it is negative, POSUE is a binary variable with a value of 1 if the mean analyst expectation is larger or equal to the EPS in the same quarter of last year and 0 otherwise. MTB is the market-to-book ratio of the firm in the previous quarter, LEV is the firm’s one-quarter lagged leverage and SIZE is the logarithm of the market capitalization on the day of the previous earnings announcement. STDEV_P is the price-scaled standard deviation of the EPS estimate, NBEST is the number of EPS estimates recorded in I/B/E/S, DOWN% is the number of downward revisions over the three months preceding the earnings release scaled by the number of estimates (NBEST). MEANREC is the mean analyst recommendation before the announcement, MBEL1, MBEL2 and MBEL3 are the one-, two- and three-quarter lagged MBE. Q2 is a dummy variable equal to 1 if the release was made in the second quarter of 2009, Q3 equals one if the release was made in the third quarter of 2009, Q4 has a value of 1 if the earnings were released in the fourth quarter of 2009 and so forth. Coefficient S.E. Wald p-value POSEE ,438 ,172 6,484 ,011 POSUE ,917 ,108 71,900 ,000 MTB -,002 ,005 ,147 ,701 LEV -,003 ,012 ,079 ,778 SIZE ,082 ,119 ,481 ,488 STDEV_P -,149 1,126 ,017 ,895 NBEST ,029 ,008 13,158 ,000 DOWN% -,870 ,114 58,639 ,000 MEANREC -,120 ,130 ,847 ,357 MBEL1 ,530 ,110 23,420 ,000 MBEL2 ,201 ,111 3,282 ,070 MBEL3 ,482 ,108 19,842 ,000 Q2 ,227 ,177 1,647 ,199 Q3 ,536 ,196 7,499 ,006 Q4 -,248 ,184 1,812 ,178 Q5 -,230 ,199 1,341 ,247 Q6 -,560 ,189 8,776 ,003 Q7 -,561 ,185 9,172 ,002 Q8 -,663 ,186 12,723 ,000 Constant (y) -,044 ,629 ,005 ,945 * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level 64 Table C OLS regression: predicting the market anticipation of the surprise Table C exhibits the results of the OLS regression (3a): CAR Q aE ( SUR Q ) bE ( MBE Q ) cQ 2 dQ3 eQ 4 fQ 5 gQ6 hQ7 iQ8 y The cumulative abnormal return (CAR) is the return calculated for the time window (-5;-1). It is obtained from adjusting the securities’ returns by the S&P500 return. The predicted surprise (E(SUR)) is the surprise that is predicted by regression (2) . The predicted probability of MBE (E(MBE)) is the predicted probability of MBE by regression (1). The seven quarter dummies are also included in this regression. CAR(-5,-1) Coefficient St. error t-test p-value Constant (y) ,011 ,004 2,373 ,018 E(SUR)** 2,964 1,174 2,525 ,012 E(MBE) -,001 ,006 -,206 ,837 Q2 -,002 ,003 -,651 ,515 Q3*** -,009 ,003 -3,547 ,000 Q4** -,006 ,003 -2,337 ,019 Q5*** -,007 ,003 -2,700 ,007 Q6*** -,011 ,003 -4,224 ,000 Q7*** -,012 ,002 -4,997 ,000 Q8*** -,013 ,002 -5,015 ,000 * : the variable is significant at the 10% level **: the variable is significant at the 5% level *** the variable is significant at the 1% level Table D Cumulative abnormal returns for firms that just beat the estimates Table XX displays the CAR (0) features for the observations with an unscaled surprise of 0, 0.1 and 0.2 95% Confidence Interval for Mean CAAR Std. Dev Std. Err Lower Bound Upper Bound SUR=0 -,013882 ,0520414 ,0030300 -,019845 -,007919 SUR=0.1 -,006874 ,0490754 ,0073157 -,021618 ,007870 SUR=0.2 -,002745 ,0517432 ,0056796 -,014044 ,008553 65 Table E OLS regression: predicting the excess surprise Table E provides the results of the OLS regression between the excess EPS surprise (actual EPS – expected EPS) and various variables. The expected surprise is produced by running regression (2). EST is the average EPS estimate of analysts, REMBE is a dummy variable equal to 1 if the three MBE lags are equal to 1. Coefficient Std. Error T test P value (Constant) -,002 ,001 -1,094 ,274 CAR(-5,-1) ,005 ,001 4,474 ,000 EST ,000 ,000 2,020 ,043 UP% -9,653E-5 ,000 -,849 ,396 DOWN% ,000 ,000 -1,532 ,126 MEANREC ,000 ,000 1,001 ,317 BUY% 1,003E-5 ,000 1,370 ,171 SELL% -1,599E-5 ,000 -1,855 ,064 LEV 2,171E-5 ,000 1,145 ,252 MTB -7,265E-6 ,000 -1,545 ,123 SIZE -8,046E-5 ,000 -,806 ,420 PRICE -6,757E-6 ,000 -3,895 ,000 STDEV_P ,363 ,037 9,896 ,000 REMBE ,000 ,000 -1,982 ,048 POSUE ,000 ,000 5,168 ,000 SURL1 -,018 ,013 -1,415 ,157 Table F LOGIT: predicting MBES Table E provides the results of the LOGIT regression between the MBES dummy variable (equal to one if the excess surprise is positive) and several other variables. EST is the average EPS estimate of analysts. Coefficient S.E. Wald p-value CAR(-5,-1) 3,561 1,087 10,734 ,001 EST ,392 ,106 13,700 ,000 DOWN% -,040 ,094 ,181 ,671 MEANREC ,530 ,429 1,525 ,217 BUY% ,010 ,007 2,241 ,134 SELL% -,006 ,008 ,647 ,421 LEV ,041 ,019 4,666 ,031 MTB -,012 ,005 6,321 ,012 PRICE -,012 ,002 33,106 ,000 STDEV_P 44,301 24,494 3,271 ,071 MBEL1 -,337 ,119 8,041 ,005 POSUE ,205 ,082 6,283 ,012 SURL1 29,370 12,232 5,765 ,016 66 Constant -1,911 1,290 2,196 ,138 Table G CAR distribution for the expected DSUR and MBES extreme quintiles This table shows the CAR (-2,0) distribution features for four quintiles. The first row concerns the lowest expected probability of MBES quintile, the second row includes the 20% observations with the lowest expected excess surprise (actual surprise – expected surprise; DSUR). MBES means “meeting or beating the expected surprise”. The third row includes the 20% observations with the highest expected probability of MBES and the fourth one includes the 20% observations with the highest expected excess surprise. Finally, the fifth row reports descriptive statistics on the CAR of the whole sample. CAAR St dev. Min Max Q1 Median Q3 Q1 MBES -,0013 ,06025 -,26 ,29 -,0356 -,0066 ,0272 Q1 DSUR -,0006 ,06051 -,24 ,29 -,0346 -,0045 ,0283 Q5 MBES ,0133 ,06109 -,24 ,31 -,0204 ,0090 ,0432 Q5 DSUR ,0143 ,06042 -,24 ,31 -,0216 ,0110 ,0440 Full sample ,0056 ,05866 -,35 ,36 -,0260 ,0026 ,0354 67